Remote Sensing Building Change Detection Based On Deep Learning

Remote sensing building change detection technology is an important remote sensing technology for identifying changes in buildings in the same geographical location in bitemporal images.It has a wealth of application scenarios.Change detection technology based on deep learning has shown good results in high-resolution datasets,and can reduce significant labor costs and time consumption.However,the sparsity of building samples in datasets and inaccuracies in manual annotation greatly affect the progress of deep learning in the field of remote sensing building change detection.Therefore,exploring suitable strategies to address these issues is a research hotspot in this field.After analyzing the problems of related deep learning algorithms,this paper proposes two strategies to complete the task of remote sensing building change detection.On one hand,this paper presents an efficient Siamese full-scale connected network to fully extract building features from sparse samples.The network includes a feature difference augmentation module composed of long-short term memory networks and a coordinated attention mechanism to assist in capturing feature information,and makes use of full-scale skip connections to capture low-level detail and high-level semantic information at various scales.To further learn feature representation from full-scale aggregate feature maps,the network includes depth supervision.The output of each layer is linked to a hybrid loss function,which helps to accurately identify changing buildings,especially buildings that appear in different proportions in the bitemporal images.The hybrid loss function is designed by multi-tasking sparse sample recognition and building edge recognition,which can effectively reduce the negative effects brought by sample sparsity and label inaccuracy.In addition,an automatic weight module is used in the training process to automatically find the optimal weight of the hybrid loss function and adjust it dynamically.On the other hand,in order to fundamentally solve the problems of sample sparsity and label accuracy in the datasets,this paper adopts the instance augmentation method to optimize the datasets.First,GAN is trained to generate buildings of similar size,and color transfer is used to make the generated buildings have the color style of the target dataset buildings.Then,the minimum fit ellipse clipping method is used to extract the effective information of the building and its edges.Finally,the target bitemporal image suitable for optimization is selected,and the virtual buildings are added harmoniously by Gaussian blending method.Compared with the data enhancement method of directly generating images,the instance augmentation method can significantly reduce the influence of sample sparsity and label inaccuracy while not increasing the time cost of network training,which has great advantages.Extensive comparative and exhaustive ablation experiments on the LEVIR-CD and WHU datasets show that the proposed method can significantly reduce the negative effects of sample sparsity and label inaccuracy,and is superior to several other state-of-the-art methods.