GLN-LRF: global learning network based on large receptive fields for hyperspectral image classification

Deep learning has been widely applied to high-dimensional hyperspectral image classification and has achieved significant improvements in classification accuracy. However, most current hyperspectral image classification networks follow a patch-based learning framework, which divides the entire image into multiple overlapping patches and uses each patch as input to the network. Such locality-based methods have limitations in capturing global contextual information and incur high computational costs due to patch overlap. To alleviate these issues, we propose a global learning network with a large receptive fields network (GLNet) to capture more comprehensive and accurate global contextual information, thereby enriching the underlying feature representation for hyperspectral image classification. The proposed GLNet adopts an encoder-decoder architecture with skip connections. In the encoder phase, we introduce a large receptive field context exploration (LRFC) block to extract multi-scale contextual features. The LRFC block enables the network to enlarge the receptive field and capture more spectral-spatial information. In the decoder phase, to further extract rich semantic information, we propose a multi-scale simple attention (MSA) block, which extracts deep semantic information using multi-scale convolution kernels and fuses the obtained features with SimAM. Specifically, on the IP dataset, GLNet achieved overall accuracies (OA) of 98.72%, average accuracies (AA) of 98.63%, and Kappa coefficients of 98.3%; similar improvements were observed on the PU and HOS18 datasets, confirming its superior performance compared to baseline models. The experimental results demonstrate that GLNet performs exceptionally well in hyperspectral image classification tasks, particularly in capturing global contextual information. Compared to traditional patch-based methods, GLNet not only improves classification accuracy but also reduces computational complexity. Future work will further optimize the model structure, enhance computational efficiency, and explore its application potential in other types of remote sensing data.