Lightweight Dual-Branch Target-Aware Network for Land-Based Hyperspectral Camouflage Target Detection

Abstract

Hyperspectral camouflaged target detection is challenging because target and background spectra are highly coupled, positive pixels are extremely scarce, and scene-dependent spectral variation limits cross-dataset robustness. Existing detectors often rely on a single spectral representation or use spatial models with substantial parameters and computation, which may discard complementary evidence or increase deployment costs. This article presents a lightweight dual-branch target-aware network for land-based hyperspectral camouflage target detection. The model preserves fine raw spectral details and whitened PCA features while integrating multi-scale spatial context. A target-aware spatial gate is further employed to suppress background-like responses, and a center-aware decision head alleviates the dilution of small targets during patch pooling. Class-weighted loss and balanced sampling address severe class imbalance. On four land-based hyperspectral camouflage datasets with 1% training, 4% validation, and 95% testing pixels, the proposed method achieves F1 scores of 0.9438–0.9601, PR-AUC values of 0.9814–0.9938, and false-alarm rates of 0.0002–0.0003, outperforming eight representative detectors. Ablation studies confirm the complementarity of the raw and PCA branches and the effectiveness of the target-aware gate and center-aware head. In cross-dataset evaluation with only 0.5% target pixels for fine-tuning and another 0.5% for validation, PR-AUC remains between 0.8833 and 0.9288. The model requires only 0.0556 M parameters and 8.0699 M FLOPs per patch, indicating a favorable trade-off between accuracy, generalization, and lightweight deployment.