CP-CRDA: central pixel guided cross-residual dual attention network for small-sample HSI classification

Transformer has achieved remarkable results in hyperspectral image (HSI) classification due to its ability to capture fine-grained details. However, under small-sample, Transformer has difficulty in achieving a balance between efficiency and accuracy. Mainly because: 1) self-attention (SA) tends to amplify abnormal features such as noise and redundant components; 2) it does not distinguish between the features of the central pixel with a real class label. To address these limitations, we propose the central pixel guided cross-residual dual attention network (CP-CRDA), which comprises: 1) pixel random transpose and flip (RTF) module, which enhances the feature invariance of the model to geometric transformations; 2) relative position embedding of the central pixel, which reduces the spatial noise interference caused by absolute positional relations; 3) an efficient channel attention (ECA) and SA cross-residual connectivity (ESCR) module, which recalibrates channel importance and mitigates the attention bias of the SA mechanism by focusing on the key features of the centre pixel; 4) finally, the fusion of multi-layer features using learnable weights. Experimental results show that under the same training environment, the overall classification accuracy of the proposed CP-CRDA model is improved by 3.83%, 0.71% and 2.91% on three datasets, respectively, compared with the current best comparison method.