HSI–LiDAR Joint Classification via Progressive Spatial–Spectral–Frequency Fusion Learning

Hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data joint classification leverages multi-modal information to improve classification performance. However, existing methods often suffer from inadequate feature representation due to their reliance on spatial-spectral domains while ignoring complementary frequency-domain cues, limiting their ability to capture cross-modal variations. Moreover, even when frequency information is incorporated, most approaches fail to achieve deep cross-domain interaction, either treating spatial and frequency features in isolation or relying on single-stage fusion, which hinders effective modality gap bridging. To address these challenges, we propose a novel progressive spatial-spectral-frequency fusion learning (PS2F2L) network, a lightweight yet powerful framework for HSI and LiDAR classification. Our method employs a multi-stage progressive fusion strategy to hierarchically integrate multimodal features, mitigating modal conflicts while capturing discriminative features. In the first stage, dual branches—spatial-spectral feature learning (S2FL) and spatial-frequency feature learning (SF2L)—jointly extract features from complementary domains, overcoming single-domain limitations. The S2FL branch combines 1D/2D convolutions to model spectral-spatial relationships, while SF2L utilizes discrete wavelet transforms to capture spatial-frequency patterns. In the second stage, an interactive spatial-spectral-frequency fusion module enhances feature discriminability by promoting deep information exchange between spatial-spectral and spatial-requency representations. Finally, adaptive decision-level fusion refines classification by consolidating multi-domain predictions. Extensive experiments on three public datasets demonstrate the superiority of PS2F2L, validating its effectiveness in achieving robust and accurate multimodal classification.