Periodic Reporting for period 1 - EO4FoodSecurity (EO4FoodSecurity: Using Earth Observation Enabled Land Cover Classification for Characterizing Global Food Security on Regional Scales)
Okres sprawozdawczy: 2023-07-01 do 2025-09-30
We are currently living in the golden era of Big Earth Observation (EO) data. Continuous, systematic monitoring of our planet, particularly through the Copernicus Sentinel missions, has transformed our understanding of environmental and socioeconomic systems. EO data is now indispensable for addressing global challenges such as urbanisation, climate change, and agricultural sustainability. Yet despite its abundance, a critical gap remains in turning vast EO datasets into actionable insights for decision-making, especially in food security.
Global and regional food security assessments continue to be limited by data scarcity, inconsistent reporting, and restricted analytical capacity. Existing approaches lack the scalability and precision needed to convert raw EO data into decision-relevant indicators of food availability, access, utilisation, and stability. Addressing these shortcomings was the core motivation of EO4FoodSecurity, which sought to build on advanced AI algorithms and large-scale EO processing frameworks developed in the ERC Starting Grant So2Sat. The project aimed to extend these foundations to characterise food security at broader scales and to refine AI-driven land cover and land use mapping for deriving key indicators of food system status and sustainability.
The overarching vision was to create innovative, scientifically robust, and scalable tools capable of extracting geospatial and socio-economic indicators relevant for food security directly from EO data. By integrating big EO datasets with other open sources through state-of-the-art AI, the project provides a foundation for a new generation of EO-driven food security analytics to support evidence-based decision-making, sustainable land management, and resilient food systems.
Global and regional food security assessments continue to be limited by data scarcity, inconsistent reporting, and restricted analytical capacity. Existing approaches lack the scalability and precision needed to convert raw EO data into decision-relevant indicators of food availability, access, utilisation, and stability. Addressing these shortcomings was the core motivation of EO4FoodSecurity, which sought to build on advanced AI algorithms and large-scale EO processing frameworks developed in the ERC Starting Grant So2Sat. The project aimed to extend these foundations to characterise food security at broader scales and to refine AI-driven land cover and land use mapping for deriving key indicators of food system status and sustainability.
The overarching vision was to create innovative, scientifically robust, and scalable tools capable of extracting geospatial and socio-economic indicators relevant for food security directly from EO data. By integrating big EO datasets with other open sources through state-of-the-art AI, the project provides a foundation for a new generation of EO-driven food security analytics to support evidence-based decision-making, sustainable land management, and resilient food systems.
During the project, both technical and market development activities were pursued in parallel. The market and policy analysis examined existing food security indexes and EO-derived indicators at global, national, and household levels. This review identified significant methodological gaps, including the limited use of high-resolution EO data in developing scalable, data-driven food security indicators.
On the technological front, the project advanced two principal technological directions for enhancing EO–based services relevant to food security: high-resolution environmental and vegetation mapping, and large-scale urban infrastructure analysis. High-resolution mapping supports robust monitoring of cropland distribution, vegetation productivity, and ecosystem condition, key determinants of food availability and ecological sustainability. Urban infrastructure analysis complements this by characterizing settlement density, built-up extent, and expansion patterns, thereby linking food demand, accessibility, and environmental pressures. Together, these components provide an integrated geospatial basis for analysing interactions between agricultural production, ecosystem function, and human consumption dynamics.
Within environmental and vegetation mapping, the project prioritised tree species mapping due to the critical role of many species as direct or supplementary food sources, especially in tropical and rural regions. Accurate information on species distribution, range shifts, and stress signals is essential for anticipating food supply risks and for assessing ecosystem services that support agriculture. However, global tree species mapping remains limited by insufficient labelled data. To address this, the project produced GlobalGeoTree, the first comprehensive global dataset for tree species classification, containing 6.3 million geolocated occurrences spanning 275 families, 2,734 genera, and 21,001 species. A baseline vision-language model, GeoTreeCLIP, pretrained on GlobalGeoTree, showed substantial improvements in zero- and few-shot classification over existing models. The dataset, model, and code have been publicly released.
For urban infrastructure analysis, the project extended earlier work from the ERC Starting Grant So2Sat, employing weakly supervised domain adaptation to mitigate the scarcity of detailed reference data. This enabled the generation of high-resolution land cover maps for ten global cities. In parallel, the project contributed to the GlobalBuildingAtlas, the first global, high-quality dataset of individual building footprints, heights, and LoD1 3D models. Comprising 2.75 billion polygons, it surpasses all existing datasets and includes a 3 × 3 m global height layer. Although not a formal EO4FoodSecurity deliverable, its development benefited directly from methodological advances consolidated within the project.
The project also performed an urban heat island (UHI) assessment across 80 German cities using remote sensing, landscape metrics, and interpretable machine learning. Results showed strong context dependence in UHI drivers, including contrasting effects of tree-cover edge density depending on local vegetation–built-up ratios. This framework provides a transferable methodology for integrating microclimate and land-cover indicators into regional food security assessments.
On the technological front, the project advanced two principal technological directions for enhancing EO–based services relevant to food security: high-resolution environmental and vegetation mapping, and large-scale urban infrastructure analysis. High-resolution mapping supports robust monitoring of cropland distribution, vegetation productivity, and ecosystem condition, key determinants of food availability and ecological sustainability. Urban infrastructure analysis complements this by characterizing settlement density, built-up extent, and expansion patterns, thereby linking food demand, accessibility, and environmental pressures. Together, these components provide an integrated geospatial basis for analysing interactions between agricultural production, ecosystem function, and human consumption dynamics.
Within environmental and vegetation mapping, the project prioritised tree species mapping due to the critical role of many species as direct or supplementary food sources, especially in tropical and rural regions. Accurate information on species distribution, range shifts, and stress signals is essential for anticipating food supply risks and for assessing ecosystem services that support agriculture. However, global tree species mapping remains limited by insufficient labelled data. To address this, the project produced GlobalGeoTree, the first comprehensive global dataset for tree species classification, containing 6.3 million geolocated occurrences spanning 275 families, 2,734 genera, and 21,001 species. A baseline vision-language model, GeoTreeCLIP, pretrained on GlobalGeoTree, showed substantial improvements in zero- and few-shot classification over existing models. The dataset, model, and code have been publicly released.
For urban infrastructure analysis, the project extended earlier work from the ERC Starting Grant So2Sat, employing weakly supervised domain adaptation to mitigate the scarcity of detailed reference data. This enabled the generation of high-resolution land cover maps for ten global cities. In parallel, the project contributed to the GlobalBuildingAtlas, the first global, high-quality dataset of individual building footprints, heights, and LoD1 3D models. Comprising 2.75 billion polygons, it surpasses all existing datasets and includes a 3 × 3 m global height layer. Although not a formal EO4FoodSecurity deliverable, its development benefited directly from methodological advances consolidated within the project.
The project also performed an urban heat island (UHI) assessment across 80 German cities using remote sensing, landscape metrics, and interpretable machine learning. Results showed strong context dependence in UHI drivers, including contrasting effects of tree-cover edge density depending on local vegetation–built-up ratios. This framework provides a transferable methodology for integrating microclimate and land-cover indicators into regional food security assessments.
The project had both technical and business/market analysis phases. In the business development phase, we carried out a thorough market analysis of existing food security indexes and EO-derived indicators at global, national, and household levels. This analysis allowed us to identify gaps in current methodologies, particularly in the integration of high-resolution EO data for scalable, data-driven food security assessments.
For technological development, we focused on two key solutions instrumental in extracting indicators for food security. First, GlobalGeoTree represents the most comprehensive global dataset for tree species classification, overcoming long-standing limitations caused by scarce labelled training data. Coupled with the GeoTreeCLIP model, the project introduces a vision-language paradigm that substantially improves zero- and few-shot classification, enabling scalable species mapping in previously data-poor regions. Second, the project extended ERC Starting Grant So2Sat methodologies, advancing urban infrastructure analysis through weakly supervised domain adaptation, generating high-resolution land cover maps across multiple global cities, and producing the GlobalBuildingAtlas, containing 2.75 billion building footprints and global LoD1 3D models, a new benchmark for completeness and spatial detail. Future work will integrate these datasets, representing human demand and socioeconomic context, with global land use classification capturing agricultural supply and environmental sustainability, forming a data-driven spatial foundation for a geo-intelligent food security index.
In terms of IPR and data governance, all algorithms, datasets, and derived products developed in this project are made publicly available to the scientific and policy communities. The project fully embraces the FAIR (Findable, Accessible, Interoperable, and Reusable) data principles to maximize transparency, reproducibility, and reusability of results.
Lastly, it is also worth mentioning that none of the work was subcontracted, i.e. all results were achieved in-house. Overall, the project can be considered a success, both in terms of its scientific outcomes and its contribution to advancing open, data-driven methodologies for global food security monitoring.
For technological development, we focused on two key solutions instrumental in extracting indicators for food security. First, GlobalGeoTree represents the most comprehensive global dataset for tree species classification, overcoming long-standing limitations caused by scarce labelled training data. Coupled with the GeoTreeCLIP model, the project introduces a vision-language paradigm that substantially improves zero- and few-shot classification, enabling scalable species mapping in previously data-poor regions. Second, the project extended ERC Starting Grant So2Sat methodologies, advancing urban infrastructure analysis through weakly supervised domain adaptation, generating high-resolution land cover maps across multiple global cities, and producing the GlobalBuildingAtlas, containing 2.75 billion building footprints and global LoD1 3D models, a new benchmark for completeness and spatial detail. Future work will integrate these datasets, representing human demand and socioeconomic context, with global land use classification capturing agricultural supply and environmental sustainability, forming a data-driven spatial foundation for a geo-intelligent food security index.
In terms of IPR and data governance, all algorithms, datasets, and derived products developed in this project are made publicly available to the scientific and policy communities. The project fully embraces the FAIR (Findable, Accessible, Interoperable, and Reusable) data principles to maximize transparency, reproducibility, and reusability of results.
Lastly, it is also worth mentioning that none of the work was subcontracted, i.e. all results were achieved in-house. Overall, the project can be considered a success, both in terms of its scientific outcomes and its contribution to advancing open, data-driven methodologies for global food security monitoring.