Remote Sensing Benchmarks

Remote sensing benchmarks provide shared datasets, splits, metrics, and protocols for testing satellite, aerial, SAR, hyperspectral, and LiDAR models on tasks such as land-cover mapping, change detection, geo-localization, and vision-language reasoning. They compare CNNs and transformers with foundation models using supervised, zero-shot, frozen-encoder, and fine-tuning evaluations, often with cross-region, cross-year, or cross-sensor splits. Results show that rankings can change across tasks and conditions: geographic and sensor shifts, fine details, temporal reasoning, and complex grounding expose weaknesses that clean in-distribution scores may hide. A central challenge is to make comparisons reproducible and realistic while measuring calibration, geometric accuracy, efficiency, robustness, and downstream utility rather than relying on a single aggregate score.