From Pixels to Semantics: Benchmarking Geospatial Foundation Model Embeddings for Vegetation Frost Damage Detection

Late frost events following anomalous warm periods pose substantial risks to Central European agriculture and forestry. The April 2024 frost in the Czech Republic caused widespread vegetation damage, yet standard optical monitoring failed to capture it due to cloud cover and subsequent vegetation recovery. This study benchmarked two Geospatial Foundation Models— AlphaEarth Foundations (AEF, 64-dimensional annual embeddings) and TESSERA (128-dimensional annual embeddings)—for detecting frost-induced vegetation anomalies across four Czech regions. AEF embeddings revealed a statistically significant anomaly gradient consistent with differential frost sensitivity: vineyards and orchards (Z = 3.07–5.16) exhibited the strongest deviation from the 2019–2023 baseline, followed by broadleaf forests (Z = 2.83–3.89 across regions), while coniferous forests showed variable responses. TESSERA annual embeddings detected no frost signal, attributable to their temporal invariance training objective. Complementary Sentinel-2 CIre time series on paired anomalous/control sites confirmed chlorophyll depression at AEF-identified locations on the first cloud-free acquisition post-frost (29 April 2024). These findings illustrate that GFMs can detect vegetation anomalies despite critical periods of obscuring cloud cover, but spectral time series remain essential for temporal attribution.