Small sample remote sensing image segmentation based on multiscale feature fusion

Aiming at the problems of high cost of drawing remote sensing image labels and low detection accuracy of remote sensing image under the condition of limited training samples in actual scenes, the deep pyramid attention network (DPA-Net) combining multiscale feature fusion and attention mechanism was proposed to extract buildings and roads from small sample remote sensing images by integrating transfer learning method. Due to limited training samples containing information is limited, first of all, on the basis of DeeplabV3+ network architecture, two sources of low-level features are added to make full use of spatial information of low-level features, and the attention mechanism is used to obtain rich context information and enhance the learning ability of the target channel, reduce the response ability to other targets and noises, and to improve the poor detection effect of the model on small samples. Finally, the transfer learning method using open remote sensing image dataset and small sample dataset for joint training reduces the impact of too few training samples on network learning performance. Experimental results show that the accuracy of the proposed method is improved by 3.69% and the annotation cost can be saved by 1/2. © 2022, Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.