Robust Dual-Model Approach for Noise-Resilient Detection of Small Camouflaged Targets in Wide-Field Hyperspectral Imaging

Hyperspectral imaging plays a crucial role in long-distance target detection, yet challenges arise from data complexity and noise interference. In the hyperspectral images acquired for this study, the camouflaged target occupies on average fewer than 25 pixels—approximately 0.0007% of the total pixels—which underscores the extreme difficulty of detecting such minute objects. This study proposes a robust dual-model fusion approach, integrating a Back Propagation (BP) neural network and Random Forest to enhance detection accuracy and noise suppression. By leveraging BP’s nonlinear pattern recognition capabilities and Random Forest’s ensemble decision-making, the method effectively identifies small, camouflaged targets in wide-field hyperspectral imagery while maintaining low false-alarm rates. Experiments using FS-22 hyperspectral data (300 spectral channels, 1920 × 1920 resolution) successfully detected camouflaged vehicle at long distances under challenging noise conditions. The dual model demonstrates superior performance in terms of probability of detection, false-alarm rate suppression, and noise removal compared to individual models. The results validate the effectiveness of the proposed approach for noise-resilient, long-range hyperspectral target detection with enhanced reliability and accuracy.