Response-Conditioned Observation Prototype Guidance for Deep Hyperspectral Unmixing

Observation-spectrum targets can supplement reconstruction in deep hyperspectral unmixing, but their effects depend on target construction and the network interface. We evaluate these effects across adapted backbones using response-conditioned observation prototype guidance (PG) as a specified response-weighted, exponential-moving-average target. The detached angular constraint adds no trainable parameters or inference branch. The original candidate scope contains 547 valid baseline–PG seed pairs in 110 of 112 planned backbone–scene combinations. Using mean paired relative error reductions, the spectral angle distance (SAD) improved in 90 combinations, abundance root mean square error (RMSE) in 70, and both in 61. The historically selected seven-backbone, four-scene collection contains 140 pairs, with corresponding counts of 27/28, 23/28, and 22/28. Mean relative RMSE reduction was 5.89% in this selected display but −0.21% in the remaining 82 valid combinations, limiting the abundance-improvement claim. Against a current unweighted top-3% target with MSE and angular terms, PG reduced both mean absolute errors in only two of eight tested combinations; the simpler current target is often competitive. This comparison and the support-dependent control do not isolate benefits of weighting or EMA. Synthetic conditions also produced opposite effects in two backbones. The study documents conditional effects and failure boundaries of observation-spectrum guidance across training interfaces; it does not establish a general advantage of response weighting or spectral memory.