Rajiv Ranjan · Udaiveer Singh · Anjali Aggarwal · Shashank Tamaskar · Dharmendra Saraswat
SPEAR is a modular, pixel-level self-supervised framework for Earth observation that learns sensor-adaptive representations across optical, radar, and climate modalities. Three dedicated learners, Spectral-MAE (multispectral), BYOL-Denoise (radar), and Climate-MAE (environmental variables), produce compact 32-D per-pixel embeddings conditioned on band wavelength, spatial scale, and geolocation. A sensor- and scale-aware fusion head integrates these via cross-attention, with learned zero-slot tokens enabling robust inference under missing-modality conditions. Despite only ~1M parameters, SPEAR achieves competitive accuracy on Dynamic World land-cover (88.6%), Sen1Floods11 flood detection (97.5%), VIIRS fire detection (99.2%), and Sickle crop classification (73.5%).
This demo shows SPEAR's per-pixel land cover predictions fused across Sentinel-2, Sentinel-1, and PlanetScope. Use the Region dropdown to select a chip, then hover over the image to preview any pixel or click to lock it. The sensor toggle lets you remove modalities in real time and see how the 3D embedding vectors shift. The consistency score measures how stable the prediction remains as sensors are dropped, directly demonstrating SPEAR's missing-modality robustness.
Sentinel-2 · 10m resolution · 9 classes · Dynamic World benchmark
Select a region above to load imagery
Legend
Pixel Inspector
Hover over the image to preview any pixel, click to lock and inspect its full embedding.
The authors acknowledge CNH Industrial for field resources and collaboration support.