Abstract:

Fully convolutional neural network has been widely used in semantic segmentation of remote sensing images, and the accuracy and efficiency of feature classification are high, but for remote sensing images with uneven distribution of features, the accuracy of feature classification is low. In order to improve the classification accuracy of remote sensing images, this paper enriches the input data features by adding priori knowledge methods, uses the dense link method to improve the reuse rate of features in the process of up and down sampling, combines the loss function Dice Loss that can optimize the intersection of union and the Focal Loss that can improve the accuracy of difficult classification categories as the loss function of the network, and uses the LayerScale module to accelerate the model convergence and suppress irrelevant features while emphasizing useful features, improves input, network structure and loss function of U-Net to optimize the effect of semantic segmentation. The results show that, compared with the original U-Net, the improved U-Net based on Gaofen image  dataset is improved by 0.023 3, 0.040 9 and 0.066 5 in terms of pixel accuracy, average pixel accuracy and mean intersection of union, respectively, which improves the classification accuracy of ground objects and achieves better classification effects.

Key words: deep learning, multi-feature, dense linking, Focal Loss, Dice Loss, LayerScale module, improved U-Net, semantic segmentation