Abstract: To address the challenge of requiring a large amount of manually annotated data for semantic segmentation of remote sensing images using deep learning, a method based on self-supervised learning is proposed. Firstly, to simultaneously learn the global and local features of remote sensing images, a self-supervised learning network structure called TBSNet (Triple-Branch Self-supervised Network) is constructed. This network comprises an image transformation prediction branch, a global contrastive learning branch, and a local contrastive learning branch. The contrastive learning part of the network employs a novel data augmentation method to simulate positive pairs of the same remote sensing images under different weather conditions, enhancing the model's performance. Meanwhile, the model integrates channel attention and spatial attention mechanisms in the projection head structure of the global contrastive learning branch, and replaces a fully connected layer with a convolutional layer in the local contrastive learning branch, thus improving the model's feature extraction ability. Secondly, to mitigate the high computational cost during the pre-training phase, an algorithm optimization strategy is proposed using the TracIn method and sequential optimization theory, which increases the efficiency of pre-training. Lastly, by fine-tuning the model with a small amount of annotated data, effective semantic segmentation of remote sensing images is achieved even with limited annotated data. The experimental results indicate that with only 10% annotated data, the overall accuracy (OA) and recall of this model have improved by 4.60% and 4.88% respectively, compared to the traditional self-supervised model SimCLR (A Simple Framework for Contrastive Learning of Visual Representations). This provides significant application value for tasks such as semantic segmentation in remote sensing imagery and other computer vision domains.

Wenbo Zhang and Achuan Wang, “Research on Semantic Segmentation Method of Remote Sensing Image Based on Self-supervised Learning” International Journal of Advanced Computer Science and Applications(IJACSA), 14(8), 2023. http://dx.doi.org/10.14569/IJACSA.2023.0140855

@article{Zhang2023,
title = {Research on Semantic Segmentation Method of Remote Sensing Image Based on Self-supervised Learning},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2023.0140855},
url = {http://dx.doi.org/10.14569/IJACSA.2023.0140855},
year = {2023},
publisher = {The Science and Information Organization},
volume = {14},
number = {8},
author = {Wenbo Zhang and Achuan Wang}
}