EN FR
EN FR

2025Activity​ reportProject-TeamEVERGREEN

RNSR:​‌ 202424505L
  • Research center Inria​​ Branch at the University​​​‌ of Montpellier
  • In partnership​ with:INRAE, CIRAD
  • Team​‌ name: Earth obserVation and​​ machine lEarning foR aGRo-Environmental​​​‌ challENges
  • In collaboration with:​Territoires, Environnement, Télédétection et​‌ Information Spatiale

Creation of​​ the Project-Team: 2024 January​​​‌ 01

Each year, Inria​ research teams publish an​‌ Activity Report presenting their​​ work and results over​​​‌ the reporting period. These​ reports follow a common​‌ structure, with some optional​​ sections depending on the​​​‌ specific team. They typically​ begin by outlining the​‌ overall objectives and research​​ programme, including the main​​​‌ research themes, goals, and​ methodological approaches. They also​‌ describe the application domains​​ targeted by the team,​​​‌ highlighting the scientific or​ societal contexts in which​‌ their work is situated.​​

The reports then present​​​‌ the highlights of the​ year, covering major scientific​‌ achievements, software developments, or​​ teaching contributions. When relevant,​​​‌ they include sections on​ software, platforms, and open​‌ data, detailing the tools​​ developed and how they​​​‌ are shared. A substantial​ part is dedicated to​‌ new results, where scientific​​ contributions are described in​​​‌ detail, often with subsections​ specifying participants and associated​‌ keywords.

Finally, the Activity​​ Report addresses funding, contracts,​​​‌ partnerships, and collaborations at​ various levels, from industrial​‌ agreements to international cooperations.​​ It also covers dissemination​​​‌ and teaching activities, such​ as participation in scientific​‌ events, outreach, and supervision.​​ The document concludes with​​​‌ a presentation of scientific​ production, including major publications​‌ and those produced during​​ the year.

Keywords

Computer​​​‌ Science and Digital Science​

  • A3.1.10. Heterogeneous data
  • A5.3.2.​‌ Sparse modeling and image​​ representation
  • A5.3.3. Pattern recognition​​​‌
  • A9.2. Machine learning
  • A9.2.1.​ Supervised learning
  • A9.2.2. Unsupervised​‌ learning
  • A9.2.6. Neural networks​​
  • A9.2.8. Deep learning
  • A9.3.​​​‌ Signal processing
  • A9.12.1. Object​ recognition

Other Research Topics​‌ and Application Domains

  • B3.1.​​ Sustainable development
  • B3.1.1. Resource​​​‌ management
  • B3.4.1. Natural risks​
  • B3.5. Agronomy
  • B3.6. Ecology​‌
  • B3.6.1. Biodiversity

1 Team​​ members, visitors, external collaborators​​​‌

Research Scientists

  • Dino Ienco​ [Team leader,​‌ INRAE, Senior Researcher​​]
  • Cassio Fraga Dantas​​​‌ [INRAE, Researcher​]
  • Raffaele Gaetano [​‌CIRAD, Researcher]​​
  • Roberto Interdonato [CIRAD​​​‌, Researcher, HDR​]
  • Diego Marcos Gonzalez​‌ [INRIA, Advanced​​ Research Position]

Post-Doctoral​​​‌ Fellow

  • Pallavi Jain [​CIHEAM, Post-Doctoral Fellow​‌]

PhD Students

  • Ananthu​​ Aniraj [INRIA]​​​‌
  • Bruno Bio Nikki Sare​ [CIRAD]
  • Christopher​‌ Jabea [UNIV COTE​​ D'AZUR, from Oct​​​‌ 2025]
  • Juan Li​ [INRAE, until​‌ Nov 2025]
  • Hugo​​ Riffaud–De Turckheim [INRIA​​​‌]
  • Pablo Ubilla Pavez​ [INRIA, from​‌ Oct 2025]
  • Quentin​​ Yeche [ATOS,​​​‌ CIFRE, until Oct​ 2025]
  • Anas Zakroum​‌ [CIRAD, from​​ Sep 2025]

Technical​​ Staff

  • Rémi Cresson [​​​‌INRAE, Engineer]‌
  • Christopher Jabea [INRAE‌​‌, Engineer, until​​ Aug 2025]
  • Anna​​​‌ Roussel [INRAE,‌ Engineer]

Interns and‌​‌ Apprentices

  • Adam Menoud [​​INRIA, Intern,​​​‌ from Oct 2025]‌
  • Julie Turpin [INRIA‌​‌, Intern, from​​ Jul 2025 until Aug​​​‌ 2025]

Administrative Assistant‌

  • Claire-Marine Parodi [INRIA‌​‌]

Visiting Scientists

  • Giuseppe​​ Guarino [UNIV NAPLES​​​‌, until Apr 2025‌]
  • Francisco Mena Toro‌​‌ [DFKI, until​​ Jan 2025]
  • Vito​​​‌ Recchia [UNIV BARI‌, from Sep 2025‌​‌]
  • Robicca Shamim [​​UNIV Turin, from​​​‌ Oct 2025]
  • Valérie‌ Zermatten [EPFL,‌​‌ until Feb 2025]​​

2 Overall objectives

Nowadays,​​​‌ modern space missions continuously‌ collect information about the‌​‌ earth surface, generating massive​​ amounts of data. The​​​‌ multitude of Earth Observation‌ (EO) systems allows to‌​‌ acquire data via different​​ sensors (e.g., optical, radar,​​​‌ LiDAR - Light Detection‌ And Ranging) at different‌​‌ spatial and temporal resolutions​​ with diverse spectral characteristics.​​​‌ On the one hand,‌ this huge volume of‌​‌ collected information opens up​​ new opportunities to better​​​‌ understand and monitor natural,‌ agricultural and other anthropized‌​‌ spaces at different scales.​​ On the other hand,​​​‌ the quantity and diversity‌ of the collected information‌​‌ sets up new challenges​​ to the remote sensing​​​‌ community. As a matter‌ of fact, in order‌​‌ to take the most​​ out of the digital​​​‌ revolution that is impacting‌ the domain, recent and‌​‌ future analysis tasks require​​ a paradigm shift towards​​​‌ data-intensive methodologies. The main‌ objectives of the EVERGREEN‌​‌ project-team is to develop​​ machine learning models and​​​‌ tools for the exploitation‌ and analysis of Earth‌​‌ Observation (EO) data in​​ accordance with the constraints​​​‌ of the operational settings‌ and in constant interaction‌​‌ with potential users and​​ targeted stakeholders. Examples of​​​‌ possible applications range from‌ land use/land cover mapping‌​‌ to natural resources monitoring,​​ including territorial planning as​​​‌ well as biodiversity mapping.‌ More broadly, that may‌​‌ include all possible EO-based​​ applications that support the​​​‌ modern agro-environmental transition.

3‌ Research program

EVERGREEN is‌​‌ an interdisciplinary team working​​ on the design and​​​‌ application of machine learning‌ techniques to deal with‌​‌ the analysis of Earth​​ observation data to support​​​‌ modern agro-environmental challenges. Our‌ research is organized along‌​‌ the three following research​​ axes:

  • Tailored Machine Learning​​​‌ methods for EO data‌ (Section 3.1) is the‌​‌ first fundamental research axis.​​ It focuses on advancing​​​‌ the methodologies related to‌ Satellite Image Time Series‌​‌ (SITS) management and on​​ tackling the multi-source exploitation​​​‌ of EO data.
  • Adoption‌ and development of advanced‌​‌ learning paradigms to support​​ Earth Observation data analysis​​​‌ (Section 3.2) is the‌ second fundamental research axis‌​‌ of the team. The​​ research objectives related to​​​‌ this axis are devoted‌ to make a step‌​‌ further with the exploitation​​ of multiple EO data​​​‌ sources, dealing with ground‌ truth paucity leveraging semi-supervised‌​‌ and self-supervised learning settings​​ and advancing the spatio/temporal​​​‌ transferability of machine learning‌ models for EO data.‌​‌
  • Interaction between Domain expert​​​‌ and Machine Learning models​ (Section 3.3) is the​‌ last research axis. Here​​ the goals are to​​​‌ introduce a priori knowledge​ to guide the learning​‌ process and design explainabilty/interpretability​​ neural network models.

3.1​​​‌ Tailored Machine Learning methods​ for EO data analysis​‌

The research objectives about​​ this topic are devoted​​​‌ to: i) advance the​ methodologies related to Satellite​‌ Image Time Series (SITS)​​ management and ii) tackle​​​‌ the multi-source exploitation of​ EO data. Improving the​‌ management of SITS data​​ requires to directly cope​​​‌ with signals that are​ non-stationary (the spectral and​‌ temporal responses over land​​ elements sharing the same​​​‌ spatio-temporal dynamics may vary​ across space and time),​‌ temporally discontinuous (sudden events​​ - e.g. human intervention​​​‌ - generally alter the​ signal responses), and/or affected​‌ by missing observations (e.g.​​ due to cloud coverage),​​​‌ but that can exhibit​ a strong spatial correlation​‌ (“close” observations of a​​ same land element are​​​‌ likely to be similar).​ To tackle such points,​‌ we aim to develop​​ new approaches capable of​​​‌ coping with missing values​ in SITS data, and​‌ to integrate external or​​ background knowledge to guide​​​‌ the learning process and​ explicitly consider the dependency​‌ of a SITS signal​​ with its spatial context.​​​‌ Concerning the multi-source exploitation​ of Earth Observation data,​‌ ad hoc solutions exist​​ but there is still​​​‌ a lack of a​ general methodological framework to​‌ leverage the complementarity of​​ different sources according to​​​‌ the considered downstream task.​ This is especially the​‌ case when one of​​ the involved sources is​​​‌ represented by SITS data.​ Our goal is to​‌ provide multi-source EO data​​ fusion paradigms related to​​​‌ a specific downstream task.​ For instance, if the​‌ downstream task is classification​​ or regression, our framework​​​‌ should reduce as much​ as possible the intermediate​‌ steps existing between the​​ raw data and their​​​‌ use for the particular​ task at hand (i.e.​‌ avoid resampling the data​​ source at the same​​​‌ spatial and/or temporal scale​ or avoid separating feature​‌ extraction and model calibration).​​ The reduction of such​​​‌ intermediate steps is directly​ related to possible bias​‌ affecting intermediate products as​​ well as the volume​​​‌ of data to manage.​

3.2 Advanced learning paradigms​‌ to support EO data​​ analysis

The research objectives​​​‌ related to this topic​ are devoted to: i)​‌ going further with the​​ complementary exploitation of multiple​​​‌ EO data sources, ii)​ dealing with ground truth​‌ paucity leveraging semi-supervised machine​​ learning settings and iii)​​​‌ advancing the spatio/temporal transferability​ of machine learning models​‌ for EO data. Ameliorating​​ the analysis of EO​​​‌ data, regarding particular applications,​ requires to intelligently manage​‌ heterogeneous and complementary information​​ taking the most out​​​‌ of the combination of​ the different sensors. To​‌ this end, we aim​​ to conceive and design​​​‌ new methodological frameworks for​ multi-source and cross-modal EO​‌ data analysis. In this​​ direction, we will investigate​​​‌ settings related to the​ general domain of Knowledge​‌ Distillation and Adversarial Training​​ to tackle the scenario​​​‌ in which some modalities​ are missing at inference​‌ time. While such methodological​​ settings are largely investigated​​ in the context of​​​‌ standard computer vision applications,‌ they are still unexplored‌​‌ in the remote sensing​​ field. Despite the huge​​​‌ amount of sensor data‌ we can dispose of‌​‌ on a study area,​​ the time and costly​​​‌ acquisition of ground truth‌ to calibrate machine and‌​‌ deep learning models can​​ negatively influence the deployment​​​‌ of such strategies in‌ an operational context. Here,‌​‌ we will provide research​​ studies contributing to the​​​‌ general domain of self-supervised‌ learning, partially labeled and‌​‌ semi-supervised scenarios (i.e. positive​​ unlabeled learning), spatial active​​​‌ learning strategies and, weakly‌ supervised setting. Last but‌​‌ not least, this axes​​ also targets the investigation​​​‌ of spatio/temporal transferability of‌ machine learning models for‌​‌ EO data analysis with​​ a particular focus on​​​‌ how to adapt a‌ classification model learnt on‌​‌ a study area to​​ generalize over another study​​​‌ area characterized by different‌ climate/environmental conditions as well‌​‌ as transfer a model​​ learnt on data coming​​​‌ from a time period‌ to data coming from‌​‌ the same, or similar,​​ study area acquired at​​​‌ a different period of‌ time.

3.3 Interaction between‌​‌ Domain expert and Machine​​ Learning models

The research​​​‌ objectives associated with this‌ topic are devoted to:‌​‌ i) integrate a priori​​ knowledge (expert or biophysical)​​​‌ in the learning process‌ of the machine learning‌​‌ models, ii) design learning​​ models that explicitly allow​​​‌ to interpret the decision‌ process under different dimensions‌​‌ (i.e. temporal, spatial, ...)​​ and iii) move towards​​​‌ multi-modal exploitation of EO‌ data. Concerning the first‌​‌ point, related to the​​ integration of a priori​​​‌ knowledge, both domain expert‌ and physical modeling can‌​‌ be exploited to guide​​ the exploration of model​​​‌ parameters, with the aim‌ to reduce the possible‌​‌ search space avoiding implausible​​ solutions supplied by the​​​‌ model. The second point‌ involves the design of‌​‌ learning models that explicitly​​ permit the interpretation and​​​‌ the explaination of the‌ decision process. This research‌​‌ direction is related to​​ the current necessity to​​​‌ get insights on how‌ machine learning models make‌​‌ their decisions with the​​ aim to supply additional​​​‌ information to the end‌ user and raising up‌​‌ the level of transparency​​ and trustworthiness associated to​​​‌ the decision process. The‌ third direction, related to‌​‌ the multi-modal exploitation of​​ EO data, cover aspects​​​‌ related to the integration‌ of EO data with‌​‌ non-EO data such as,​​ for instance, text or​​​‌ audio modalities. Such a‌ multi-modal analysis raises new‌​‌ challenges related to how​​ to integrate data coming​​​‌ from a remote sensing‌ modality with non spatially‌​‌ explicit information going further​​ with the analysis of​​​‌ heterogeneous data sources and‌ opening novel research questions‌​‌ about multi-modal data integration​​ and mining.

4 Application​​​‌ domains

The application scope‌ of the team is‌​‌ mainly guided by the​​ application domains of the​​​‌ INRAE et CIRAD partners‌ with applications related to‌​‌ agricultural and environmental monitoring​​ and assessment. In addition,​​​‌ the application domains of‌ the team is constantly‌​‌ growing and changing with​​ the aim to answer​​​‌ to societal questions related‌ to the modern agro-environmental‌​‌ transition and the challenges​​​‌ it is raising up.​

4.1 Food Security

The​‌ role of remote sensing​​ in the assessment of​​​‌ food security indicators, especially​ those concerning food supply​‌ (one of the four​​ pillars for food security​​​‌ along with accessibility, quality​ and stability) through the​‌ monitoring of agricultural activities,​​ has been long proven​​​‌ in the last decades.​ Products typically targeted by​‌ these activities go, across​​ scales, from cropland and​​​‌ crop type mapping, to​ the detection of anomalies​‌ in vegetation growth as​​ well as crop yield​​​‌ estimation and forecast. Leveraging​ remote sensing for the​‌ design of novel spatio-temporal​​ indicators related to agricultural​​​‌ production and food security​ is of paramount importance​‌ to support policy makers​​ and social actors in​​​‌ their decision processes. Additionally,​ remote sensing derived information​‌ can provide up-to-date information​​ in order to assess​​​‌ the underlying sustainability of​ the agricultural production. This​‌ is even more important​​ in the context of​​​‌ tropical agricultural systems.

This​ domain application is at​‌ the core of the​​ EVERGREEN activity with multiple​​​‌ research efforts devoted to​ the analysis of land​‌ use and land cover​​ mapping that are of​​​‌ fundamental importance in order​ to subsequently extract spatio-temporal​‌ indicators to characterize agricultural​​ production. For instance, actions​​​‌ related to this application​ domain are conducted in​‌ the context of the​​ CIFRE PhD thesis of​​​‌ Quentin Yeche (INRAE/ATTOS) on​ the topic of parcel​‌ identification/extraction and characterization, the​​ PhD thesis of Azza​​​‌ Abidi (University of Montpellier,​ University of Manouba/Tunisie) and​‌ Bruno Bio Nikki (CIRAD)​​ related to the analysis​​​‌ of multi-temporal/multi-source remote sensing​ data for land cover​‌ mapping in conventional and​​ tropical agricultural systems.

4.2​​​‌ Forest monitoring

New challenges​ are arising about the​‌ quantity and quality of​​ the information about forest​​​‌ cover which can effectively​ support decision making processes​‌ at global, national and​​ local scales. The need​​​‌ for a deeper description​ of the forest cover​‌ seems to emerge, by​​ means of a larger​​​‌ set of biophysical or​ structural indicators carrying information​‌ about its diversity in​​ terms of species and​​​‌ their “role” in the​ landscape. Notable examples are​‌ i) the possibility of​​ discriminating between proper forests​​​‌ and tree cover related​ to agricultural exploitation, which​‌ do not provide the​​ same carbon sequestration potential,​​​‌ and ii) to precisely​ identify the spatial extent​‌ of a forest cover,​​ also with respect to​​​‌ changes in its density​ and “greenness”, especially in​‌ transition areas between different​​ ecosystems and/or climatic zones.​​​‌ Among these indicators, many​ have been proven to​‌ be derived from EO​​ data. The EVERGREEN team​​​‌ is increasing its activities​ related to the analysis​‌ and monitoring of forest​​ areas due to the​​​‌ paramount importance of this​ natural resource. For instance,​‌ members of the EVERGREEN​​ team are involved in​​​‌ both European (HORIZON Eco2Adapt)​ and National projects (ANR​‌ PREDISPOSE) related to the​​ analysis of forest disturbances​​​‌ (i.e. forest fires) or​ they are collaborating with​‌ international partners (i.e. DLR/Germany,​​ Wageningen University/Netherlands) on the​​​‌ analysis of forest covers​ and its properties in​‌ Southern Countries (i.e. Africa).​​

4.3 Biodiversity mapping and​​ monitoring

Biodiversity loss is​​​‌ now considered to be‌ an existential threat to‌​‌ humankind on par with​​ climate change. In order​​​‌ to advance the understanding‌ of the underlying phenomena‌​‌ behind this phenomena, we​​ first need to have​​​‌ a clear picture of‌ the current ranges and‌​‌ population densities of species​​ globally. This is a​​​‌ crucial challenge that will‌ require vast amounts of‌​‌ data in the form​​ of species observations coupled​​​‌ with Earth observation-based habitat‌ suitability. To this end,‌​‌ our objective is to​​ link ground-level data to​​​‌ remote sensing imagery in‌ order to map fundamental‌​‌ niches of species and​​ monitor their spatial shifts​​​‌ under climate change and‌ other anthropogenic pressure. Actions‌​‌ related to this application​​ domain cover the collaboration​​​‌ with the Iroko team,‌ via the co-superivision of‌​‌ a post-doc researcher on​​ species distribution modeling and​​​‌ spatial biases, international collaborations‌ covering the development of‌​‌ visual-language model for ecology​​ mapping with the EPFL​​​‌ University (Switzerland) through the‌ visiting research stay of‌​‌ Valerie Zermatten and the​​ national project IMPACT, funded​​​‌ by the OFB (Office‌ français de la biodiversité)‌​‌ on the detection of​​ possible plant disease outbreaks​​​‌ via remote sensing multi-temporal‌ data.

5 Social and‌​‌ environmental responsibility

5.1 Footprint​​ of research activities

  • Work​​​‌ trips. While the‌ sanitary crisis had drastically‌​‌ cut the number of​​ work trips of the​​​‌ team, recent years have‌ seen an increase in‌​‌ the physical participation in​​ conferences and various committees.​​​‌ However compared to the‌ pre-covid period, one can‌​‌ note that the majority​​ of movements are national​​​‌ or at best European,‌ with very few trips‌​‌ outside of Europe and,​​ when it is possible,​​​‌ trains are preferred to‌ planes.
  • Utilization of computing‌​‌ resources. Being a team​​ specialized in computer vision​​​‌ and machine learning for‌ remote sensing data, a‌​‌ recurrent task in EVERGREEN​​ is to run CPU/GPU-intensive​​​‌ algorithms on large data‌ collections. To this end,‌​‌ our strategy towards computing​​ resources is driving us​​​‌ to increase the use‌ of regional, national and‌​‌ institutional infrastructures (i.e. Jean​​ Zay, ABACA) in order​​​‌ to leverage sustainable computing‌ platforms instead of local‌​‌ server/workstation with a general​​ positive effect on energy​​​‌ consumption.

5.2 Impact of‌ research results

We estimate‌​‌ that our research work​​ can have several impacts​​​‌ on the society due‌ to the fact that‌​‌ EVERGREEN is working on​​ challenges related to modern​​​‌ agro-environmental analysis. We give‌ below two examples of‌​‌ impact of our research​​ results:

  • Most of the​​​‌ research work is conducted‌ in collaboration with scientists‌​‌ from environmental and agricultural​​ sciences considering both applied​​​‌ research and operational scenarios.‌ Such interdisciplinary work paves‌​‌ the way to the​​ deployment of our research​​​‌ contributions in projects related‌ to a more sustainable‌​‌ and reasoned management of​​ natural (i.e. forest, water,​​​‌ ...) and agricultural resources.‌
  • A part of our‌​‌ research work is conducted​​ in partnership with companies,​​​‌ through CIFRE PhDs and‌ collaboration actions. Hence, the‌​‌ addressed research problems concern​​ an important challenge for​​​‌ the companies, and the‌ solutions proposed are evaluated‌​‌ on their relevance to​​​‌ tackle this challenge.

6​ Highlights of the year​‌

  • Article at the Thirty-Ninth​​ AAAI Conference on Artificial​​​‌ Intelligence (AAAI-25) on hybrid​ phenology modeling for predicting​‌ temperature effects on tree​​ dormancy 24.
  • Article​​​‌ in the Nature Communications​ journal on the effective​‌ integration of drone technology​​ for mapping and managing​​​‌ palm species in the​ Peruvian Amazon.
  • Organization of​‌ the 7th edition of​​ the MACLEAN (Machine Learning​​​‌ of Earth Observation) workshop​ co-located with the ECML/PKDD​‌ 2025 conference (see link​​) and the co-organization​​​‌ of the MVEO (Machine​ Vision for Earth Observation​‌ and Environment Monitoring) workshop​​ co-located with BMVC 2025​​​‌ conference (see link).​

6.1 Awards

  • Best presentation​‌ award at the British​​ Machine Vision Conference (BMVC)​​​‌ 2025.

7 Latest software​ developments, platforms, open data​‌

7.1 New platforms

7.1.1​​ MORINGA: an open-source platform​​​‌ for automatic land cover​ classification from multi-sensor imagery​‌

Participants: Raffaele Gaetano.​​

Started in 2015 in​​​‌ the framework of the​ activities of the Land​‌ Cover Scientific Expertise Center​​ as part of the​​​‌ French Land Surface Data​ and Services Hub -​‌ THEIA, the developement of​​ the MORINGA 1 processing​​​‌ chain was initially aimed​ at providing a “turn​‌ key” solution, addressed to​​ thematic specialists with relatively​​​‌ low programming skills, for​ the automatic land cover​‌ classification from multi-sensor, multi-resolution​​ and multi-temporal satellite imagery.​​​‌ It particularly targets the​ needs for accurate land​‌ cover mapping in the​​ context of tropical agricultural​​​‌ systems, where several specificities​ (landscape heterogeneity and fragmentation,​‌ small field sizes, high​​ cloud coverage during cropping​​​‌ seasons) call for the​ combination of different resolutions​‌ and acquisition modes to​​ both capture spatial details​​​‌ and temporal profiles. Leveraging​ an object-based approach and​‌ a suitable supervised classification​​ framework based on legacy​​​‌ machine learning techniques, the​ MORINGA processing chain takes​‌ in charge imagery provided​​ by different satellite missions,​​​‌ both at high (Sentinel-1​ and -2, Landsat 8/9,​‌ etc.) and very high​​ (Pléiades, SPOT6/7) spatial resolution,​​​‌ and automatically manages their​ pre-processing and ingestion in​‌ the object based machine​​ learning framework, with limited​​​‌ user interaction. Reference data​ are also automatically processed​‌ to provide the best​​ possible validation also in​​​‌ cases of data paucity,​ which are rather common​‌ in the targeted application.​​ To date, MORINGA has​​​‌ become a feature-rich, modular​ platform for remote sensing​‌ image analysis, which can​​ also be used as​​​‌ a lower-level API to​ ease common image processing​‌ tasks. Thanks to the​​ support of thematic experts​​​‌ and cartography specialists, it​ has since been used​‌ for the production of​​ high quality land cover​​​‌ maps in many different​ scientific and dissemination contexts​‌ (see 33 for a​​ notable example in 2024).​​​‌ The software package is​ currently bound to evolve​‌ to a larger remote​​ sensing based land cover​​​‌ workbench, including novel deep​ learning techniques for both​‌ image pre-processing/enhancement and multi-sensor​​ classification.

8 New results​​​‌

In this section, we​ briefly summarize and reference​‌ the major research results​​ published in 2025. The​​​‌ research works are organized​ into three subsections: i)​‌ Tailored Machine Learning Methods​​ for EO Data Analysis;​​ ii) Advanced Learning Paradigms​​​‌ to Support EO Data‌ Analysis; and iii) Interaction‌​‌ Between Domain Experts and​​ Machine Learning Models.

8.1​​​‌ Tailored Machine Learning methods‌ for EO data analysis‌​‌

8.1.1 Geographical Context Matters:​​ Bridging Fine and Coarse​​​‌ Spatial Information to Enhance‌ Continental Land Cover Mapping‌​‌

Participants: Cássio Fraga Dantas​​, Raffaele Gaetano,​​​‌ Dino Ienco.

Collaborators‌: Babak Ghassemi (BOKU‌​‌ - University of Natural​​ Resources and Life Sciences,​​​‌ Vienna, Austria), Omid Ghorbanzadeh‌ (BOKU - University of‌​‌ Natural Resources and Life​​ Sciences, Vienna, Austria), Emma​​​‌ Izquierdo-Verdiguier (BOKU - University‌ of Natural Resources and‌​‌ Life Sciences, Vienna, Austria),​​ Francesco Vuolo (BOKU -​​​‌ University of Natural Resources‌ and Life Sciences, Vienna,‌​‌ Austria).

Keywords: Geospatial​​ Analysis, Deep Learning, Large-Scale​​​‌ Analysis, Land Cover Mapping.‌

Land use and land‌​‌ cover mapping from Earth​​ Observation (EO) data is​​​‌ a critical tool for‌ sustainable land and resource‌​‌ management as, for instance,​​ in domains like biodiversity​​​‌ and agricultural food production.‌ While advanced machine learning‌​‌ and deep learning algorithms​​ excel at analyzing EO​​​‌ imagery data, they often‌ overlook crucial geospatial metadata‌​‌ information that could enhance​​ scalability and accuracy across​​​‌ regional, continental, and global‌ scales. To address this‌​‌ limitation, we propose BRIDGE-LC​​ (Bi-level Representation Integration for​​​‌ Disentangled GEospatial Land Cover),‌ a novel deep learning‌​‌ framework that explicitly integrates​​ multi-scale geospatial information into​​​‌ the land cover classification‌ process. By simultaneously leveraging‌​‌ fine-grained (latitude/longitude) and coarse-grained​​ (biogeographical region) spatial information,​​​‌ our lightweight multi-layer perceptron‌ architecture learns from both‌​‌ multi-scale information during training​​ but only requires fine-grained​​​‌ information for inference, allowing‌ it to disentangle region-specific‌​‌ from region-agnostic land cover​​ features while maintaining computational​​​‌ efficiency comparable with standard‌ machine learning approaches. To‌​‌ assess the quality of​​ our framework, we use​​​‌ an open-access in-situ dataset‌ spanning the 27 countries‌​‌ of the European Union​​ and we adopt several​​​‌ competing classification approaches commonly‌ considered for large-scale land‌​‌ cover mapping. A visual​​ sketch of the proposed​​​‌ framework is depicted in‌ Figure 1. We‌​‌ evaluated all the approaches​​ through two scenarios: an​​​‌ extrapolation scenario in which‌ training data encompasses samples‌​‌ coming from all the​​ biogeographical regions and a​​​‌ leave-one-region-out scenario where samples‌ from all the regions,‌​‌ except one, are employed​​ for the training stage.​​​‌ Additionally, we also explore‌ the spatial representation learned‌​‌ by the proposed deep​​ learning model, highlighting a​​​‌ connection between its internal‌ manifold and the geographical‌​‌ information used during the​​ training stage. Our results​​​‌ demonstrate that integrating geospatial‌ information improves land cover‌​‌ mapping performances, with the​​ most substantial gains achieved​​​‌ by jointly leveraging both‌ fine-grained and coarse-grained spatial‌​‌ information.

Figure 1

The image depicts​​ a machine learning model​​​‌ architecture for spatial data‌ analysis. It starts with‌​‌ latitude and longitude inputs​​ that undergo fixed positional​​​‌ encoding, enhanced by a‌ multi-layer perceptron (MLP) to‌​‌ produce learned positional encoding.​​ This encoding, combined with​​​‌ satellite input features, generates‌ region-invariant and region-specific embeddings.‌​‌ The region-invariant embedding is​​ processed by an invariant​​​‌ encoder and a land‌ cover (LC) classifier, optimizing‌​‌ for classification loss (Lclf)​​​‌ and contrastive loss (Lcon).​ The region-specific embedding is​‌ processed by a specific​​ encoder and a zone​​​‌ classifier, optimizing for zone​ classification loss (Lzone). The​‌ model integrates fine-grained spatial​​ information for detailed analysis​​​‌ and coarse-grained spatial info​ for broader spatial context.​‌ Regions are classified into​​ categories like Continental, Atlantic,​​​‌ and Boreal.

Figure 1​: Our framework, in​‌ addition to the standard​​ land cover classification branch​​​‌ (yellow), which processes reflectance-based​ input features, we introduce​‌ two complementary components: (1)​​ a fine-grained spatial information​​​‌ branch (green) utilizing latitude-longitude​ coordinates, and (2) a​‌ coarse-grained spatial information branch​​ (red) based on biogeographical​​​‌ region labels. During training,​ the red branch guides​‌ the model to learn​​ region-invariant embeddings that are​​​‌ more robust and generalizable​ for land cover classification.​‌ This branch is used​​ only during training and​​​‌ is successively discarded at​ inference time.

This work​‌ has been published at​​ the Science of Remote​​​‌ Sensing journal (Elsevier) 14​.

8.1.2 Environmental and​‌ bioclimatic data for epidemiological​​ analysis over French Mediterranean​​​‌ areas

Participants: Dino Ienco​.

Collaborators: Camille​‌ Portes (UR BioSP, INRAE,​​ France), Eric Verdin (UMR​​​‌ Pathologie Végétale, INRAE, France),​ Edith Gabriel (UR BioSP,​‌ INRAE, France)

Keywords:​​ climate, epidemiology, land type,​​​‌ machine learning.

Risk-based surveillance​ is now a well-established​‌ paradigm in epidemiology, involving​​ the distribution of sampling​​​‌ efforts differentially in time,​ space, and within populations,​‌ based on multiple risk​​ factors. To assess and​​​‌ map the risk of​ the presence of the​‌ bacterium Xylella fastidiosa, we​​ have compiled a dataset​​​‌ that includes factors influencing​ plant development and thus​‌ the spread of such​​ harmful organism. To this​​​‌ end, we have collected,​ preprocessed, and gathered information​‌ and data related to​​ land types, soil compositions,​​​‌ and climatic conditions to​ predict and assess the​‌ probability of risk associated​​ with X. fastidiosa in​​​‌ relation to environmental features.​ This resource can be​‌ of interest to researchers​​ conducting analyses on X.​​​‌ fastidiosa and, more generally,​ to researchers working on​‌ geospatial modeling of risk​​ related to plant infectious​​​‌ diseases.

This work has​ been published at the​‌ Environmental Data Science journal​​ (Cambridge Press) 21.​​​‌

8.1.3 Assessing habitat suitability​ for black grouse broods​‌ at the bioregional scale​​

Participants: Dino Ienco.​​​‌

Collaborators: Alexandre Defossez​ (UMR TETIS, INRAE, France),​‌ Samuel Alleaume (UMR TETIS,​​ INRAE, France), Marc Montadert​​​‌ (OFB - Office français​ de la biodiversité, France),​‌ Josselin Giffard-Carlet (UMR TETIS,​​ INRAE, France), Nadia Guiffant​​​‌ (UMR TETIS, INRAE, France),​ Sandra Luque (UMR TETIS,​‌ INRAE, France)

Keywords:​​ Lyrurus tetrix, brood habitat​​​‌ suitability, species distribution models,​ remote sensing, wildlife monitoring.​‌

The black grouse Lyrurus​​ tetrix, a galliform species​​​‌ emblematic of the European​ Alps, is currently threatened​‌ by habitat change, particularly​​ given the closure of​​​‌ heathland linked to the​ rising tree line at​‌ higher altitudes. The presence​​ of heathlands in good​​​‌ ecological condition is, however,​ imperative for the species'​‌ reproduction. In this study,​​ we attempted to map​​​‌ black grouse brood habitat​ suitability at a bioregional​‌ scale in the French​​ Alps, coupling a species​​ distribution model with multi-source​​​‌ remote sensing data. To‌ predict brood habitat suitability,‌​‌ we used a random​​ forest ensemble model. Altitude,​​​‌ ericaceous heathland, and the‌ annual maximum normalized difference‌​‌ vegetation index (NDVI) emerged​​ as the three most​​​‌ important variables, consistent with‌ the ecological needs of‌​‌ black grouse. The proportion​​ of ericaceous heathland was​​​‌ especially representative of the‌ foraging and vegetation cover‌​‌ needs of black grouse​​ hens. The resulting map​​​‌ was evaluated by black‌ grouse experts and found‌​‌ to be consistent with​​ their local knowledge in​​​‌ the context of the‌ French Alps.

This work‌​‌ has been published at​​ the Wildlife Biology journal​​​‌ (Nordic Council for Wildlife‌ Research) 13.

8.1.4‌​‌ Rapeseed mapping using machine​​ learning methods and Sentinel-1​​​‌ time series coupled with‌ growing degree-days information

Participants:‌​‌ Cássio Fraga Dantas,​​ Dino Ienco.

Collaborators​​​‌: Saeideh Maleki (UMR‌ TETIS, INRAE, France), Nicolas‌​‌ Baghdadi (UMR TETIS, INRAE,​​ France), Sami Najem (UMR​​​‌ TETIS, INRAE, France), Ya‌ Gao (NSSC - National‌​‌ Space Science Center [Beijing],​​ China), Hassan Bazzi (UMR​​​‌ TETIS, AgroParisTech, France)

Keywords‌: InceptionTime, Random Forest,‌​‌ Sentinel-1 SAR image time​​ series, Growing degree days.​​​‌

In light of recent‌ escalations of geopolitical conflicts‌​‌ around the world, mapping​​ rapeseed areas has garnered​​​‌ great interest given its‌ importance to food security.‌​‌ Sentinel-1 (S1) SAR data​​ was used for timely​​​‌ and regular rapeseed mapping.‌ By coupling S1 data‌​‌ with GDD (Growing Degree​​ Days) information, S1 GDD​​​‌ series were also created‌ and assessed. Rapeseed classification‌​‌ was realized using random​​ forest (RF) and inception​​​‌ time (IT). An alignment‌ method based on detected‌​‌ flowering dates was proposed​​ with the aim of​​​‌ alleviating the possible shifts‌ in the growth cycle‌​‌ between the different sites​​ and years. The spatial​​​‌ (cross-regional) transferability of the‌ models was tested accordingly,‌​‌ before and after alignment.​​ The results showed that​​​‌ using the S1 time‌ series before alignment, the‌​‌ overall F1-score achieved by​​ RF was 81.3%, and​​​‌ the overall F1-score of‌ IT was 89.2%. After‌​‌ alignment, RF achieved an​​ overall F1-score of 90.3%,​​​‌ while the overall F1-score‌ of IT was 91.7%.‌​‌ Using the S1 GDD​​ series, before alignment, the​​​‌ overall score of RF‌ was 58.2%, while the‌​‌ overall F1-score achieved by​​ IT was 86.6%. After​​​‌ the alignment of the‌ S1 GDD series, the‌​‌ F1-score achieved by RF​​ was 73.9%, and the​​​‌ F1-score of IT was‌ 86.8%. The best configuration‌​‌ for rapeseed mapping was​​ using IT with S1​​​‌ time series after alignment,‌ as it gave the‌​‌ highest overall F1-score and​​ the best consistency with​​​‌ the lowest standard deviation.‌ Overall, the S1 time‌​‌ series provided better results​​ than the S1 GDD​​​‌ series, meaning that employing‌ thermal time does not‌​‌ enhance the classification performance.​​ The results indicate that​​​‌ the proposed method enables‌ reliable, timely and continuous‌​‌ rapeseed monitoring, Paving the​​ way for more effective​​​‌ food stock management and‌ planning by policymakers and‌​‌ stakeholders.

This work has​​ been published at the​​​‌ Science of Remote Sensing‌ journal (Elsevier) 20.‌​‌

8.1.5 Sentinel-1 (S1) time​​​‌ series alignment method for​ rapeseed fields mapping

Participants:​‌ Cássio Fraga Dantas,​​ Dino Ienco.

Collaborators​​​‌: Saeideh Maleki (UMR​ TETIS, INRAE, France), Nicolas​‌ Baghdadi (UMR TETIS, INRAE,​​ France), Sami Najem (UMR​​​‌ TETIS, INRAE, France), Hassan​ Bazzi (UMR TETIS, AgroParisTech,​‌ France)

Keywords: InceptionTime,​​ Random Forest, classification algorithm,​​​‌ Synthetic Aperture Radar, machine​ learning algorithms.

Mapping rapeseed​‌ fields plays a crucial​​ role in agricultural management​​​‌ as rapeseed being a​ major source for oilseed,​‌ protein meal, livestock feed,​​ and industrial liquid biofuels.​​​‌ Accurately monitoring the distribution​ and characteristics of rapeseed​‌ fields enables farmers and​​ decision-makers to make informed​​​‌ decisions regarding fertilizer application,​ optimize harvest dates, and​‌ estimate yield. The conducted​​ research presents a comprehensive​​​‌ analysis of rapeseed fields​ mapping using Sentinel-1 (S1)​‌ time series data. We​​ applied a time series​​​‌ alignment method to enhance​ the accuracy of rapeseed​‌ fields detection, even in​​ scenarios where reference label​​​‌ data are limited or​ not available.

This work​‌ has been published at​​ the Frontiers in Remote​​​‌ Sensing journal (Frontiers) 16​.

8.1.6 Effective integration​‌ of drone technology for​​ mapping and managing palm​​​‌ species in the Peruvian​ Amazon

Participants: Diego Marcos​‌.

Collaborators: Ximena​​ Tagle (Wageningen University, The​​​‌ Netherlands), Rodolfo Cardenas-Vigo (IIAP,​ Perú), Martin Herold (GFZ​‌ Potsdam, Germany), Timothy R.​​ Baker (University of Leeds,​​​‌ UK) and others

Keywords​: Tree species mapping,​‌ UAV imagery, instance segmentation,​​ Peruvian Amazonia.

Remote sensing​​​‌ data could increase the​ value of tropical forest​‌ resources by helping to​​ map economically important species.​​​‌ However, current tools lack​ precision over large areas,​‌ and remain inaccessible to​​ stakeholders. Here, we work​​​‌ with the Protected Areas​ Authority of Peru to​‌ develop and implement precise,​​ landscape-scale, species-level methods to​​​‌ assess the distribution and​ abundance of economically important​‌ arborescent Amazonian palms using​​ field data, visible-spectrum drone​​​‌ imagery and deep learning.​ We compare the costs​‌ and time needed to​​ inventory and develop sustainable​​​‌ fruit harvesting plans in​ two communities using traditional​‌ plot-based and our drone-based​​ methods. Our approach detects​​​‌ individual palms of three​ species, even when densely​‌ clustered (average overall score,​​ 74%), with high accuracy​​​‌ and completeness for Mauritia​ flexuosa (precision; 99% and​‌ recall; 81%). Compared to​​ plot-based methods, our drone-based​​​‌ approach reduces costs per​ hectare of an inventory​‌ of Mauritia flexuosa for​​ a management plan by​​​‌ 99% (USD 5/ha versus​ USD 411/ha), and reduces​‌ total operational costs and​​ personnel time to develop​​​‌ a management plan by​ 23% and 36%, respectively.​‌ These findings demonstrate how​​ tailoring technology to the​​​‌ scale and precision required​ for management, and involvement​‌ of stakeholders at all​​ stages, can help expand​​​‌ sustainable management in the​ tropics.

This work has​‌ been published in Nature​​ Communications 22.

8.1.7​​​‌ Fully automatic extraction of​ morphological traits from the​‌ web: Utopia or reality?​​

Participants: Diego Marcos.​​​‌

Collaborators: Robert van​ de Vlasakker (Wageningen University,​‌ The Netherlands), Ioannis N.​​ Athanasiadis (Wageningen University, The​​​‌ Netherlands), Pierre Bonnet (UMR​ AMAP, Cirad, France), Hervé​‌ Goëau (UMR AMAP, Cirad,​​ France), Alexis Joly (IROKO,​​ Inria, France), W. Daniel​​​‌ Kissling (University of Amsterdam,‌ The Netherlands), César Leblanc‌​‌ (IROKO, Inria, France), André​​ S. J. van Proosdij​​​‌ (Wageningen University, The Netherlands)‌

Keywords: Tree species‌​‌ mapping, UAV imagery, instance​​ segmentation, Peruvian Amazonia.

Plant​​​‌ morphological traits, their observable‌ characteristics, are fundamental to‌​‌ understanding the role played​​ by each species within​​​‌ its ecosystem; however, compiling‌ trait information for even‌​‌ a moderate number of​​ species is a demanding​​​‌ task that may take‌ experts years to accomplish.‌​‌ At the same time,​​ online species descriptions contain​​​‌ massive amounts of information‌ about morphological traits, but‌​‌ the lack of structure​​ makes this source of​​​‌ data impossible to use‌ at scale. To overcome‌​‌ this, we propose to​​ leverage recent advances in​​​‌ large language models and‌ devise a mechanism for‌​‌ gathering and processing plant​​ trait information in the​​​‌ form of unstructured textual‌ descriptions, without manual curation.‌​‌ We evaluate our approach​​ by automatically replicating three​​​‌ manually created species–trait matrices.‌ Our method found values‌​‌ for over half of​​ all species–trait pairs, with​​​‌ an F1 score of‌ over 75%. Our results‌​‌ suggest that large-scale creation​​ of structured trait databases​​​‌ from unstructured online text‌ is now feasible due‌​‌ to the information extraction​​ capabilities of large language​​​‌ models. However, the process‌ is currently limited by‌​‌ the availability of textual​​ descriptions that cover all​​​‌ traits of interest.

This‌ work has been published‌​‌ in Applications in Plant​​ Sciences 17.

8.2​​​‌ Advanced learning paradigms to‌ support EO data analysis‌​‌

8.2.1 Deep learning interpretability​​ for understanding forest disturbance​​​‌ driver classification from Sentinel-1‌ and -2 data

Participants:‌​‌ Diego Marcos.

Collaborators​​: Laura Elena Cué​​​‌ La Rosa (Wageningen University,‌ The Netherlands), Jonas van‌​‌ Duijvenbode (WWF), Zillah Calle​​ (Wageningen University, The Netherlands),​​​‌ Jorn Dallinga (WWF) and‌ Johannes Reiche (Wageningen University,‌​‌ The Netherlands)

Keywords:​​ Explainable AI, Sentinel-1, Sentinel-2,​​​‌ Forest Disturbance driver, Feature-level‌ fusion.

Monitoring the drivers‌​‌ of tropical forest disturbances​​ using remote-sensing data has​​​‌ become increasingly critical to‌ supporting actionable law enforcement‌​‌ and sustainable land management.​​ Using information captured by​​​‌ multi-source Earth Observation data,‌ deep learning-based fusion models‌​‌ are state-of-the-art in many​​ remote sensing applications. Despite​​​‌ their efficacy, the inherent‌ black-box nature of these‌​‌ deep neural networks poses​​ challenges to our understanding​​​‌ of their decision-making processes.‌ To enhance their interpretability,‌​‌ we applied eXplainable Artificial​​ Intelligence (XAI) methods for​​​‌ several deep learning-based models‌ including single and multi-modal‌​‌ approaches. We evaluated six​​ XAI methods: Integrated Gradients,​​​‌ GradientShap, Saliency, Deconvolution, Guided‌ Grad-CAM, and Guided Backpropagation.‌​‌ Using both quantitative and​​ qualitative assessments, we conducted​​​‌ extensive experiments to evaluate‌ the capability of each‌​‌ XAI method to interpret​​ the proposed models. Our​​​‌ analysis included variable importance,‌ single- and multi-class explanations,‌​‌ cloud cover analysis, and​​ instances of misclassification. We​​​‌ identified Guided Grad-CAM as‌ the most reliable of‌​‌ these methods. In addition,​​ we gained deeper insight​​​‌ into how positive and‌ negative attribution scores influence‌​‌ the interpretation of model​​ output, highlighting the need​​​‌ for more research on‌ the significance of negative‌​‌ values. Our study improves​​​‌ the understanding of deep​ learning model decisions in​‌ the context of forest​​ disturbance driver classification, shedding​​​‌ light on the interpretability​ of fusion models and​‌ dataset characteristics. It establishes​​ a connection between remote​​​‌ sensing applications and XAI​ methodologies. This work was​‌ supported by the Open​​ Domain Science project Forest​​​‌ Carbon Crime under Grant​ OCENW.M.21.203; Nederlandse Organisatie voor​‌ Wetenschappelijk Onderzoek (NWO); Norway’s​​ Climate and Forest Initiative​​​‌ (NICFI) and World Wide​ Fund for Nature (WWF)​‌ the Netherlands.

This work​​ has been published at​​​‌ the International Journal of​ Remote Sensing (Taylor &​‌ Francis) 11.

8.2.2​​ SAHARA: Heterogeneous Semi-Supervised Transfer​​​‌ Learning with Adversarial Adaptation​ and Dynamic Pseudo-Labeling

Participants:​‌ Cássio Fraga Dantas,​​ Raffaele Gaetano, Dino​​​‌ Ienco.

Collaborators:​ Giuseppe Guarino (University of​‌ Naples Parthenope), Gemine Vivone​​ (CNR - National Research​​​‌ Council of Italy), Matteo​ Ciotola (University of Naples​‌ Parthenope), Giuseppe Scarpa (University​​ of Naples Parthenope)

Keywords​​​‌: Domain Adaptation, Heterogeneous​ data, Pseudo-labeling, Feature Disentanglement,​‌ Adversarial Learning.

Semi-supervised domain​​ adaptation aims to transfer​​​‌ knowledge from a labeled​ source domain to a​‌ scarcely labeled target domain,​​ despite distribution shifts. The​​​‌ challenge becomes greater when​ source and target data​‌ differ in acquisition modality,​​ as in remote sensing​​​‌ where variations in sensor​ type (e.g., optical vs.​‌ radar), spectral properties (e.g.,​​ RGB vs. multispectral), or​​​‌ spatial resolution are common.​ This challenging scenario, known​‌ as Semi-Supervised Heterogeneous Domain​​ Adaptation (SSHDA), requires learning​​​‌ across modalities with limited​ target labels. In this​‌ work, we have proposed​​ SAHARA (Semi-supervised Adaptation in​​​‌ Heterogeneous domains via conditional​ Adversarial Representation disentanglement and​‌ Adaptive pseudo-labeling), a new​​ method for SSHDA that​​​‌ combines conditional adversarial feature​ adaptation with dynamic pseudo-labeling​‌ to learn domain-invariant features​​ and handle extremely scarce​​​‌ target annotations. Experiments on​ two heterogeneous remote sensing​‌ benchmarks for scene classification,​​ conducted with both convolutional​​​‌ and transformer-based backbones, demonstrate​ that SAHARA consistently outperforms​‌ existing SSHDA and semi-supervised​​ methods.

This work has​​​‌ been published at the​ IEEE Geoscience and Remote​‌ Sensing Letter journal (IEEE)​​ 15.

8.2.3 Multi-Modal​​​‌ Co-Learning for Earth Observation:​ Enhancing single-modality models via​‌ modality collaboration

Participants: Cássio​​ Fraga Dantas, Roberto​​​‌ Interdonato, Dino Ienco​.

Collaborators: Francisco​‌ Mena Toro (DFKI -​​ Deutsches Forschungsinstitut für Künstliche​​​‌ Intelligenz), Andreas Dengel (DFKI​ - Deutsches Forschungsinstitut für​‌ Künstliche Intelligenz)

Keywords:​​ Multi-sensor model, Missing sensor​​​‌ data, Deep Learning.

Multi-modal​ co-learning is emerging as​‌ an effective paradigm in​​ machine learning, enabling models​​​‌ to collaboratively learn from​ different modalities to enhance​‌ single-modality predictions. Earth Observation​​ (EO) represents a quintessential​​​‌ domain for multi-modal data​ analysis, wherein diverse remote​‌ sensors collect data to​​ sense our planet. This​​​‌ unprecedented volume of data​ introduces novel challenges. Specifically,​‌ the access to the​​ same sensor modalities at​​​‌ both training and inference​ stages becomes increasingly complex​‌ based on real-world constraints​​ affecting remote sensing platforms.​​​‌ In this context, multi-modal​ co-learning presents a promising​‌ strategy to leverage the​​ vast amount of sensor-derived​​​‌ data available at the​ training stage to improve​‌ single-modality models for inference-time​​ deployment. Most current research​​ efforts focus on designing​​​‌ customized solutions for either‌ particular downstream tasks or‌​‌ specific modalities available at​​ the inference stage. To​​​‌ address this, we have‌ proposed a novel multi-modal‌​‌ co-learning framework capable of​​ generalizing across various tasks​​​‌ without targeting a specific‌ modality for inference. A‌​‌ visual sketch of the​​ proposed framework is depicted​​​‌ in Figure 2.‌ Our approach combines contrastive‌​‌ and modality discriminative learning​​ together to guide single-modality​​​‌ models to structure the‌ internal model manifold into‌​‌ modality-shared and modality-specific information.​​ We evaluate our framework​​​‌ on four EO benchmarks‌ spanning classification and regression‌​‌ tasks across different sensor​​ modalities, where only one​​​‌ of the modalities available‌ during training is accessible‌​‌ at inference time. Our​​ results demonstrate consistent predictive​​​‌ improvements over state-of-the-art approaches‌ from the recent machine‌​‌ learning and computer vision​​ literature, as well as​​​‌ EO-specific methods. The obtained‌ findings validate our framework‌​‌ in the single-modality inference​​ scenarios across a diverse​​​‌ range of EO applications.‌

Figure 2

The image displays a‌​‌ schema overview of a​​ multi-modal sensor data processing​​​‌ system. It shows two‌ sensor modalities: Sensor 1‌​‌ (optical image) and Sensor​​ 2 (radar satellite image​​​‌ time series (SITS)). Each‌ sensor has two encoders:‌​‌ one for unique information​​ and one for common​​​‌ information. These encoders extract‌ specific, shared, and unused‌​‌ features represented by distinct​​ colors. The extracted features​​​‌ are then fed into‌ prediction heads for each‌​‌ sensor. Various losses (main,​​ auxiliary, contrastive, modality discrimination,​​​‌ and orthogonality) are applied‌ to optimize the processing‌​‌ pipeline, ensuring accurate and​​ discriminative feature extraction. The​​​‌ system aims to combine‌ data from different sensors‌​‌ effectively for improved predictions.​​ (Description generated at January​​​‌ 22nd, 2026 by Albert‌ AI with the model‌​‌ Mistral-Small-3.2-24B)

Figure 2:​​ Illustration of MDiCo framework​​​‌ with shared, specific, and‌ unused features. The main‌​‌ prediction per modality is​​ shown in regular arrows,​​​‌ while the multiple losses‌ are shown with dashed‌​‌ lines. Two sensor modalities​​ are shown: an optical​​​‌ image and a radar‌ Satellite Image Time Series‌​‌ (SITS).

This work has​​ been published at the​​​‌ Machine Learning journal (Springer)‌ 18.

8.2.4 Revisiting‌​‌ Cross-Modal Knowledge Distillation: A​​ Disentanglement Approach for RGBD​​​‌ Semantic Segmentation

Participants: Cássio‌ Fraga Dantas, Dino‌​‌ Ienco.

Collaborators:​​ Roger Ferrod (UNITO -​​​‌ Università degli studi di‌ Torino = University of‌​‌ Turin), Luigi di Caro​​ (UNITO - Università degli​​​‌ studi di Torino =‌ University of Turin)

Keywords‌​‌: Knowledge Distillation, Cross-modal​​ learning, RGBD data, Semantic​​​‌ Segmentation.

Multi-modal RGB and‌ Depth (RGBD) data are‌​‌ predominant in many domains​​ such as robotics, autonomous​​​‌ driving and remote sensing.‌ The combination of these‌​‌ multi-modal data enhances environmental​​ perception by providing 3D​​​‌ spatial context, which is‌ absent in standard RGB‌​‌ images. Although RGBD multi-modal​​ data can be available​​​‌ to train computer vision‌ models, accessing all sensor‌​‌ modalities during the inference​​ stage may be infeasible​​​‌ due to sensor failures‌ or resource constraints, leading‌​‌ to a mismatch between​​ data modalities available during​​​‌ training and inference. Traditional‌ Cross-Modal Knowledge Distillation (CMKD)‌​‌ frameworks, developed to address​​​‌ this task, are typically​ based on a teacher/student​‌ paradigm, where a multi-modal​​ teacher distills knowledge into​​​‌ a single-modality student model.​ However, these approaches face​‌ challenges in teacher architecture​​ choices and distillation process​​​‌ selection, thus limiting their​ adoption in real-world scenarios.​‌ To overcome these issues,​​ we introduce CroDiNo-KD (Cross-Modal​​​‌ Disentanglement: a New Outlook​ on Knowledge Distillation), a​‌ novel cross-modal knowledge distillation​​ framework for RGBD semantic​​​‌ segmentation (Figure 3).​ Our approach simultaneously learns​‌ single modality RGB and​​ Depth models by exploiting​​​‌ disentanglement representation, contrastive learning​ and decoupled data augmentation​‌ with the aim to​​ structure the internal manifolds​​​‌ of neural network models​ through interaction and collaboration.​‌ We evaluated CroDiNo-KD on​​ three RGBD datasets across​​​‌ diverse domains, considering recent​ CMKD frameworks as competitors.​‌ Our findings illustrate the​​ quality of CroDiNo-KD, and​​​‌ they suggest reconsidering the​ conventional teacher/student paradigm to​‌ distill information from multi-modal​​ data to single modality​​​‌ neural networks. Source code​ is available here.

Figure 3

The​‌ image shows a schematic​​ of a dual-pathway neural​​​‌ network model that processes​ RGB and depth images.​‌ RGB images are encoded​​ by an RGB encoder​​​‌ and decoded by an​ RGB decoder, while depth​‌ images are encoded by​​ a depth encoder and​​​‌ decoded by a depth​ decoder. A mix-up process​‌ combines features from both​​ pathways using orthogonal constraints​​​‌ to separate invariant and​ specific features for each​‌ modality. Contrastive learning is​​ applied to ensure that​​​‌ the representations are distinct.​ An auxiliary decoder further​‌ processes the invariant features.​​ The model aims to​​​‌ learn complementary information from​ RGB and depth inputs​‌ for tasks like semantic​​ segmentation. (Description generated at​​​‌ January 22nd, 2026 by​ Albert AI with the​‌ model Mistral-Small-3.2-24B)

Figure 3​​: Overview of the​​​‌ CroDiNo-KD architecture, composed by​ two encoder-decoder models, for​‌ both RGB and Depth​​ modalities. In addition an​​​‌ auxiliary decoder and a​ set of loss functions​‌ are adopted to enforce​​ the desired disentanglement properties​​​‌ between modalities, i.e., modality-invariant​ and modality-specific features for​‌ both RGB and Depth​​ information.

This work has​​​‌ been published at the​ European Conference on Machine​‌ Learning and Principles and​​ Practice of Knowledge Discovery​​​‌ in Databases (ECML/PKDD) 2025​ 26.

8.2.5 Multi-sensor​‌ Model for Earth Observation​​ Robust to Missing Data​​​‌ via Sensor Dropout and​ Mutual Distillation

Participants: Cássio​‌ Fraga Dantas, Roberto​​ Interdonato, Dino Ienco​​​‌.

Collaborators: Francisco​ Mena Toro (DFKI -​‌ Deutsches Forschungsinstitut für Künstliche​​ Intelligenz), Andreas Dengel (DFKI​​​‌ - Deutsches Forschungsinstitut für​ Künstliche Intelligenz)

Keywords:​‌ Multi-sensor model, Missing sensor​​ data, Deep Learning.

Multi-sensor​​​‌ data has become a​ foundation of Earth Observation​‌ (EO) research, offering models​​ with enhanced accuracy via​​​‌ optimal fusion strategies. However,​ the unavailability of sensor​‌ data at the regional​​ or country scale during​​​‌ inference can significantly undermine​ model performance. The literature​‌ explores diverse approaches to​​ increasing model robustness to​​​‌ missing sensor scenarios, i.e.,​ to reducing the decline​‌ in accuracy caused by​​ missing data at inference​​​‌ time. Nevertheless, most of​ them have suboptimal behavior​‌ when a single-sensor is​​ available for prediction. To​​ address this challenge, we​​​‌ propose a novel method‌ for multi-sensor modeling, Decision-level‌​‌ Sensor Dropout with mutual​​ distillation, named DSensD+ (Figure​​​‌ 4). This employs‌ a decision-level fusion, ignoring‌​‌ predictions from missing sensors​​ and incorporating the Sensor​​​‌ Dropout (SensD) technique. Unlike‌ works that use the‌​‌ SensD at the input​​ or feature level, we​​​‌ use it at the‌ decision level. Moreover, we‌​‌ include a mutual distillation​​ strategy to improve the​​​‌ robustness. From a practical‌ viewpoint, the additional components‌​‌ in the DSensD+ method​​ are incorporated only for​​​‌ the training phase. During‌ inference, it operates as‌​‌ a standard decision-level fusion​​ model that ignores missing​​​‌ sensors.We validate our method‌ on three EO datasets,‌​‌ spanning binary, multi-class, and​​ multi-label classification tasks for​​​‌ crop- and tree-mapping related‌ applications. Notably, DSensD+ outperforms‌​‌ several state-of-the-art methods, achieving​​ consistent improvements across moderate​​​‌ (single-sensor missing) and extreme‌ (single-sensor available) conditions, as‌​‌ well as with full-sensor​​ data. These results demonstrate​​​‌ the robustness of DSensD+‌ and highlight the effectiveness‌​‌ of our method for​​ the missing sensor problem,​​​‌ advancing the field of‌ multi-sensor modeling in EO.‌​‌

Figure 4

The image depicts a​​ machine learning framework using​​​‌ multiple sensors (s1, s2,‌ ..., s). Each sensor‌​‌ has a dedicated model​​ that generates predicted class​​​‌ logits (C1, C2, C3).‌ These logits are normalized‌​‌ and fused, with mutual​​ distillation losses calculated between​​​‌ each pair of models.‌ The fused logits are‌​‌ compared to ground truth​​ labels to compute full-sensor​​​‌ data loss. Additionally, a‌ missing sensor data loss‌​‌ is calculated for robustness.​​ The framework includes components​​​‌ for sensor dropout, indicating‌ a training strategy to‌​‌ handle missing data scenarios.​​ (Description generated at January​​​‌ 22nd, 2026 by Albert‌ AI with the model‌​‌ Mistral-Small-3.2-24B)

Figure 4:​​ Illustration of the DSensD+​​​‌ model during training, where‌ the fusion, sensor dropout,‌​‌ and mutual distillation occur​​ at the decision-level. During​​​‌ inference, the forward pass‌ of the model corresponds‌​‌ to only the dark​​ arrows.

This work has​​​‌ been published at the‌ IEEE ACCESS journal (IEEE)‌​‌ 19.

8.2.6 Transfer​​ Land Cover Maps Across​​​‌ Years: A Time Series-based‌ Semantic Segmentation Approach

Participants:‌​‌ Cássio Fraga Dantas,​​ Roberto Interdonato, Dino​​​‌ Ienco.

Collaborators:‌ Christopher Jabea (UMR TETIS,‌​‌ INRAE, France), Flavie Cernasson​​ (UMR TETIS, AgroParisTech, France),​​​‌ Eric Barbe (UMR TETIS,‌ INRAE, France), Nadia Guiffant‌​‌ (UMR TETIS, INRAE, France),​​ Christiane Weber (UMR TETIS,​​​‌ CNRS, France)

Keywords:‌ Transfer learning, Satellite image‌​‌ time series, Semantic Segmentation.​​

The widespread availability of​​​‌ satellite imagery data has‌ enabled advancements in Land‌​‌ Use/Land Cover (LULC) and​​ Urban Fabric (UF) mapping​​​‌ through deep learning. However,‌ maintaining up-to-date urban land‌​‌ cover maps is challenged​​ by the high cost​​​‌ and operational constraints of‌ continuous field data collection.‌​‌ This study explores the​​ feasibility of updating urban​​​‌ LULC maps using SITS-based‌ semantic segmentation models trained‌​‌ on historical data, specifically​​ examining a transfer scenario​​​‌ where a model trained‌ on 2015 data is‌​‌ applied to 2020 imagery.​​ We benchmark the performance​​​‌ of two convolution-based architectures‌ (Unet and Unet3D), plus‌​‌ a recent spatiotemporal transformer-based​​​‌ approach (TSViT) and a​ proposed variant, named TSViT+SW,​‌ which incorporates a shifted​​ window attention scheme. Experimental​​​‌ evaluations covering the urban​ area of Lyon, France,​‌ reveal that the proposed​​ TSViT+SW model achieves the​​​‌ best results among transferred​ models, minimizing performance degradation​‌ compared to the ideal​​ in-year training scenario. This​​​‌ work offers insights into​ the potential and limitations​‌ of using historical data​​ to update urban land​​​‌ cover in the absence​ of fresh labeled data.​‌

This work has been​​ published at the Joint​​​‌ Urban Remote Sensing Event​ (JURSE) 2025 27.​‌

8.2.7 SenCLIP: Enhancing zero-shot​​ land-use mapping for Sentinel-2​​​‌ with ground-level prompting

Participants:​ Diego Marcos, Roberto​‌ Interdonato, Dino Ienco​​.

Collaborators: Pallavi​​​‌ Jain (CIHEAM-IAMM - Centre​ International de Hautes Etudes​‌ Agronomiques Méditerranéennes - Institut​​ Agronomique Méditerranéen de Montpellier),​​​‌ Tristan Berchoux (CIHEAM-IAMM -​ Centre International de Hautes​‌ Etudes Agronomiques Méditerranéennes -​​ Institut Agronomique Méditerranéen de​​​‌ Montpellier)

Keywords: Visual​ Language Model, Remote Sensing​‌ Analysis, Zero shot prompting.​​

Pre-trained vision language models​​​‌ (VLMs) like CLIP have​ demonstrated impressive zero-shot classification​‌ capabilities based on free-form​​ prompts, showing some generalization​​​‌ capabilities even in specialized​ domains. However, these models​‌ face limitations when applied​​ to satellite imagery due​​​‌ to the relatively low​ prevalence of such data​‌ in their training datasets,​​ compared to ground level​​​‌ and natural images. Furthermore,​ due to the nature​‌ of image-text associations found​​ in these datasets, current​​​‌ prompting techniques are limited​ to simplistic overhead view​‌ prompts like "a satellite​​ image of...", limiting​​​‌ their applicability for zero-shot​ land-use/land-cover mapping. To address​‌ these challenges, we create​​ a large dataset of​​​‌ satellite (Sentinel-2) and geotagged​ ground-level images across the​‌ whole European Union to​​ transfer the richer ground​​​‌ level representation in the​ CLIP representation to satellite​‌ imagery. This dataset enables​​ a representation of satellite​​​‌ images that captures ground​ level concepts and allows​‌ using rich, ground level​​ perspective prompts (Figure 5​​​‌). We explore prompt​ style variations from both​‌ satellite and ground level​​ views. Our approach results​​​‌ in a substantial improvement​ on zero-shot land use/land​‌ cover classification on two​​ Sentinel-2 benchmarks, EuroSAT and​​​‌ BigEarthNet, compared to directly​ using CLIP or specialized​‌ remote sensing VLMs, opening​​ the doors to zero-shot​​​‌ land-use/land-cover mapping by using​ free-form textual descriptions.

Figure 5

The​‌ image depicts a framework​​ for training and selecting​​​‌ prompts for satellite image​ analysis using machine learning​‌ models. It involves three​​ main steps: Training (1),​​​‌ Prompt Selection (2), and​ Zero-shot Evaluation (3). In​‌ the training phase, both​​ ground-level and satellite images​​​‌ are processed by CLIP​ models to generate embeddings.​‌ These are stored in​​ a memory queue. In​​​‌ the prompt selection phase,​ diverse text prompts are​‌ generated by a language​​ model and evaluated for​​​‌ their similarity to the​ image embeddings. The top​‌ performing prompts are selected.​​ In the zero-shot evaluation​​​‌ phase, these selected prompts​ are used to make​‌ predictions about new images.​​ (Description generated at January​​​‌ 22nd, 2026 by Albert​ AI with the model​‌ Mistral-Small-3.2-24B)

Figure 5:​​ Architecture: The figure illustrates​​ the three-step pipeline consisting​​​‌ of Pre-Training, Prompt Selection,‌ and Zero-shot Predictions. It‌​‌ also demonstrates the prompt​​ generation process from LLMs,​​​‌ which is utilized for‌ prompt selection and then‌​‌ selected prompts for zero-shot​​ prediction.

This work has​​​‌ been published at the‌ IEEE/CVF Winter Conference on‌​‌ Applications of Computer Vision​​ (WACV) 2025 28.​​​‌

8.2.8 MARA: a deep‌ learning based framework for‌​‌ multilayer graph simplification

Participants:​​ Roberto Interdonato, Dino​​​‌ Ienco.

Collaborators:‌ Cheick Tidiane Ba (Queen‌​‌ Mary University of London,​​ London, United Kingdom), Sabrina​​​‌ Gaito (Department of Computer‌ Science, University of Milan,‌​‌ Milan, Italy)

Keywords:​​ Graph Neural Network, Graph​​​‌ Simplification, Multilayer Graph.

In‌ many scientific fields, complex‌​‌ systems are characterized by​​ a multitude of heterogeneous​​​‌ interactions/relationships that are challenging‌ to model. Multilayer graphs‌​‌ constitute valuable tools that​​ can represent such complex​​​‌ systems, thus making possible‌ their analysis for downstream‌​‌ decision-making processes. Nevertheless, modeling​​ such complex information still​​​‌ remains challenging in real-world‌ scenarios. On the one‌​‌ hand, holistically including all​​ relationships may lead to​​​‌ noisy or computationally intensive‌ graphs. On the other‌​‌ hand, limiting the amount​​ of information to model​​​‌ through the selection of‌ a portion of the‌​‌ available relationships can introduce​​ boundary specification biases. However,​​​‌ the current research studies‌ are demonstrating that it‌​‌ is more beneficial to​​ retain as much information​​​‌ as possible and at‌ a later stage perform‌​‌ graph simplification i.e., removing​​ uninformative or redundant parts​​​‌ of the graph to‌ facilitate the final analysis.‌​‌ While simplification strategies, based​​ on deep learning methods,​​​‌ have been already extensively‌ explored in the context‌​‌ of single-layer graphs, only​​ a limited amount of​​​‌ efforts have been devoted‌ to simplification strategies for‌​‌ multilayer graphs. In this​​ work, we have proposed​​​‌ the MultilAyer gRaph simplificAtion‌ (MARA) framework, a Graph‌​‌ Neural Network (GNN) based​​ approach designed to simplify​​​‌ multilayer graphs based on‌ the downstream task. MARA‌​‌ generates node embeddings for​​ a specific task by​​​‌ training jointly two main‌ components: i) an edge‌​‌ simplification module and ii)​​ a (multilayer) graph neural​​​‌ network. We tested MARA‌ on different real-world multilayer‌​‌ graphs for node classification​​ tasks (Figure 6).​​​‌ Experimental results show the‌ effectiveness of the proposed‌​‌ approach: MARA reduces the​​ dimension of the input​​​‌ graph while keeping and‌ even improving the performance‌​‌ of node classification tasks​​ in different domains and​​​‌ across graphs characterized by‌ different structures. Moreover, deep‌​‌ learning-based simplification allows MARA​​ to preserve and enhance​​​‌ important graph properties for‌ the downstream task. To‌​‌ our knowledge, MARA represents​​ the first simplification framework​​​‌ especially tailored for multilayer‌ graphs analysis .

Figure 6

The‌​‌ image depicts a flowchart​​ of a graph neural​​​‌ network (GNN) process. It‌ starts with two layered‌​‌ graphs that are simplified​​ by a module, resulting​​​‌ in a single graph.‌ This simplified graph undergoes‌​‌ processing through GNN layers,​​ leading to a final​​​‌ prediction. The process involves‌ simplifying complex graph structures,‌​‌ applying graph neural network​​ layers to extract information,​​​‌ and making predictions based‌ on the processed data.‌​‌ The nodes and edges​​​‌ in the graphs represent​ data points and their​‌ relationships, which are refined​​ and analyzed sequentially. (Description​​​‌ generated at January 22nd,​ 2026 by Albert AI​‌ with the model Mistral-Small-3.2-24B)​​

Figure 6: Multilayer​​​‌ graph simplification with multilayer​ GNN. A multilayer simplification​‌ module detects the links​​ to remove by taking​​​‌ into account the whole​ input multilayer graph, while​‌ a GNN is used​​ to generate node embeddings​​​‌ for a downstream task.​

This work has been​‌ published at the Neurocomputing​​ journal (Elsevier) 9.​​​‌

8.2.9 Hybrid phenology modeling​ for predicting temperature effects​‌ on tree dormancy

Participants:​​ Diego Marcos.

Collaborators​​​‌: Ron van Bree​ (Wageningen University, The Netherlands),​‌ Ioannis Athanasiadis (Wageningen University,​​ The Netherlands)

Keywords:​​​‌ Plant phenology, hybrid models.​

Biophysical models offer valuable​‌ insights into climate-phenology relationships​​ in both natural and​​​‌ agricultural settings. However, there​ are substantial structural discrepancies​‌ across models which require​​ site-specific recalibration, often yielding​​​‌ inconsistent predictions under similar​ climate scenarios. Machine learning​‌ methods offer data-driven solutions,​​ but often lack interpretability​​​‌ and alignment with existing​ knowledge. We present a​‌ phenology model describing dormancy​​ in fruit trees, integrating​​​‌ conventional biophysical models with​ a neural network to​‌ address their structural disparities.​​ We evaluate our hybrid​​​‌ model in an extensive​ case study predicting cherry​‌ tree phenology in Japan,​​ South Korea and Switzerland.​​​‌ Our approach consistently outperforms​ both traditional biophysical and​‌ machine learning models in​​ predicting blooming dates across​​​‌ years. Additionally, the neural​ network’s adaptability facilitates parameter​‌ learning for specific tree​​ varieties, enabling robust generalization​​​‌ to new sites without​ site-specific recalibration. This hybrid​‌ model leverages both biophysical​​ constraints and data-driven flexibility,​​​‌ offering a promising avenue​ for accurate and interpretable​‌ phenology modeling.

This work​​ has been presented in​​​‌ the AAAI conference 24​.

8.2.10 EcoWikiRS: Learning​‌ Ecological Representation of Satellite​​ Images from Weak Supervision​​​‌ with Species Observations and​ Wikipedia

Participants: Diego Marcos​‌.

Collaborators: Valerie​​ Zermatten (EPFL, Switzerland), Javiera​​​‌ Castillo-Navarro (EPFL, Switzerland), Devis​ Tuia (EPFL, Switzerland)

Keywords​‌: Biodiversity, vision-language models,​​ aerial imagery, Wikipedia.

The​​​‌ presence of species provides​ key insights into the​‌ ecological properties of a​​ location such as land​​​‌ cover, climatic conditions or​ even soil properties. We​‌ propose a method to​​ predict such ecological properties​​​‌ directly from remote sensing​ (RS) images by aligning​‌ them with species habitat​​ descriptions. We introduce the​​​‌ EcoWikiRS dataset, consisting of​ high-resolution aerial images, the​‌ corresponding geolocated species observations,​​ and, for each species,​​​‌ the textual descriptions of​ their habitat from Wikipedia.​‌ EcoWikiRS offers a scalable​​ way of supervision for​​​‌ RS vision language models​ (RS-VLMs) for ecology. This​‌ is a setting with​​ weak and noisy supervision,​​​‌ where, for instance, some​ text may describe properties​‌ that are specific only​​ to part of the​​​‌ species' niche or is​ irrelevant to a specific​‌ image. We tackle this​​ by proposing WINCEL, a​​​‌ weighted version of the​ InfoNCE loss. We evaluate​‌ our model on the​​ task of ecosystem zero-shot​​​‌ classification by following the​ habitat definitions from the​‌ European Nature Information System​​ (EUNIS). Our results show​​ that our approach helps​​​‌ in understanding RS images‌ in a more ecologically‌​‌ meaningful manner.

This work​​ has been presented in​​​‌ the CVPR EarthVision workshop‌ 30.

8.2.11 Atomizer:‌​‌ Generalizing to unseen modalities​​ by breaking images down​​​‌ to a set of‌ scalars

Participants: Roberto Interdonato‌​‌, Diego Marcos.​​

Collaborators: Sylvain Lobry​​​‌ (Université de Paris Cité,‌ France)

Keywords: Multi-modal‌​‌ remote sensing, transformers, foundation​​ models.

The growing number​​​‌ of Earth observation satellites‌ has led to increasingly‌​‌ diverse remote sensing data,​​ with varying spatial, spectral,​​​‌ and temporal configurations. Most‌ existing models rely on‌​‌ fixed input formats and​​ modality-specific encoders, which require​​​‌ retraining when new configurations‌ are introduced, limiting their‌​‌ ability to generalize across​​ modalities. We introduce Atomizer,​​​‌ a flexible architecture that‌ represents remote sensing images‌​‌ as sets of scalars,​​ each corresponding to a​​​‌ spectral band value of‌ a pixel. Each scalar‌​‌ is enriched with contextual​​ metadata (acquisition time, spatial​​​‌ resolution, wavelength, and bandwidth),‌ producing an atomic representation‌​‌ that allows a single​​ encoder to process arbitrary​​​‌ modalities without interpolation or‌ resampling. Atomizer uses structured‌​‌ tokenization with Fourier features​​ and non-uniform radial basis​​​‌ functions to encode content‌ and context, and maps‌​‌ tokens into a latent​​ space via cross-attention. Under​​​‌ modality-disjoint evaluations, Atomizer outperforms‌ standard models and demonstrates‌​‌ robust performance across varying​​ resolutions and spatial sizes.​​​‌

This work has been‌ presented in the British‌​‌ Machine Vision Conference 29​​, where it was​​​‌ awarded the Best Presentation‌ Award.

8.2.12 Predicting Near-future‌​‌ Deforestation Across the Tropics​​ Using Deep Learning: Insights​​​‌ from the Forest Foresight‌ Project

Participants: Diego Marcos‌​‌.

Collaborators: Laura​​ Elena Cué La Rosa​​​‌ (Wageningen University, The Netherlands),‌ Jonas van Duijvenbode (WWF),‌​‌ Zillah Calle (Wageningen University,​​ The Netherlands), Jorn Dallinga​​​‌ (WWF) and Johannes Reiche‌ (Wageningen University, The Netherlands)‌​‌

Keywords: Deforestation forecasting,​​ tropical forests

Tropical deforestation​​​‌ continues to threaten biodiversity,‌ carbon storage, and climate‌​‌ regulation. While satellite-based near-real-time​​ monitoring systems track deforestation​​​‌ across local to global‌ scales, they only detect‌​‌ forest loss after it​​ occurs. Proactively identifying areas​​​‌ at risk can help‌ support timely mitigation efforts.‌​‌ As part of the​​ Forest Foresight initiative led​​​‌ by the World Wide‌ Fund for Nature (Netherlands),‌​‌ we developed a deep​​ learning model to predict​​​‌ near-future deforestation risk. The‌ model produces monthly risk‌​‌ maps at a 400-meter​​ spatial resolution with a​​​‌ six-month prediction horizon, based‌ on near real-time deforestation‌​‌ alerts and various open-access​​ geospatial datasets used as​​​‌ predictors. It was tested‌ in 17 countries across‌​‌ humid tropical forests in​​ South America, Africa, and​​​‌ Southeast Asia, and achieved‌ an average F0.5 score‌​‌ of 64.8%. It demonstrates​​ a modest performance gain​​​‌ over both a rule-based‌ baseline model (4% global‌​‌ improvement) and an XGBoost​​ decision forest model (1.4%​​​‌ improvement globally), as well‌ as greater temporal and‌​‌ cross-country prediction stability. Prediction​​ performance was highest in​​​‌ areas near recent deforestation‌ and declined with increasing‌​‌ distance from these areas,​​ highlighting the model's dependence​​​‌ on past deforestation patterns‌ as the main limitation.‌​‌ The model is less​​​‌ effective in predicting new​ deforestation events in previously​‌ undisturbed forests, such as​​ regions where new logging​​​‌ roads are being developed​ after the prediction date.​‌ To substantially improve prediction​​ performance, it is essential​​​‌ to integrate frequently updated,​ region-specific, and novel data​‌ sources, particularly real-time indicators​​ of human activity, such​​​‌ as mobile phone movements​ or economic signals.

This​‌ work has been accepted​​ at Environmental Research Communications​​​‌ 10.

8.2.13 Learning​ transferable land cover semantics​‌ for open vocabulary interactions​​ with remote sensing images​​​‌

Participants: Diego Marcos.​

Collaborators: Valerie Zermatten​‌ (EPFL, Switzerland), Javiera Castillo-Navarro​​ (EPFL, Switzerland), Devis Tuia​​​‌ (EPFL, Switzerland)

Keywords:​ Biodiversity, vision-language models, aerial​‌ imagery, Semantic Segmentation.

Why​​ should we confine land​​​‌ cover classes to rigid​ and arbitrary definitions? Land​‌ cover mapping is a​​ central task in remote​​​‌ sensing image processing, but​ the rigorous class definitions​‌ can sometimes restrict the​​ transferability of annotations between​​​‌ datasets. Open vocabulary recognition,​ i.e. using natural language​‌ to define a specific​​ object or pattern in​​​‌ an image, breaks free​ from predefined nomenclature and​‌ offers flexible recognition of​​ diverse categories with a​​​‌ more general image understanding​ across datasets and labels.​‌ The open vocabulary framework​​ opens doors to search​​​‌ for concepts of interest,​ beyond individual class boundaries.​‌ In this work, we​​ propose to use Text​​​‌ As supervision for COntrastive​ Semantic Segmentation (TACOSS), and​‌ we design an open​​ vocabulary semantic segmentation model​​​‌ that extends its capacities​ beyond that of a​‌ traditional model for land​​ cover mapping: In addition​​​‌ to visual pattern recognition,​ TACOSS leverages the common​‌ sense knowledge captured by​​ language models and is​​​‌ capable of interpreting the​ image at the pixel​‌ level, attributing semantics to​​ each pixel and removing​​​‌ the constraints of a​ fixed set of land​‌ cover labels. By learning​​ to match visual representations​​​‌ with text embeddings, TACOSS​ can transition smoothly from​‌ one set of labels​​ to another and enables​​​‌ the interaction with remote​ sensing images in natural​‌ language. Our approach combines​​ a pretrained text encoder​​​‌ with a visual encoder​ and adopts supervised contrastive​‌ learning to align the​​ visual and textual modalities.​​​‌ We explore several text​ encoders and label representation​‌ methods and compare their​​ abilities to encode transferable​​​‌ land cover semantics. The​ model’s capacity to predict​‌ a set of different​​ land cover labels on​​​‌ an unseen dataset is​ also explored to illustrate​‌ the generalization capacities across​​ domains of our approach.​​​‌ Overall, TACOSS is a​ general method and permits​‌ adapting between different sets​​ of land cover labels​​​‌ with minimal computational overhead.​

This work has been​‌ presented in the ISPRS​​ Journal of Photogrammetry and​​​‌ Remote Sensing 23.​

8.3 Interaction between Domain​‌ expert and Machine Learning​​ models

8.3.1 Evaluation of​​​‌ geographical distortions in language​ models

Participants: Roberto Interdonato​‌.

Collaborators: Rémy​​ Decoupes (UMR TETIS, INRAE,​​​‌ France), Mathieu Roche (UMR​ TETIS, CIRAD, France), Maguelonne​‌ Teisseire (UMR TETIS, INRAE,​​ France), Sarah Valentin (UMR​​​‌ TETIS, CIRAD, France)

Keywords​: NLP, LLM, Spatial​‌ information, Bias.

Geographic bias​​ in language models (LMs)​​ is an underexplored dimension​​​‌ of model fairness, despite‌ growing attention being given‌​‌ to other social biases.​​ We investigate whether LMs​​​‌ provide equally accurate representations‌ across all global regions‌​‌ and propose a benchmark​​ of four indicators to​​​‌ detect undertrained and underperforming‌ areas: (i) indirect assessment‌​‌ of geographic training data​​ coverage via tokenizer analysis,​​​‌ (ii) evaluation of basic‌ geographic knowledge, (iii) detection‌​‌ of geographic distortions, and​​ (iv) visualization of performance​​​‌ disparities through maps. Applying‌ this framework to ten‌​‌ widely used encoder- and​​ decoder-based models, we find​​​‌ systematic overrepresentation of Western‌ countries and consistent underrepresentation‌​‌ of several African, Eastern​​ European, and Middle Eastern​​​‌ regions, leading to measurable‌ performance gaps. We further‌​‌ analyze the impact of​​ these biases on downstream​​​‌ tasks, particularly in crisis‌ response, and show that‌​‌ regions most vulnerable to​​ natural disasters are often​​​‌ those with poorer LM‌ coverage. Our findings underscore‌​‌ the need for geographically​​ balanced LMs to ensure​​​‌ equitable and effective global‌ applications.

This work has‌​‌ been published in the​​ Discovery Science conference 25​​​‌ and succcessively extended in‌ a publication on the‌​‌ Machine Learning journal (Springer​​ Nature) 12.

9​​​‌ Bilateral contracts and grants‌ with industry

9.1 Bilateral‌​‌ contracts with industry

  • ATOS​​ - Cifre thesis

    Participants:​​​‌ Raffaele Gaetano, Diego‌ Marcos, Dino Ienco‌​‌.

    This Cifre Ph.D.​​ thesis project, entitled “Multi-source​​​‌ satellite image segmentation for‌ the extraction of geometric‌​‌ landscape objects with an​​ application to the extraction​​​‌ of agricultural land parcels”,‌ started in September 2023,‌​‌ for a total duration​​ of 3 years.

    Context​​​‌: Delineating agricultural field‌ plot accurately and efficiently‌​‌ is important not only​​ for the declaration-based subsidy​​​‌ systems such as the‌ European Common Agricultural Policy,‌​‌ but also for monitoring​​ agricultural activities on several​​​‌ scales (environmental impact, territorial‌ development, crop monitoring and‌​‌ precision farming, etc.) and​​ get useful information regarding​​​‌ the status of agricultural‌ production. To this end,‌​‌ the necessity of precise​​ and timely spatialized products,​​​‌ such as land use‌ and land cover maps‌​‌ and the estimation of​​ agricultural yields at field​​​‌ level, are essential. These‌ tools are part of‌​‌ a process of developing​​ value-added services linked to​​​‌ digital agriculture. The accuracy‌ and freshness of these‌​‌ products could prove to​​ be a key factor​​​‌ in supporting decision-making by‌ a wide range of‌​‌ stakeholders, including farmers, land​​ managers and political decision-makers.​​​‌

    Objectives: Initial work‌ on the extraction of‌​‌ agricultural land parcels from​​ satellite imagery using deep​​​‌ learning techniques has recently‌ been proposed but these‌​‌ are mainly studies that​​ directly deploy techniques from​​​‌ the state of the‌ art in computer vision‌​‌ in this field of​​ application, and therefore with​​​‌ a limited adaptation to‌ the field of satellite‌​‌ imagery, particularly with regard​​ to taking into account​​​‌ the multi-source, multi-temporal and‌ multi-scale information that is‌​‌ accessible via modern Earth​​ observation missions. It is​​​‌ in this context that‌ this CIFRE thesis aims‌​‌ to tackle the problem​​ of the automatic extraction​​​‌ of agricultural fields from‌ remotely sensed data on‌​‌ a territory and its​​​‌ characterization in terms of​ land use and land​‌ cover. To this end,​​ the thesis project plans​​​‌ to leverage deep learning​ techniques such as semantic​‌ segmentation and instance segmentation​​ to propose new approaches​​​‌ tailored to the analysis​ of satellite data for​‌ the task of extracting​​ geometric contours for the​​​‌ delineation of agricultural fields,​ as well as for​‌ characterizing the corresponding land​​ use (in terms of​​​‌ cropping practices).

10 Partnerships​ and cooperations

10.1 International​‌ initiatives

  • OBSYDYA

    Participants: Raffaele​​ Gaetano, Roberto Interdonato​​​‌.

    Official website of​ the OBSYDYA project

    Title​‌ : Observatoire Pilote des​​ Paysages et des Dynamiques​​​‌ Agricoles du Bénin (Pilot​ Observatory of Landscapes and​‌ Agricultural Dynamics in Benin).​​

    Duration: From January​​​‌ 1, 2022 to December​ 31, 2026.

    Summary:​‌ The OBSYDYA project, Observatoire​​ Pilote des Paysages et​​​‌ Dynamiques Agricoles (Pilot​ Observatory of Landscapes and​‌ Agricultural Dynamics in Benin​​), is a DeSIRA​​​‌ (Development Smart Innovation through​ Research in Agriculture) project​‌ funded by the European​​ Union. The project aims​​​‌ at taking advantage of​ the possibilities offered by​‌ recent satellite images to​​ monitor changes in the​​​‌ landscape and provide information​ on the agricultural practices​‌ that induce them, in​​ order to finally have​​​‌ more reliable and frequent​ agricultural statistics to guide​‌ agricultural advisory services and​​ infrastructure planning. The overall​​​‌ aim of the project​ is to set up​‌ a pilot observatory of​​ agrarian systems and landscapes,​​​‌ in the form of​ a platform of services​‌ dedicated to the capitalization​​ of spatialized information (maps,​​​‌ satellite images, surveys), the​ production of indicators (regular​‌ and inexpensive) and the​​ mapping of the dynamics​​​‌ of agrarian systems and​ landscapes in Central and​‌ Northern Benin.

    This project​​ funds the PhD of​​​‌ Bruno Bio Nikki.

  • Land​ Matrix Initiative

    Participants: Roberto​‌ Interdonato.

    Official website​​ of the Land Matrix​​​‌ Initiative

    Title : Land​ Matrix Phase IV -​‌ Increased Transparency and Accountability​​ on Land Acquisitions: Indigenous​​​‌ peoples and local communities,​ biodiversity hotspots and new​‌ pressures on land in​​ the context of climate​​​‌ change and sustainable supply​ chains.

    Duration: From​‌ January 1, 2024 to​​ December 31, 2027

    Summary​​​‌: The Land Matrix​ Initiative (LMI) is an​‌ independent global land monitoring​​ initiative, consisting of several​​​‌ global and regional partners,​ including CIRAD. The LMI​‌ was created in 2009​​ to address the lack​​​‌ of robust data on​ Large Scale Land Acquisitions​‌ (LSLAs). The first version​​ of the Land Matrix​​​‌ database was launched in​ April 2012, providing a​‌ systematic overview of large-scale​​ agricultural investments. Today, in​​​‌ addition to a large​ collection of global data​‌ illustrating the magnitude of​​ LSLAs at an international​​​‌ scale, the collection of​ country-specific data is carried​‌ on by the four​​ regional focal points in​​​‌ Africa, Asia, Eastern Europe​ and Latin America, as​‌ well as the national​​ land observatories in Argentina,​​​‌ Cameroon, the Philippines, Senegal​ and Uganda. The project,​‌ funded by IFAD -​​ International Land Coalition (ILC),​​​‌ has recently completed its​ Phase III (Open​‌ data for transparency and​​ accountability on land and​​ investment, 2019-2023) and​​​‌ entered its Phase IV‌ in January 2024, under‌​‌ the theme Increased Transparency​​ and Accountability on Land​​​‌ Acquisitions: Indigenous peoples and‌ local communities, biodiversity hotspots‌​‌ and new pressures on​​ land in the context​​​‌ of climate change and‌ sustainable supply chains.‌​‌

    EVERGREEN participates in this​​ project concerning the modeling​​​‌ and analysis of complex‌ network models issued by‌​‌ data about the global​​ land trade market.

  • CNPq/MCTI/FNDCT​​​‌ (Brazil) Collaboration Network

    Participants:‌ Cássio Fraga Dantas.‌​‌

    Title: Collaboration Network​​ for the Development of​​​‌ Methods in Responsible Artificial‌ Intelligence.

    Duration: From‌​‌ September 1, 2025 to​​ August 31, 2027

    Summary​​​‌: A major challenge‌ in artificial intelligence (AI)‌​‌ is to reconcile all​​ the benefits of its​​​‌ applications with a responsible‌ use of the algorithms.‌​‌ For example, it is​​ well known that, when​​​‌ dealing with historical data,‌ machine learning algorithms can‌​‌ learn (and replicate) biased​​ decisions toward different segments​​​‌ of society, such as‌ race and gender. Furthermore,‌​‌ an important challenge when​​ working with nonlinear models,​​​‌ such as neural networks,‌ is understanding the relationships‌​‌ between the attributes used​​ as model inputs and​​​‌ the predicted outputs. These‌ challenges motivate the problems‌​‌ addressed in this project,​​ in which the central​​​‌ topic is the development‌ of ethical (or responsible)‌​‌ AI solutions. In the​​ first research topic, we​​​‌ shall develop and analyze‌ new machine learning methods‌​‌ for bias mitigation and​​ identification, especially considering the​​​‌ unsupervised paradigm. The second‌ line addresses the development‌​‌ of explainability algorithms capable​​ of providing insights into​​​‌ the functioning of the‌ so-called black-box models. As‌​‌ a major objective, we​​ aim to jointly develop​​​‌ new solutions in these‌ two areas, exploring the‌​‌ synergies between the teams​​ based in Brazil and​​​‌ in the partner institutions‌ abroad. Additionally, aiming to‌​‌ establish a long-term cooperation​​ network, the project also​​​‌ has as an important‌ objective to establish graduate‌​‌ projects to be jointly​​ supervised by a researcher​​​‌ based in Brazil and‌ a researcher based abroad.‌​‌ The project is funded​​ by a CNPq/MCTI/FNDCT call​​​‌ for project Brazil Knowledge‌ Program aimed at fostering‌​‌ and supporting collaborative research​​ networks involving Brazilian researchers​​​‌ abroad. The funded project‌ establishes partnerships with research‌​‌ groups in the state​​ of São Paulo, such​​​‌ as the University of‌ Campinas (with the project‌​‌ lead, Leonardo Tomazeli Duarte)​​ and the Federal University​​​‌ of ABC.

10.2 European‌ initiatives

10.2.1 Horizon Europe‌​‌

  • Participants: Dino Ienco,​​ Cássio Fraga Dantas.​​​‌

    Eco2Adapt

    Eco2Adapt project on‌ cordis.europa.eu

    Title : Ecosystem-based‌​‌ Adaptation and Changemaking to​​ Shape, Protect and Maintain​​​‌ the Resilience of Tomorrow’s‌ Forests

    Duration: From‌​‌ September 1, 2022 to​​ August 31, 2027

    Summary​​​‌: Forests can be‌ destroyed through climatic events‌​‌ such as storms or​​ drought, or attacked by​​​‌ pests and pathogens, leaving‌ a devastated landscape and‌​‌ despairing local populations. The​​ EU-funded eco2adapt project will​​​‌ develop the ecosystem-based adaptation‌ framework derived from nature-based‌​‌ solutions and work in​​ Living Labs located in​​​‌ climate hotspots in Europe‌ and China. It will‌​‌ use an advanced Decision​​​‌ Theatre approach to investigate​ how to integrate disturbance​‌ and vulnerability into forest​​ management by developing changemaking​​​‌ scenarios. Furthermore, it will​ create innovative technical, economic​‌ and governance mechanisms at​​ a regional level, apply​​​‌ semantic technology to establish​ a knowledge base for​‌ hosting FAIR data, create​​ a smartphone application (OneForest​​​‌ ToolBox) to enable users​ to add data on​‌ climate-resilient forests, and provide​​ cutting-edge tools to monitor​​​‌ vulnerability and resilience.

    EVERGREEN​ is involved in Task​‌ 3.3 "Defining and estimating​​ forest ecosystem services" that​​​‌ D. Ienco co-leads. The​ interest for the team​‌ is to advance spatio-temporal​​ analysis for the analysis​​​‌ of forest disturbances at​ large scale with the​‌ exploitation of satellite derived​​ products. We have a​​​‌ post-doctoral researcher for 18​ months funded by this​‌ project.

10.3 National initiatives​​

  • #DigitAg: Digital Agriculture

    Participants:​​​‌ Cássio Fraga Dantas,​ Raffaele Gaetano, Roberto​‌ Interdonato, Diego Marcos​​, Dino Ienco.​​​‌

    Duration: From September​ 2017 to August 2026​‌

    Summary: #DigitAg is​​ a “Convergence Institute” dedicated​​​‌ to the increasing importance​ of digital techniques in​‌ agriculture. Its goal is​​ twofold: First, making innovative​​​‌ research on the use​ of digital techniques in​‌ agriculture in order to​​ improve competitiveness, preserve the​​​‌ environment, and offer correct​ living conditions to farmers.​‌ Second, preparing future farmers​​ and agricultural policy makers​​​‌ to successfully exploit such​ technologies. While #DigitAg is​‌ based in Montpellier, Rennes​​ is a satellite of​​​‌ the institute focused on​ cattle farming.

    EVERGREEN is​‌ involved in the “agricultural​​ territoy management” challenge of​​​‌ the institute, which D.​ Ienco leads. The interest​‌ for the team is​​ to design novel methods​​​‌ and frameworks to analyze​ and manage heterogeneous remote​‌ sensing data for agricultural​​ applications. In 2024, we​​​‌ had an internship funded​ by this initiative.

  • PEPR​‌ IRiMa

    Participants: Cássio Fraga​​ Dantas, Diego Marcos​​​‌, Dino Ienco.​

    Duration: From September​‌ 2023 to August, 2031​​

    Summary: The exploratory​​​‌ IRiMa PEPR aims to​ formalize a "science of​‌ risk" to contribute to​​ the development of a​​​‌ new strategy for the​ management of risks and​‌ disasters and their impacts​​ in a context of​​​‌ global change. To achieve​ this, it implements a​‌ series of research projects​​ and expert assessments (involving​​​‌ observation, analysis or decision​ support) to accelerate the​‌ transition to a society​​ capable of facing a​​​‌ range of threats (hydro-climatic,​ telluric, technological, health-related, coupled),​‌ by adapting and becoming​​ more resilient and sustainable.​​​‌ In order to face​ this challenge, which is​‌ increased by climate change,​​ it is necessary to​​​‌ consolidate, stimulate and coordinate​ the national research effort.​‌

    EVERGREEN is involved in​​ the project Intelligent Mapping,​​​‌ funded by the PEPR​ IRIMA, for the analysis​‌ and management of multi-source​​ remote sensing data for​​​‌ the monitoring of coastal​ areas. To this end,​‌ a post-doctoral researcher, for​​ a duration of 3​​​‌ years, is scheduled in​ the framework of the​‌ Intelligent Mapping project.

  • REUSTIS​​ - TOSCA - CNES​​​‌

    Participants: Cássio Fraga Dantas​, Raffaele Gaetano,​‌ Dino Ienco.

    Title​​: Re-Use of historical​​ ground truth data to​​​‌ improve land cover mapping‌ based on satellite imagery‌​‌ time series

    Duration:​​ From January 2024 to​​​‌ December 2026

    Summary:‌ This proposal explores the‌​‌ possibility of reusing historical​​ ground truth data to​​​‌ improve the land cover‌ mapping process based on‌​‌ single and multi-source satellite​​ imagery time series. To​​​‌ this end, the objectives‌ of this project are:‌​‌ i) to develop and​​ implement a deep learning​​​‌ methodology that can reuse‌ historical ground truth data‌​‌ to improve land cover​​ mapping results by explicitly​​​‌ taking into account differences‌ and/or shifts in the‌​‌ distribution of remote sensing​​ data associated with each​​​‌ field campaign; ii) propose‌ this new methodology both‌​‌ in a single-sensor context​​ (exploiting only Sentinel-2 optical​​​‌ imagery time series) and,‌ in a second phase,‌​‌ extend the proposed approach​​ to the multi-sensor radar/optical​​​‌ context, in particular by‌ targeting the simultaneous exploitation‌​‌ of Sentinel-1 and Sentinel-2​​ time series.

    EVERGREEN is​​​‌ leading this project that‌ is devoted to the‌​‌ reuse of historical ground​​ truth data for the​​​‌ current analysis of remote‌ sensing information for the‌​‌ downstream task of land​​ use and land cover​​​‌ mapping. This project funds‌ several master/engineering internships and‌​‌ publications fees.

10.3.1 ANR​​

  • ANR GLOURB

    Participants: Cássio​​​‌ Fraga Dantas, Roberto‌ Interdonato, Dino Ienco‌​‌.

    Title: Floodplain​​ urbanization at global scale​​​‌

    Duration: From December‌ 2022 to November, 2026‌​‌

    Summary: GloUrb addresses​​ the issue of floodplain​​​‌ urbanization at global scale‌ since the 1980s based‌​‌ on an interdisciplinary and​​ integrated approach. Floodplains are​​​‌ amongst the most threatened‌ and vulnerable ecosystems, but‌​‌ vital for human society.​​ GloUrb builds on the​​​‌ need for global references‌ to support understanding of‌​‌ floodplain urbanization processes and​​ their socio-ecosystem consequences (biodiversity,​​​‌ flood risk, urban resilience,‌ environmental justice); explain such‌​‌ trends and distinguish between​​ local and global scale​​​‌ drivers; inform and increase‌ the public awareness; monitor‌​‌ to prevent threats and​​ manage future changes. GloUrb​​​‌ will use existing local‌ and global information, web‌​‌ data mining, remote sensing​​ and innovative signal analysis​​​‌ techniques. We will inform‌ people on urbanization processes‌​‌ to support sustainable, integrated​​ and adaptive management, developing​​​‌ a global online information‌ system with a monitoring‌​‌ interface showing urbanization trends​​ and targeting potential threats.​​​‌

    EVERGREEN is involved in‌ the project for the‌​‌ analysis of remote sensing​​ time series data with​​​‌ the aim to conceive,‌ design and develop techniques‌​‌ for the downstream task​​ of land use land​​​‌ cover mapping with a‌ focus on how the‌​‌ process can be transferred​​ temporally from one period​​​‌ of time to another‌ period of time.

  • ANR‌​‌ GeoReSeT

    Participants: Roberto Interdonato​​, Diego Marcos,​​​‌ Dino Ienco.

    Title‌: Generalized Earth Observation‌​‌ with Remote Sensing and​​ Text

    Duration: From​​​‌ September 2023 to August,‌ 2027

    Summary: This‌​‌ research proposal aims to​​ develop a versatile foundation​​​‌ model for geo-spatial data‌ that can be used‌​‌ for any task and​​ with any data modality.​​​‌ By using location on‌ the Earth's surface as‌​‌ the common link between​​​‌ different modalities, the model​ will be able to​‌ incorporate a variety of​​ data sources, including remote​​​‌ sensing imagery, textual descriptions​ of places, and features​‌ in maps. Through self-supervised​​ learning methods such as​​​‌ contrastive learning or multi-modal​ masked autoencoders, the model​‌ will leverage the large​​ amounts of unlabeled geo-spatial​​​‌ data from these different​ sources to learn a​‌ better representation of any​​ geo-spatial location and convey​​​‌ a semantic representation of​ the information.

    EVERGREEN is​‌ co-leading the project with​​ D. Marcos as co-PI​​​‌ (principal investigator) of the​ project. The project objectives​‌ are at the core​​ of the research activities​​​‌ of the team.

  • ANR​ PREDISPOSE

    Participants: Dino Ienco​‌.

    Title: Fire​​ prevention: preparing the French​​​‌ regions

    Duration: From​ March 2025 to Februray,​‌ 2029

    Summary: This​​ research project aims to​​​‌ address the escalating risk​ of wildfires in France,​‌ as high-lighted by the​​ 6th IPCC report and​​​‌ corroborated by studies predicting​ an increase in fire​‌ occurrence, particularly in the​​ north and west at​​​‌ an accelerated rate. We​ aim to develop a​‌ comprehensive methodology based on​​ scientific research and tools,​​​‌ some new and some​ adapted, with the overall​‌ objective of improving the​​ prevention of forest fires​​​‌ in forested and semi-natural​ ecosystems.

    EVERGREEN contributes to​‌ the project by analyzing​​ remote sensing data to​​​‌ extract large-scale information about​ vegetation strata. This analysis​‌ supports landscape-level assessments and​​ enables the extraction of​​​‌ essential landscape variables that​ are crucial for wildfire​‌ forest monitoring.

10.4 Public​​ policy support

  • IMPACT -​​​‌ OFB (Office Français pour​ la Biodivérsité)

    Participants: Dino​‌ Ienco.

    Title:​​ Integrating the mosaic of​​​‌ landscapes mapped by remote​ sensing, and the associated​‌ epidemic risk, for more​​ agro-ecological management of regulated​​​‌ diseases of perennial crops​

    Duration: From September​‌ 2023 to August, 2031​​

    Summary: This project​​​‌ focuses on regulated vector-borne​ diseases of perennial plants.​‌ Epidemiosurveillance of these diseases,​​ centered on a crop​​​‌ of interest, will be​ extended to the diversity​‌ of landscape mosaics in​​ order to reduce compulsory​​​‌ treatments and promote preventive​ management of reservoirs. Remote​‌ sensing and modeling will​​ be used to map​​​‌ the risk of 3​ diseases: HLB, sharka and​‌ flavescence dorée (FD). More​​ detailed characterization of the​​​‌ FD pathosystem (experiments on​ the vector, sequencing to​‌ distinguish between epidemic and​​ non-epidemic variants) will also​​​‌ improve its management.

    EVERGREEN​ is in charge of​‌ conducting analysis and exploration,​​ through the analysis of​​​‌ multi-temporal remote sensing data,​ to detect the current​‌ status of vineyard plantations.​​ To this end, a​​​‌ research engineer, for a​ duration of 18 months,​‌ is scheduled in the​​ framework of the project.​​​‌

11 Dissemination

11.1 Promoting​ scientific activities

11.1.1 Scientific​‌ events: organization

Local

National

  • Dino​​​‌ Ienco has co-organized the​ IA+Remote Sensing event in​‌ the frame of the​​ AI Prospects (half a​​​‌ day of workshop and​ discussion) of the INEE​‌ institute of CNRS.

International​​

  • Cássio Fraga Dantas ,​​ Roberto Interdonato , and​​​‌ Dino Ienco have organized‌ the seventh edition of‌​‌ the MACLEAN (Machine Learning​​ for Earth Obervation Data)​​​‌ workshop co-located with the‌ European conference on Machine‌​‌ Learning and Data Mining​​ (ECML/PKDD).
  • Roberto Interdonato organized​​​‌ the special session “Network‌ Science Meets AI” as‌​‌ part of the ESANN​​ 2025 conference, European Symposium​​​‌ on Artificial Neural Networks,‌ Computational Intelligence and Machine‌​‌ Learning, Bruges (Belgium), April​​ 23-25, 2025. The session​​​‌ was co-organized with Matteo‌ Zignani, University of Milan‌​‌ (Italy), Fragkiskos D. Malliaros,​​ Paris-Saclay University (France), Ingo​​​‌ Scholtes, Julius-Maximilians-Universitat Wuurzburg (Germany)‌ and Manuel Dileo, University‌​‌ of Milan (Italy) .​​
  • Roberto Interdonato co-organized the​​​‌ SAI4OID worksop, 1st International‌ Workshop on Sustainable Artificial‌​‌ Intelligence for addressing Online​​ Information Disorder, co-located with​​​‌ the 24th International Conference‌ on Web Intelligence and‌​‌ Intelligent Agent Technology (WI-IAT​​ 2025, 17 November, 2025,​​​‌ London, United Kingdom). Workshop‌ is co-organized with Francesco‌​‌ Scala (ICAR-CNR, Italy), Liliana​​ Martirano (ICAR-CNR, Italy), Marco​​​‌ Minici (ICAR-CNR, Italy), Luca‌ Luceri (USC Information Sciences‌​‌ Institute, USA), Sergio Flesca​​ (Università della Calabria, Italy).​​​‌

11.1.2 Scientific events: selection‌

Chair of conference program‌​‌ committees
  • Roberto Interdonato is​​ Short Paper Program Chair​​​‌ of WSDM 2026, The‌ 19th ACM International Conference‌​‌ on Web Search and​​ Data Mining (Boise, Idaho,​​​‌ USA, February 22 –‌ 26, 2026).
Member of‌​‌ the conference program committees​​
  • Dino Ienco : DS25​​​‌ (Discovery Science), ECML/PKDD25 -‌ Journal Track (European Conference‌​‌ on Machine Learning and​​ Principle and Practice of​​​‌ Knowledge Discovery), PAKDD25 (Pacific‌ Asian Conference on Knoweldge‌​‌ Discovery from Data), IJCAI25​​ (International Joint Conference on​​​‌ Artificial Intelligence), Senior PC‌ AAAI25 (AAAI Conference on‌​‌ Artificial Intelligence), EGC25 (Extraction​​ et Gestion des Connaissances),​​​‌ APIA25 (Conférence Nationale sur‌ les Applications Pratiques de‌​‌ l’Intelligence Artificielle).
  • Roberto Interdonato​​ : BMVC25 (British Machine​​​‌ Vision Conference 2025), ECML/PKDD25‌ (European Conference on Machine‌​‌ Learning and Principle and​​ Practice of Knowledge Discovery),​​​‌ AAAI25 - Special Track‌ on AI for Social‌​‌ Impact (AAAI Conference on​​ Artificial Intelligence), DS25 (Discovery​​​‌ Science), CCS25 (Conference on‌ Complex Systems 2024), Complex‌​‌ Networks 2025 (International Conference​​ on Complex Networks and​​​‌ their Applications), IC2S2'24 (International‌ Conference for Computational Social‌​‌ Science), CARI'2025 (African Conference​​ on Research in Computer​​​‌ Science), FRCCS 2025 (French‌ Regional Conference on Complex‌​‌ Systems 2025), IGARSS 2025​​ (International Geoscience and Remote​​​‌ Sensing Symposium 2025).
  • Cássio‌ Fraga Dantas : International‌​‌ Geoscience and Remote Sensing​​ Symposium (IGARSS), IEEE International​​​‌ Conference on Multimedia (ICME),‌ International Conference on Computer‌​‌ Vision (ICCV), IEEE/CVF Winter​​ Conference on Applications of​​​‌ Computer Vision (WACV), Workshop‌ MVEO (Machine Vision for‌​‌ Earth Observation and Environment​​ Monitoring, in conjunction with​​​‌ BMVC'25), Workshop TerraBytes (Towards‌ global datasets and models‌​‌ for Earth Observation, in​​ conjunction with ICML'25).
  • Diego​​​‌ Marcos : Conference on‌ Computer Vision and Pattern‌​‌ Recognition (CVPR), International Conference​​ on Learning Representations (ICLR),​​​‌ NeurIPS, International Conference on‌ Computer Vision (ICCV), IEEE/CVF‌​‌ Winter Conference on Applications​​ of Computer Vision (WACV),​​​‌ Workshop MVEO (Machine Vision‌ for Earth Observation and‌​‌ Environment Monitoring, in conjunction​​ with BMVC'25).

11.1.3 Journal​​​‌

Member of the editorial‌ boards
  • Dino Ienco :‌​‌ Associate Editor for the​​​‌ Scientific Report journal (Springer​ Nature), Associate Editor for​‌ the Remote Sensing journal​​ (MDPI), member of the​​​‌ Editorial Board of the​ ISPRS Journal of Photogrammetry​‌ and Remote Sensing, Action​​ Editor for the Machine​​​‌ Learning journal (Springer Nature).​
  • Roberto Interdonato : Associate​‌ Editor for the Applied​​ Network Science journal (Springer​​​‌ Nature), member of the​ Editorial Board of the​‌ Frontiers in Big Data​​ journal.
Reviewer - reviewing​​​‌ activities
  • Dino Ienco :​ IEEE Geoscience and Remote​‌ Sensing Magazine, IEEE Geoscience​​ and Remote Sensing Letters,​​​‌ IEEE Journal of Selected​ Topics in Applied Earth​‌ Observations and Remote Sensing,​​ IEEE Transactions on Geoscience​​​‌ and Remote Sensing, ISPRS​ Journal of Photogrammetry and​‌ Remote Sensing (Elsevier), Data​​ Mining and Knowledge Discovery​​​‌ (Springer), Machine Learning Journal​ (Springer), Remote Sensing of​‌ Environment (Elsevier), Multimedia Tools​​ and Applications (Elsevier), Applied​​​‌ Intelligence (Elsevier), International Journal​ of Remote Sensing (Taylor​‌ and Francis), Computers and​​ Electronics in Agriculture (Elsevier),​​​‌ Neural networks (Elsevier).
  • Roberto​ Interdonato : IEEE Transactions​‌ on Geoscience and Remote​​ Sensing, IEEE Journal of​​​‌ Selected Topics in Applied​ Earth Observations and Remote​‌ Sensing, PLOS One, Expert​​ Systems with Applications (Elsevier),​​​‌ ISPRS Journal of Photogrammetry​ and Remote Sensing (Elsevier),​‌ International Journal of Applied​​ Earth Observation and Geoinformation​​​‌ (Elsevier), ACM Computing Surveys,​ Artificial Intelligence (Springer), World​‌ Development (Elsevier), Data Mining​​ and Knowledge Discovery (Springer),​​​‌ Neurocomputing (Elsevier), Geo-spatial Information​ Science (Taylor & Francis),​‌ Artificial Intelligence in Geosciences​​ (Elsevier).
  • Cássio Fraga Dantas​​​‌ : IEEE Journal of​ Selected Topics in Applied​‌ Earth Observations and Remote​​ Sensing (JSTARS), IEEE Transactions​​​‌ on Multimedia (TMM), IEEE​ Transactions on Artificial Intelligence​‌ (TAI).

11.1.4 Invited talks​​

National

  • Diego Marcos:​​​‌ Adapting Maxent species distribution​ models to deep learning​‌ and remote sensing at​​ Prospective Biodiversité & IA,​​​‌ CNRS, Paris.

International

  • Dino​ Ienco: Collaborative Cross-Modal​‌ Knowledge Distillation via Disentanglement​​ Representation for imagery data​​​‌ at the Jožef Stefan​ Institute, Ljubjana, Slovenia.

11.1.5​‌ Scientific expertise

  • Dino Ienco​​ : Remote Reviewer for​​​‌ the Research Foundation Flanders​ - FWO post-doctoral project.​‌
  • Dino Ienco : Member​​ of an ANR committee​​​‌ for project evaluations.
  • Roberto​ Interdonato : Reviewer for​‌ “Starting Grant” for Young​​ Researchers, University of Cagliari​​​‌ (UniCA), Italy.

11.1.6 Research​ administration

  • Dino Ienco is​‌ participating to the activities​​ of the Doctoral School​​​‌ I2S (Information Structures Systèmes)​ as a sub referee​‌ of the Computer Science​​ speciality. Among the other​​​‌ tasks related to this​ activity, this involves the​‌ participation to several Doctoral​​ advisory committee per year​​​‌ (around 20/25).
  • Dino Ienco​ is member of the​‌ CEP (Committee of the​​ Project Team) of the​​​‌ Inria Center at Université​ Côte d'Azur.
  • Dino Ienco​‌ is member of the​​ BCEP (Bureau of the​​​‌ the Project Team Committee)​ of the Inria Center​‌ at Université Côte d'Azur.​​
  • Dino Ienco is Member​​​‌ of the Scientific Direction​ Council (Conseil de Direction​‌ Scientifique) at the UMR​​ TETIS.
  • Roberto Interdonato is​​​‌ head of the MISCA​ team (Modélisation de l'Information​‌ Spatiale, extraction de Connaissances​​ et Analyse) at UMR​​​‌ TETIS.
  • Roberto Interdonato is​ Research Scientists’ representative at​‌ the UMR TETIS Council​​ (Conseil d’Unité).
  • Roberto Interdonato​​ is Member of the​​​‌ Scientific Direction Council (Conseil‌ de Direction Scientifique) at‌​‌ the UMR TETIS.

11.2​​ Teaching - Supervision -​​​‌ Juries - Educational and‌ pedagogical outreach

11.2.1 Teaching‌​‌

  • Cássio Fraga Dantas :​​ Introduction to Machine Learning​​​‌ module (12h of practical‌ sessions) for M1 students‌​‌ from the Energy and​​ Environment masters at EPF​​​‌ engineering school, Montpellier. Introduction‌ to Machine Learning module‌​‌ (9h courses, 12h practical​​ sessions) for M1 students​​​‌ from the Data Engineering‌ masters at EPF engineering‌​‌ school.
  • Raffaele Gaetano :​​ at MSc level, responsible​​​‌ for the teaching unit‌ “Spatial imagery for the‌​‌ management of environmental resources”,​​ M2 Géomatique, jointly​​​‌ organized by Université de‌ Montpellier, Université Paul Valéry‌​‌ and AgroParistech (7h courses,​​ 15h practical sessions). For​​​‌ professional training, participation to‌ the teaching unit “Information‌​‌ extraction from remote sensing​​ imagery”, Mastère SILAT,​​​‌ AgroParistech (2h courses, 5h‌ practical session). R. Gaetano‌​‌ also delivered a tutorial​​ on the use of​​​‌ the MORINGA processing chain‌ to partners in the‌​‌ OBSYDYA project (24h practical​​ sessions) in February 2025.​​​‌
  • Diego Marcos : Advanced‌ Data Science (4.5h of‌​‌ courses and 15h of​​ practical sessions) and Artificial​​​‌ Intelligence at Université de‌ Montpellier.

11.2.2 Supervision

  • PhD‌​‌ in progress: Ananthu Aniraj,​​ Explainable image classification through​​​‌ supervised and unsupervised part‌ detection, co-advised by Diego‌​‌ Marcos , Cássio Fraga​​ Dantas and Dino Ienco​​​‌ , funded by the‌ CPJ OBTEA, since April‌​‌ 2023.
  • PhD in progress:​​ Camille Portes, Increased risk-based​​​‌ epidemio-surveillance for Xylella fastidiosa,‌ advised by Dino Ienco‌​‌ , in collaboration with​​ Edith Gabriel (BioSP -​​​‌ Biostatistique et Processus Spatiaux‌ Unit), funded by École‌​‌ Universitaire de Recherche (EUR)​​ Implanteus, since October 2021.​​​‌
  • PhD in progress: Bruno‌ Bio Nikki, Deep learning‌​‌ and multi-sensor remote sensing​​ imagery for land use​​​‌ and land cover mapping‌ of agricultural systems in‌​‌ Northern Benin, co-advised by​​ Raffaele Gaetano , Roberto​​​‌ Interdonato (co-supervised by Prof.‌ Yvon C. Hountoundji, Univeristy‌​‌ of Parakou - Benin),​​ funded by the EU​​​‌ OBSYDYA project, since October‌ 2023.
  • PhD in progress:‌​‌ Ron van Bree, Hybrid​​ AI for Food Security,​​​‌ co-advised by Ioanis Athanasiadis‌ (Wageningen University) and Diego‌​‌ Marcos , since April​​ 2024.
  • PhD in progress:​​​‌ Valerie Zermatten, co-advised by‌ Devis Tuia (EPFL) and‌​‌ Diego Marcos , since​​ 2023
  • PhD in progress:​​​‌ Christopher Jabea, AI approaches‌ for the identification of‌​‌ morphological patterns in coastal​​ areas from Earth observation​​​‌ data, co-advised by Cássio‌ Fraga Dantas and Dino‌​‌ Ienco , in collaboration​​ with Isabelle Manighetti (University​​​‌ Côté D'Azur), Bruno Castelle‌ (CNRS).
  • PhD in progress:‌​‌ Anas Zakroum, Analysis of​​ complex networks for landscape​​​‌ dynamics analysis, co-advised by‌ Roberto Interdonato , Pascal‌​‌ Degenne, Danny Loseen and​​ Mathieu Roche (UMR TETIS,​​​‌ CIRAD, France).
  • PhD in‌ progress: Hugo Riffaud de‌​‌ Turckheim, Fundamental geospatial models​​ for Earth observation: integration​​​‌ of textual and remote‌ sensing data, co-advised by‌​‌ Diego Marcos , Roberto​​ Interdonato and Sylvain Lobry​​​‌ (Université Paris Cité, France).‌
  • PhD in progress: Pablo‌​‌ Ubilla, Deep learning-based species​​ distribution models for integrating​​​‌ different data sources, co-advised‌ by Diego Marcos ,‌​‌ Roberto Interdonato and Christophe​​​‌ Botella (Inria Montpellier, IROKO​ team).

11.2.3 Juries

HDR​‌
  • Dino Ienco was a​​ member of the following​​​‌ HDR Jury in 2025​ (2): Nicolas Audebert, University​‌ Paris-Est, Paris, France (examiner);​​ Lionel Bombrun, University of​​​‌ Bordeaux, Bordeaux, France (reviewer).​
PhD
  • Dino Ienco was​‌ a member of the​​ following PhD Juries in​​​‌ 2025 (8): George Killick,​ University of Glasgow, Glasgow,​‌ UK (opponent); Ataollah Kamal,​​ University of Lyon, Lyon,​​​‌ France (reviewer); Colin Prieur,​ University of Montpellier, Montpellier,​‌ France (president); Matteo Contini,​​ University of Montpellier, Montpellier,​​​‌ France (president); Marion Boyer,​ Inria Center of the​‌ University Côte d'Azur, Sophia-Antipolis,​​ France (reviewer); Mahdi Djama​​​‌ Rayaleh, University of Montpellier,​ Montpellier, France (president); Sarah​‌ Mauny, University Paris-Saclay, Paris,​​ France (examiner); Marjan Stoimchev,​​​‌ University of Ljubjana, Ljubjana,​ Slovenia (reviewer).
  • Roberto Interdonato​‌ was a member of​​ the following PhD juries​​​‌ in 2025 (3): Florian​ Teste, AgroParisTech, France (examiner);​‌ Corentin Dufourg, Université de​​ Bretagne Sud, France (reviewer);​​​‌ Asma Mesdour, CIRAD, France​ (examiner).
Doctoral advisory commitee​‌
  • Dino Ienco was a​​ member of the following​​​‌ PhD mid-term evaluation committees​ (21): Axel Dubar (Univ.​‌ Montpellier), Guillaume Fouret (Univ.​​ Montpellier), Elliot Butz (Univ.​​​‌ Montpellier), Charlotte Fabre (Univ.​ Montpellier), Khelian Larvet (Univ.​‌ Montpellier), Erwan Reinders (Univ.​​ Montpellier), Guillaume Picaud (Univ.​​​‌ Montpellier), Florian Lecourt (Univ.​ Montpellier), Théo Larcher (Univ.​‌ Montpellier), Loai Gandeel (Univ.​​ Montpellier), Guillaume Coulaud (Univ.​​​‌ Montpellier), Charles Berger (Univ.​ Montpellier), Kawtar Zaher (Univ.​‌ Montpellier), Sebastien Gigot (Univ.​​ Montpellier), Imran Meghazi (Univ.​​​‌ Montpellier), Eugenio Dias Riberto​ Neto (Univ. Montpellier), Anas​‌ Zakroum (Univ. Montpellier), Nathan​​ Guilhot (Univ. Montpellier), Charles​​​‌ Ngom (Univ. Montpellier), Tiziano​ Maisonhaute (Univ. Montpellier), Remy​‌ Decoupes (Univ. Montpellier).
  • Roberto​​ Interdonato was a member​​​‌ of the following PhD​ mid-term evaluation committees (4):​‌ Camille Portes (Univ. Avignon),​​ Hussam Ghanem (Université Bourgogne​​​‌ Europe), Théo Morel (Univ.​ Normandie Le Havre), Nicolas​‌ Houdré (Université Paris Cité).​​
Recruitment
  • Dino Ienco :​​​‌ Member of the recruitment​ jury for an INRIA​‌ Junior Research - Chargé​​ de Recherche - (INRIA).​​​‌

11.3 Popularization

12 Scientific​​ production

12.1 Major publications​​​‌

12.2 Publications of the​​​‌ year

International journals

International​​ peer-reviewed conferences

Reports &​ preprints

12.3 Cited publications​​

  • 33 articleA.Améline​​​‌ Vallet, S.Stéphane​ Dupuy, M.Matthieu​‌ Verlynde and R.Raffaele​​ Gaetano. Generating high-resolution​​​‌ land use and land​ cover maps for the​‌ greater Mariño watershed in​​ 2019 with machine learning​​​‌.Scientific Data 11​12024, 915​‌HALDOIback to​​ text