Answers are grounded in this paper and cite their sources.
Multisource remote sensing data, such as hyperspectral imagery, synthetic aperture radar, and light detection and ranging, provide complementary spectral, structural, and geometric information for land-cover classification and semantic segmentation. However, effective fusion remains challenging because of the large modality gap across remote sensors and the need to jointly model local details and global contextual structures. Existing deep learning methods still face difficulties in balancing cross-modal interaction, representation capability, and computational efficiency. Although Kolmogorov–Arnold networks (KANs) offer strong nonlinear modeling ability, their heavy parameterization and unstable computation limit practical use in remote sensing. In response, we introduce spectral-KAN fusion network (SKANet), a unified single-stream framework that synergizes frequency-domain global perception with learnable high-order nonlinearity for multisource remote sensing interpretation. The core of SKANet is the SpectralKAN module, which features a dual-branch topology, i.e., a frequency gating branch utilizing fast Fourier transform to capture global structural information, and a KAN branch employing learnable B-spline transformations to model complex semantics. To mitigate task-specific dimensional disparities, we incorporate adaptive configurations, such as a channel–spectral approach for categorization to manage dimensionally reduced inputs, and a spatial–spectral approach for delineation to retain detailed spatial integrity. In addition, an adaptive multimodal integration component is engineered to flexibly adjust sensor-derived attributes during initial processing phases. Extensive experiments on three benchmark datasets (Berlin, MUUFL, and Houston2018) demonstrate that SKANet establishes state-of-the-art performance, achieving overall accuracies of 73.70%, 89.24%, and 67.84%, respectively, demonstrating the effectiveness of SKANet for multisource remote sensing classification and segmentation.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: