Hyperspectral image classification based on adaptive spectral feature decoupling with global local feature fusion network

Deep learning-based methods are widely used in hyperspectral image (HSI) classification and have achieved excellent classification performance. However, hyperspectral data from different categories exhibit strong nonlinear coupling, which results in low spatial distinguishability between samples from different categories. Under the condition of limited sample size, how to extract spectral-spatial features and reduce the coupling of hyperspectral data from different categories is the key to achieving high-precision classification. Some methods based on Convolutional Neural Networks (CNN) tend to focus on local information within hyperspectral cubes. Transformers have excellent performance in modeling global dependencies between sequences. To solve the above problems, this paper proposes a global local feature fusion network (GLF2Net) for hyperspectral classification. To effectively integrate global information, this method introduces frequency domain statistical methods into the field of hyperspectral image classification. Firstly, this paper utilizes Fast Fourier Transform (FFT) to obtain frequency domain information from HSI data. Then, an improved adaptive 13-dimensional frequency domain statistical feature is applied as a supplement to the information after Principal Component Analysis (PCA) dimensionality reduction. To fully capture local-global hyperspectral features from HSI data, a dual-branch structure with a Transformer encoder Convolution Mixer Branch (TCM) and a CNN Branch is designed. Through extensive experiments on real HSI datasets, it is proven that the classification performance of GLF2Net is superior to several classic HSI classification methods.