A DEEP LEARNING APPROACH FOR ROBUST CROP CLASSIFICATION UNDER MIXED-PIXEL CONDITIONS IN HYPERSPECTRAL IMAGES SIMPLIFIED RESEARCH MANUSCRIPT BASED ON THE SUPPLIED SYNTHETIC EXPERIMENT

Abstract

Hyperspectral imaging records reflectance across many narrow wavelength bands and can help distinguish crop types. However, a pixel may contain signals from more than one crop, soil, shadow, or other material. This paper reviews the supplied synthetic benchmark for a mixture-aware spectral–spatial fusion concept called MA-SSFN. The experiment is described as using 48 abstract spectral features and 16 Salinas-inspired class labels; it does not establish the use of official Salinas imagery. Five methods were compared using overall accuracy (OA), macro-F1, Cohen’s kappa, and macro recall. RBF-SVM achieved the highest benchmark scores (OA = 0.939643; macro-F1 = 0.939509; κ = 0.935619). In the ablation study, spectral features combined with smoothed context achieved the highest scores (OA = 0.93643; macro-F1 = 0.93631). The full spectral-context-mixture configuration performed less well. These results do not show that the fusion-inspired model is superior. Because raw data, implementation details, repeated runs, and real-scene validation are not supplied, the findings should be treated as descriptive results from a synthetic experiment.