Remote Scene Image Classification Based on Vision-LSTM Model

As the latest development of extended long short-term memory (xLSTM) architecture in the field of image classification, it has shown promising results compared to traditional convolutional neural networks (CNN) and Vision Transformer (ViT) models. Inspired by this progress, this paper proposes a remote scene image classification method based on Vision-LSTM. To reduce the impact of intra-class diversity and inter-class similarity, we improved a label smoothing loss function and combined it with a memory bank to further improve accuracy in complex scenes. Experiments conducted on three different remote sensing datasets show that the method performs well in classification results. It improved by 0.48% on the UCM dataset, and 0.13% on the NWPU dataset.