Extracting spatially and spectrally coherent regions from multispectral images

Extracting spectrally homogeneous regions as features from hyperspectral and multispectral raster data has unique challenges when accurate shape preservation is a priority. We tackle this task by representing neighborhoods that contain heterogeneously classified pixels as a graph. We then use graph-cut based combinatorial optimization to eliminate spuriously classified pixels. After the region of interest is uniformly classified, we use a vectorization step to extract it as a feature.