To address the problem of gradient vanishing in recurrent neural networks and the limited receptive field of traditional convolutional neural networks, this paper proposes a spectral–spatial feature extraction method that incorporates multi-scale convolutional filters. The method consists of two main components: spectral feature extraction and spatial feature extraction. In the spectral feature extraction stage, a bidirectional long short-term memory (Bi-LSTM) network is combined with a band-grouping strategy. This approach mitigates the gradient vanishing issue caused by excessive network depth. In the spatial feature extraction stage, multi-scale convolutional filters are introduced based on a convolutional neural network (CNN), allowing the model to capture both fine details and global structural information. Additionally, shallow features are fused with deep features to further enhance classification performance. Experimental results on two datasets demonstrate that the proposed method effectively improves classification accuracy.