AN INFORMATION EXTRACTION METHOD OF SUBURBAN INDUSTRIAL LAND USING IMPROVED DENSENET NETWORK IN REMOTE SENSING IMAGES

Aiming at the problem that traditional classification algorithm and shallow learning algorithm are not suitable for the information extraction of suburban industrial land in remote sensing images. In this paper, a method of extracting industrial land information from remote sensing images based on improved DenseNet network is proposed. Firstly, the dense neural network (DenseNet) is improved, and the SE block is embedded into DenseNet. Using the characteristics of DenseNet feature reuse and efficient information flow, the effect of SE block to extract effective remote sensing scene image features is improved. Then, the learning rate is set in the way of degenerate learning rate. At the beginning of training, the learning rate is used to accelerate training of network model, and learning rate is reduced to seek optimal solution and improve accuracy of information extraction. Finally, in order to prevent the lack of fitting caused by lack of data, translation, rotation and other operations are used to expand training data set. Experiments are carried out in Tensorflow framework to demonstrate the proposed method. The results show that the proposed method can achieve fast convergence, high extraction efficiency, and the accuracy and loss are better than other comparison methods.