Label-Free Degradation Diagnosis for Classifier Selection in Hyperspectral Scenes.

In hyperspectral image classification, the most suitable classifier depends on the magnitude and the type of degradation present in the scene; this information, however, cannot be obtained without labels at the stage where the choice has to be made. A two-layer diagnosis computed before classification and without labels is proposed in this study: a severity index reports the magnitude of the degradation, while three scene-derived indicators separate independent, band-dependent, spectrally correlated and striped structures from one another, while classification itself remains supervised. The approach is evaluated on ten benchmark scenes under a leakage-aware spatial protocol. Classifier fragility is found to follow two opposing regimes: tree ensembles are robust to spatially local degradation but fragile to degradation spread across the spectrum, whereas the opposite is observed for the convolutional network. Regime-based classifier selection produces the better classification map in every combination examined. Replacing the spatial protocol with a random pixel split is further shown to inflate accuracy by between 0.162 and 0.249 and to change the best classifier in six of the ten scenes.