calebrob6/similar-but-different

Similar But Different 数据集是一个精心策划的 Sentinel-2 L2A 多光谱补丁集合，包含 72,897 个 32×32 的补丁（12个波段重采样至10米分辨率）。这些补丁的选择标准是：它们至少参与一对补丁，这些补丁在真彩色RGB中看起来几乎完全相同，但在近红外反射率上存在显著差异。这使得该数据集成为对仅依赖可见光波段的模型进行压力测试的小型但有用的工具：当只有RGB可用时，许多“明显”的类别（如森林与农田、水体与深色屋顶）变得真正难以区分。数据集基于2627个Sentinel-2 L2A场景随机采样构建，覆盖82个手动选择的全球种子区域（包括森林、城市、农业、水体、沙漠、湿地、山脉、稀树草原等）。每个补丁使用ESA WorldCover 10米2021年土地覆盖数据分配了模式类别作为标签，并按照类别组织存储为12波段uint16压缩GeoTIFF文件。数据集还提供了按类别分层的80/10/10训练/验证/测试分割，以及基线结果，展示了使用不同波段组合的模型性能。Similar But Different is a curated set of 72,897 32×32 Sentinel-2 L2A multispectral patches (12 bands resampled to 10 m) selected because they participate in at least one pair of patches that look nearly identical in true-colour RGB but differ substantially in near-infrared (NIR) reflectance. This makes the dataset a small but useful stress-test for models that rely on the visible bands alone: many obvious classes (forest vs. cropland, water vs. dark roof) become genuinely hard when only RGB is available. The dataset is built from randomly sampled windows across 2627 Sentinel-2 L2A scenes from 82 hand-picked global seeds (forest, urban, agriculture, water, desert, wetland, mountain, savanna). Each patch is labeled with the mode class from ESA WorldCover 10 m 2021 land-cover distribution and stored as 12-band uint16 deflate-compressed GeoTIFF files organized by class. It includes a stratified 80/10/10 train/val/test split and baseline results showing model performance with different band combinations.