Hierarchical Sparsity-Guided DeepLassoNet: Interpretable PolSAR Feature Selection for Typical Land-Cover Classification in the Hunshandake Sandy Land

The Hunshandake Sandy Land is located in the transition zone between typical steppe and desertified grassland in Inner Mongolia, where grassland and sandy land are the predominant land-cover types, accompanied by scattered buildings, roads, and lakes. Its ecosystem is fragile and exhibits significant spatiotemporal dynamics. Given the need for reliable land-cover monitoring in such a fragile and dynamically changing ecosystem, spaceborne polarimetric synthetic aperture radar (PolSAR) is adopted for typical land-cover classification to support grassland ecological monitoring and desertification control. However, high-dimensional PolSAR features derived from multiple decomposition strategies usually contain redundancy and correlation, which may weaken the representation of critical discriminative scattering information. To address this issue, we propose a DeepLassoNet-based feature selection method for typical land-cover classification in the Hunshandake Sandy Land. First, a set of conventional polarimetric decomposition features derived from PolSARpro software is extracted to construct an initial high-dimensional feature pool. Next, a dual-module DeepLassoNet is designed by integrating a nonlinear spatial representation module and an explicit feature–response module, so as to learn discriminative spatial representations while preserving direct channel-level relationships between polarimetric features and classification responses. Finally, by introducing hierarchical sparsity constraints and combining gradient updating with proximal mapping, redundant and irrelevant features are progressively eliminated along a sparse feature screening path. Experimental results demonstrate that the proposed method can effectively select discriminative PolSAR features, reduce feature redundancy interference, and improve the classification accuracy and robustness in complex sandy land–grassland transition scenarios.