A multi-level deformable gated aggregated network for hyperspectral image classification

Deep learning has dominated hyperspectral image (HSI) classification due to its modular design and powerful feature extraction capabilities. Recently, a modern macro-architecture-based framework with high-order feature interactions has been proposed, inspiring the design of HSI classification models. As a spatial mixer in a modern macro-architecture, the high-order feature interaction facilitates the aggregation of discriminative information by gated mechanisms with standard convolutions. However, the homogeneous operators of standard convolution are challenging to consider the interaction information of different spatial locations. Furthermore, the macro architecture designed for RGB image classification tasks performs poorly with limited training samples. To address these issues, we propose a multi-level deformable gated aggregated network (MDGA) for HSI classification. First, we present axis decomposition convolutions with deformable sampling for adaptive feature interactions to extract invariant features, suppressing the redundant and mutually exclusive information. Then, we introduce the inverted residual block into the macro architecture, which allows its channel mixer to extract spatial features, reducing the depth and complexity of the model. Extensive experiments conducted on four widely used HSI datasets demonstrate that the proposed MDGA effectively mitigates the interference of redundant information and achieves satisfactory classification accuracy.