Label-Efficient Change Detection with Precomputed Satellite Embeddings under Spatial and Temporal Shift

Preetam Chhimpa

Abstract: Label scarcity is a practical bottleneck for satellite change detection. Models must generalize across geography and time with limited human supervision. Using precomputed 64-D annual satellite embeddings from Google Earth Engine, we study Brazil forest-loss detection from MapBiomas land-cover transitions (2017$-$2022) under tile-level spatial holdout and a held-out future transition (2020$\rightarrow$2021). A simple linear probe on embeddings is already strong (F1 $\approx$ 0.98). We then simulate pool-based active learning and show that uncertainty sampling reaches 99.5\% of full-data performance with substantially fewer labels than random selection: 3.73$\times$--10.68$\times$ fewer labels when random reaches the target, and lower bounds up to $>$19$\times$ when it does not. These results support a practical label-efficient workflow in this proxy-label setting: precomputed embeddings + linear probe + uncertainty-based querying.

Submission Number: 10