Leveraging adaptive spatial shifts multi-layer perceptron with parameterized spatial offsets for hyperspectral image classification

The architectures based on Multi-Layer Perceptron (MLP) have attracted great attention in hyperspectral image (HSI) classification recently, due to their simplified and efficient architectures. However, such architectures are qualified by the rigid positional relationships between weights and feature elements, inhibiting their capacity to effectively extract diversified features. To address these challenges, An adaptive spatial-shift MLP (AS2MLP) is presented to dynamically modify spatial features by parameterizing learnable spatial offsets. In this way, the AS2MLP can facilitate sample-specific spatial shifts, aligning spatial structures more effectively. Then, An innovative adaptive spatial-shift block (AS2block) is designed to adaptively shift spatial features along distinct spatial axes, enabling the extraction of diversified features separately. It also implements a re-weighting strategy to mitigate redundant features. Building on this foundation, the proposed adaptive spatial-shift network (AS2Net) is for HSI classification. The dual-path AS2Net employs AS2blocks and MLPs for channel mixing, facilitating an adaptive integration of dynamic spatial contextual information dispersed across a range of spectra. The effectiveness of this model is demonstrated using five widely used HSI datasets.