Clay v1.5 + U-Net hybrid boosts landslide detection accuracy by 7.7%
Geo-foundational model as auxiliary context yields F1 of 64.5% on extreme class imbalance.
Researchers evaluated Clay v1.5, a Geo-Foundational Model (GFM) pretrained on satellite imagery, to improve automated landslide mapping from Sentinel-2 and terrain data. The Landslide4Sense benchmark presents extreme class imbalance (only ~2% positive pixels out of 3,799 training chips), making rapid post-event detection notoriously difficult.
Three strategies were compared: Clay as a standalone encoder with multi-scale terrain fusion, a U-Net backbone augmented with Clay’s semantic context at the bottleneck via Low-Rank Adaptation (LoRA), and a standard U-Net baseline. The hybrid U-Net + Clay model achieved the top test F1 score of 64.5 ± 1.8% over three seeds, surpassing the U-Net baseline (59.9%) and the Clay-only backbone (55.2%). The standalone Clay encoder underperformed due to lack of multi-scale skip connections, confirming that GFMs are most effective when complementing spatially detailed CNNs rather than replacing them.
- Hybrid U-Net + Clay v1.5 with LoRA achieved 64.5% F1, beating both Clay-only (55.2%) and U-Net baseline (59.9%)
- Clay-as-encoder underperformed due to missing multi-scale skip connections, despite pretrained representations
- Study used 14 Sentinel-2 and terrain bands from 3,799 training chips with only 2% positive landslide pixels
Why It Matters
GFMs like Clay can drastically improve disaster response mapping when hybridized, not just used as standalone models.