DLMBMT: A Deep Learning Method Based on the Multi-Scale Technique for Hyperspectral Image Classification with Limited Training Samples

In hyperspectral imaging (HSI) classification tasks, Convolutional Neural Network (CNN) and 3-D Swin Transformer (3DST) methods have shown strong HSI classification performance, yet CNN is limited by local receptive fields that restrict global–spatial and multi-scale modeling, while 3DST may underutilize continuous spectral characteristics. With limited training samples, deep learning models also tend to overfitting, degrading HSI classification performance. The multi-scale feature processing technique alleviates this by capturing both global semantics and local spatial details. Thus, we propose a deep learning method based on the multi-scale technique for HSI classification using limited training samples (DLMBMT), which combines the spectral–spatial HSI Feature Extraction Module (SHFEM), Improved 3DST (I3DST), and the Multi-Scale Module with Residual Framework (MMRF) methods. SHFEM can effectively perform spectral–spatial HSI feature extraction. I3DST performs HSI spatial feature processing. In the MMRF, the U-Net structure facilitates multi-scale HSI spectral feature processing. Additionally, in the MMRF, the limited training sample issue is slightly eased by two semantic feature enhancement modules (SFEMs), based on the Mamba structure, and a spectral attention mechanism module (SAMM), based on the channel attention module. DLMBMT employs the SHFEM, I3DST, the MMRF and CONCAT operation to enable spectral–spatial HSI feature processing. Experiments were conducted on four datasets. The performance of DLMBMT was evaluated using overall accuracy (OA) and other standard metrics. The HSI classification results of DLMBMT are good. For example, the OA value on the HU dataset is 97.43%. The results demonstrate that DLMBMT and some state-of-the-art (SOTA) methods have a certain level of competitiveness with each other, demonstrating DLMBMT’s effectiveness and potential for HSI classification tasks.