Abstract:Aiming at the problems of high cost of manual labeling of river data samples in remote sensing images and difficulty in obtaining a large number of them, as well as poor effect of network extraction of river image edge details, a self-supervised comparative learning method was proposed to use a large number of unlabeled remote sensing river image data for encoder pre-training, and a small amount of label data was used to fine-tune the encoder after pre-training. Meanwhile, a semantic segmentation network based on non-uniform sampling was used in the codec structure. Self-supervised comparative learning can use a large number of unlabeled data for pre-task model training, and only a small amount of label data was needed to fine-tune the downstream river extraction task model. The non-uniform sampling method can obtain clear boundary information between different categories in the image and details in the same category by intensive sampling in the high frequency region and sparse sampling in the low frequency region, thus reducing the redundancy of the model. Experiments on river data sets showed that the pixel accuracy, intersection over union and recall rate of the pre-trained network can reach 90.4%, 68.6% and 83.2%, respectively, when 360 labeled datas was used to fine-tune the network, which was comparable to the performance of the supervised AFR-LinkNet network trained with 1200 labeled datas. After fine-tuning with all data labels, the pixel accuracy, intersection ratio and recall rate of the network reached 93.7%, 73.2% and 88.5%, respectively. Compared with AFR-LinkNet, DeepLabv3+, LinkNet, ResNet50 and UNet networks, the pixel accuracy was increased by 3.1, 7.6, 12.3, 14.9, 19.8 percentage points, the intersection over union was increased by 3.5, 8.7, 10.5, 16.9, 24.0 percentage points, and the recall rate was increased by 2.1, 4.8, 6.7, 9.4, 12.9 percentage points, respectively. The effectiveness of the model to accurately extract rivers from river images was verified. This algorithm model was of great significance for solving the lack of a large number of labeled data and analyzing the distribution of rivers in cold and arid regions, water disaster warning, rational utilization of water resources and agricultural irrigation development.
