Land degradation, which manifests as desertification in arid and semi-arid regions, affects approximately 40% of the world's fertile land. Drylands are predominantly occupied by rangelands, which are highly sensitive to degradation processes that may trigger non-linear ecological transitions and long-term productivity losses. Achieving Land Degradation Neutrality under Sustainable Development Goal 15 requires moving beyond regreening approaches to address multifunctional indicators of degradation, such as those categorized by the Land Degradation Surveillance Framework (LDSF), including land use, land cover, land degradation, soil health, impact on habitat, and topography. This study, conducted in the Karamoja border region of Kenya and Uganda, integrates LDSF ground observations with Earth Observation ancillary gridded data (precipitation, temperature, and evapotranspiration) and 64-component spatial embeddings derived from the AlphaEarth Foundations Model using similarity-based clustering. Variable interactions were analyzed using random forest models to classify AlphaEarth-derived clusters and to predict the primary drivers of land degradation, soil erosion rate and herbaceous cover rating, at both regional and cluster-specific scales. Model outputs were interpreted using SHapley Additive exPlanations values. The classification of AlphaEarth-derived clusters reveals their correspondence to eco-climatic groupings, translating the high-dimensional embeddings into field-interpretable units (Accuracy = 0.98). The feature importance of soil erosion rate and herbaceous cover rating yielded R2 = 0.61 and R2 = 0.72, respectively. Additionally, feature importance of the land degradation drivers in the cluster-specific scales enables the identification of site-specific degradation patterns. This framework supports the identification of field-scale indicators for monitoring land degradation and defining context-specific restoration priorities for rangeland management through the combined use of deep learning, earth observation data, and field data.
AlphaEarth foundations model, geospatial derived data and land degradation surveillance framework combined with random forest for targeted rangeland restoration in Karamoja region / Bile, M.V., Mwangi Mureithi, S., Fiorentino, N., D'Urso, G., Belfiore, O.R.. - In: REMOTE SENSING APPLICATIONS. - ISSN 2352-9385. - 44:(2026). [10.1016/j.rsase.2026.102303]
AlphaEarth foundations model, geospatial derived data and land degradation surveillance framework combined with random forest for targeted rangeland restoration in Karamoja region
Maria Virginia Bile;Nunzio Fiorentino;Guido D'Urso;Oscar Rosario Belfiore
2026
Abstract
Land degradation, which manifests as desertification in arid and semi-arid regions, affects approximately 40% of the world's fertile land. Drylands are predominantly occupied by rangelands, which are highly sensitive to degradation processes that may trigger non-linear ecological transitions and long-term productivity losses. Achieving Land Degradation Neutrality under Sustainable Development Goal 15 requires moving beyond regreening approaches to address multifunctional indicators of degradation, such as those categorized by the Land Degradation Surveillance Framework (LDSF), including land use, land cover, land degradation, soil health, impact on habitat, and topography. This study, conducted in the Karamoja border region of Kenya and Uganda, integrates LDSF ground observations with Earth Observation ancillary gridded data (precipitation, temperature, and evapotranspiration) and 64-component spatial embeddings derived from the AlphaEarth Foundations Model using similarity-based clustering. Variable interactions were analyzed using random forest models to classify AlphaEarth-derived clusters and to predict the primary drivers of land degradation, soil erosion rate and herbaceous cover rating, at both regional and cluster-specific scales. Model outputs were interpreted using SHapley Additive exPlanations values. The classification of AlphaEarth-derived clusters reveals their correspondence to eco-climatic groupings, translating the high-dimensional embeddings into field-interpretable units (Accuracy = 0.98). The feature importance of soil erosion rate and herbaceous cover rating yielded R2 = 0.61 and R2 = 0.72, respectively. Additionally, feature importance of the land degradation drivers in the cluster-specific scales enables the identification of site-specific degradation patterns. This framework supports the identification of field-scale indicators for monitoring land degradation and defining context-specific restoration priorities for rangeland management through the combined use of deep learning, earth observation data, and field data.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


