Scene recognition of remote sensing image with residuals

Scene recognition of remote sensing images can be used for weapon guidance, forest fire prevention, etc., and has wide applications in environmental testing, urban planning and other fields. The development of deep learning technology has provided new possibilities for improving the accuracy of scene recognition in remote sensing images. Three network models for scene recognition in remote sensing images are designed based on residual networks and convolutional neural networks (CNNs). Taking the UCMerced_LandUse dataset as an example, the network depth is determined through pre-training. In the UCMerced_LandUse test set, the accuracy of the residual network model is 85% under 100 training cycles, which is about 50% higher than the residual network without BN layer and about 20% higher than the CNN model. By increasing the training cycle, the accuracy of the residual network model on all datasets reaches 96%. The residual network model designed in this paper does not significantly improve the accuracy in remote sensing image scene recognition tasks, but the model has the advantage of being lightweight. The depth of only 8 layers of residual blocks can be comparable to the accuracy of typical networks such as ResNet50, VGG16, MobileNet etc. The application prospects of lightweight deep learning technology in the field of remote sensing image processing are very broad, providing strong support for the further development of this field.