EGAS-Net: Entropy-Guided Adaptive Selection Network for Hyperspectral–LiDAR Classification

Abstract

The complementarity of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data provides significant advantages in land-cover classification. In current methods, the cropping size is typically determined by empirical selection, which readily results in the redundancy or loss of land-cover feature statistics. Information entropy, a key measure of data uncertainty and richness, plays a crucial role in selecting feature channels with abundant texture cues and strong discriminability. Motivated by this, this article proposes an entropy-guided adaptive selection network (EGAS-Net) for hyperspectral–LiDAR classification. With information entropy as the core, this method adaptively selects the optimal size for each land cover by calculating spatial information entropy and spatial–spectral information entropy to maximize information content. Then, the spatial and frequency information entropies of HSI and LiDAR data are obtained after cropping. Adaptive weighting is implemented based on the information content ratio of each entropy feature to establish a cross-modal spatial–frequency information interaction mechanism. During the feature fusion stage, the final features are generated through entropy feature weighted coupling to fully exploit the complementary value of HSI–LiDAR data. Experimental results on the Houston 2013, MUUFL, Augsburg, and Yancheng datasets show that the proposed method outperforms other empirical size-based networks, achieving higher classification scores. The code of this project will be available at https://github.com/ZhaoYuQing01/EGASNet