Land cover type extraction based on the combination of UNet++and BP neural network
In the extraction process of soil erosion-oriented remote sensing image ground surface object,there are serious"pepper and salt"phenomenon and insensitivity to small surface objects.In order to solve this problem,a two-branch neural network deep learning method was constructed based on back propagation(BP),UNet++neural network and object-oriented thought.The method can be used to adaptively extract soil erosion-oriented land cover types in different complexity regions.Taking the GF-2(Gaofen-2 satellite)remote sensing image of Shenzhen as an example,object-oriented segmentation and optimal feature index subset screening were carried out.The segmentation result and the optimal feature subset were used as the input data of the two-branch neural network to extract land cover types.The proposed method was compared with the object-oriented k-nearest neighbor(KNN),BP and UNet++neural network algorithms.The results show that the overall accuracy of the two-branch neural network for ground object extraction reaches 91.73％,and the comprehensive evaluation index F measure of all kinds of ground object recognition effect is more than 0.8.It can effectively inhibit the phenomenon of"pepper and salt"and the misclassification of small ground objects,and the evaluation results are better than those of the control group.
BP neural networkUNet++object-orientedfeature selectionextraction of land cover types
NETL
NSTL
万方数据
维普
