Hyper-LKCNet: Exploring the Utilization of Large Kernel Convolution for Hyperspectral Image Classification

Recently, transformers have garnered significant attention due to their exceptional capability to capture long-range dependencies in data. A critical factor contributing to their superior performance is their ability to operate over large receptive fields. As such, a natural question arises as to how to expand the receptive fields in convolutional neural networks to achieve the superior performance comparable with that of transformers. Large kernel convolution provides the inspiration for the above issue. To explore the potential of large kernel convolution, we propose a hyperspectral image (HSI) classification algorithm in this article that utilizes a large kernel convolution module combined with multiscale coattention and an adaptive geometric feature (AGF) classifier, named Hyper-LKCNet. By integrating this feature enhancement module, our method effectively adjusts the contributions of various spectral and spatial features, ensuring that the network captures critical but easily overlooked information across both dimensions and improving the performance to classify HSI. The AGF classifier, derived by neural collapse theory, alleviates the sample imbalance problem and incorporates the label smoothing focal loss function to enhance generalization ability. Extensive experiments on four HSI datasets demonstrate that the proposed method outperforms the state-of-the-art approaches. In addition, our algorithm maintains a low parameter count and reduced floating point of operations.