CGSNet: Cross-consistency guiding semi-supervised semantic segmentation network for remote sensing of plateau lake

Abstract

Analyzing the geographical information for the Plateau Lake region with remote sensing images (RSI) is an emerging technology to monitor the changes of the ecological environment. To alleviate the requirement of abundant labels for supervised RSI segmentation, the Cross-consistency Guiding Semi-supervised Learning (SSL) Semantic Segmentation Network is proposed, and it can perform high-quality multi-category semantic segmentation for complex remote sensing scenes with limited quantity of labeled images. Firstly, based on the SSL semantic segmentation framework, through the cross-consistency method training a teacher model with less annotated images and plentiful unannotated images, then generating higher-quality pseudo labels to guide the learning process of the student model. Secondly, dense conditional random field and mask hole repair are used to patch and fill the flaw areas of pseudo-labels based on the pixel features of position, color, and texture, further improving the granularity and reliability of the student model training dataset. Additionally, to improve the accuracy of the model, we designed a strong data augmentation (SDA) method based on a stochastic cascaded strategy, which connects multiple augmentation techniques in random order and probability cascade to generate new training samples. It mimics a variety of image transformations and noise conditions that occur in the real world to enhance the robustness in complex scenarios. To validate the effectiveness of CGSNet in complex remote sensing scenes, extended experiments are conducted on the self-built plateau lake RSI dataset and two public multi-category RSI datasets. The experiment results demonstrate that, compared with other state-of-the-art SSL methods, the proposed CGSNet achieves the highest 77.47% mIoU and 87.06% F1 scores with a limited quantity of annotated data.