DFDSCF: A Dual-Frequency and Dynamic Sparse Cross-Path Fusion Network for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification is a challenging task due to the complex spatial–spectral properties and the high-dimensional nature of the data. Existing models still face challenges in simultaneously capturing fine-grained structural details and discriminative spectral patterns, as well as in enabling sufficiently deep interaction between spatial and spectral modalities, thereby limiting their effectiveness in complex scenes. To address these challenges, this article presents a novel dual-frequency and dynamic sparse cross-path fusion network (DFDSCF), which is constructed with three key components: First, a dual-frequency multidirection enhancement module, which utilizes dual-frequency transforms and direction-aware convolutions to capture structural details, mitigating spatial detail loss; Second, a dynamic instance-level sparse channel self-attention module, which employs a multiscale complexity gating mechanism to perform instance-adaptive sparse channel selection, filtering out redundant bands and extracting discriminative spectral features; Finally, a cross-path adaptive fusion (CPAF) module, which leverages CPAF strategies to achieve the deep interaction and integration of spatial and spectral features. Comprehensive experiments on four benchmark HSI datasets demonstrate that DFDSCF achieves competitive classification performance, attaining overall accuracies of 99.92%, 99.21%, 99.90%, and 99.90% on the Pavia University, Indian Pines, WHU-Hi-HanChuan, and WHU-Hi-HongHu datasets, respectively. These results validate the effectiveness and robustness of the proposed method for HSI classification.