MSFE-Mamba: Multiscale Frequency-Enhanced Mamba for Hyperspectral Image Classification

Hyperspectralimage classification is crucial for environmental monitoring, land-cover analysis, and other remote sensing applications. Existing classification approaches are limited by convolutional neural network capturing only short-range spatial context, transformer suffering from quadratic complexity on high-dimensional spectra, and the frequency domain being largely underutilized, restricting their effectiveness in complex scenes. To address these issues, we propose a multiscale frequency-enhanced Mamba (MSFE-Mamba) unified framework that jointly models spatial, spectral, and frequency-domain information. MSFE-Mamba employs a directional-aware adaptive multiscale fusion module to adaptively extract multiscale and direction-sensitive spatial features, a spectrofrequency modulation module to encode and modulate spectral features in the frequency domain, and a frequency-enhanced Mamba module that dynamically injects frequency-guided statistics into state–space transitions for bidirectional scanning, enabling efficient long-range dependency modeling. Extensive experiments on Pavia University, Houston University 2013, and WHU-Hi-LongKou datasets demonstrate that the proposed MSFE-Mamba achieves state-of-the-art classification accuracy, confirming the effectiveness of integrating spatial, spectral, and frequency cues in a unified framework.