ConvNeXt-Mask2Former: A Semantic Segmentation Model for Land Classification in Remote Sensing Images

With the rapid development of remote sensing technology and deep learning, significant progress has been made in land classification of remote sensing images using semantic segmentation models. However, it still faces some deficiencies and challenges in terms of multi-scale features of images, the computational burden of high resolution images, and long-range dependencies. Therefore, in this paper, I design a semantic segmentation network based on encoder and decoder (Encoder-Decoder) architecture, using ConvNeXt as the encoder, i.e., the backbone network, and Mask2Former as the decoder, with the addition of an auxiliary segmentation module (ASM), and at the same time, an image enhancement pipeline is used to assist the training in the image preprocessing stage. Mask2Former is an architecture with good flexibility and generalization ability, which can be easily used in different visual tasks. At the same time, as a convolutional network without the addition of an attention mechanism, ConvNeXt has excellent feature extraction capabilities. I evaluate my model on the widely used high resolution image datasets Potsdam and Vaihinge and compare the results with the current mainstream semantic segmentation models, and the experimental results show that my model can be optimized for the accurate segmentation of land categories. Specifically, my model has an optimal mPA of $\mathbf{9 3 . 7 9 \%}$ and mIoU of $86.80 \%$ on the Potsdam dataset, and an optimal mPA of $82.32 \%$ and mIoU of $\mathbf{9 0 . 0 5 \%}$ on the Vaihinge dataset, which is both significantly higher than the existing models, and provide important references for land classification tasks and land use.