Optimizing Spectral Inputs for Deep Learning-Based Vegetation Destruction Detection Using Sentinel-2 Imagery

Deep learning models have been widely used for vegetation destruction detection, but their performance depends on input features in addition to sample size and model architecture. Using Sentinel-2 imagery, we constructed multiple feature combinations from visible, near-infrared, and short-wave infrared bands and several spectral indices, and evaluated them with nine deep learning models. The results show that adding extra bands to RGB input improved most models, with RGBS yielding the most robust performance improvement among the tested band combinations, consistently outperforming the RGB baseline across the majority of deep learning architectures. Spectral indices further improved most models, especially BVDI, NBR, and SVI. Some single-index or two-index inputs achieved accuracy comparable to or higher than that of RGB input, indicating their potential for information compression and band substitution. However, model responses varied, and more input channels did not guarantee stable gains. Accuracy–efficiency analysis showed that RGBS, RGB + BVDI, RGB + NBR, and selected RGBNS + index combinations achieved favorable performance at low computational cost. Overall, optimized spectral inputs improve vegetation destruction detection, but their selection should balance model architecture, accuracy, and computational cost.