When Do Geospatial FoundationModels Help? : Evaluating Satellite Image Embeddings for WetlandClassification in Stockholm County

Sammanfattning

Wetlands are ecologically critical but their mapping remains complex and labour-intensive. In Sweden, wetlands maps are produced using BIOTOP SE, a two-step workflow that combines automated classification of remote sensing data with expert interpretation of aerial colour-infrared photographs. Recent geospatial foundation models (GeoFMs), such as AlphaEarth (multimodal annual embeddings) and TESSERA (Sentinel1/Sentinel-2 time-series embeddings), offer new opportunities to support habitat and wetland mapping. By providing general-purpose satellite-derived features, these models may help to reduce the manual interpretation and improve the scalability of wetland classification workflows. This thesis evaluates both AlphaEarth and TESSERA models against a handcrafted Sentinel-1/2 seasonal baseline for predicting five BIOTOP SE wetland structural classes across approximately 55,000 polygons in Stockholm County for the year 2019. All reported accuracies measure agreement with BIOTOP SE Step 1 structural labels under spatially disjoint evaluation, rather than against independent ecological ground truth. The results show that, under polygon-mean feature extraction and 5-fold spatial block cross-validation, AlphaEarth with LightGBM achieves the highest macro 𝐹1 (0.668±0.008), followed by TESSERA (0.641±0.009) and the handcrafted Sentinel baseline (0.637 ± 0.012). The AlphaEarth advantage is real but modest, and persists across label budgets, classifier families, multiyear stacking, and large-region geographic transfer. The thesis’s most transferable finding is methodological. Under singlepoint (single-pixel) extraction the comparative ranking of representations becomes unstable: TESSERA gains substantially more from within-polygon averaging than the other feature sets, plausibly because pixel-level syntheticaperture-radar speckle is suppressed by zonal aggregation. Polygon-mean (or equivalent zonal) sensitivity analysis is therefore warranted as a robustness check in labelled-polygon GeoFM benchmarks of this kind.