High-resolution Remote Sensing Image Change Detection Methods Based On Deep Learning

Remote sensing images are the main means of earth observation.The task of remote sensing image change detection plays an important guiding role in vegetation coverage,urban planning,and emergency and disaster relief.With the increasing resolution of remote sensing images,traditional change detection methods can no longer meet the needs.Therefore,thesis uses deep learning technology to study the high-resolution remote sensing image change detection methods from the following three aspects:1.The method of remote sensing image change detection based on metric learning.Thesis constructs a siamese change detection network based on metric learning.The multi-level features of two-phase remote sensing images are extracted through the siamese network,then increase the distance between changed pixel pairs and decrease the distance between unchanged pixel pairs through metric learning.In order to balance the details and semantic information in the feature map,atrous convolution and feature pyramid network are introduced,which effectively solve the problem of unclear boundaries of changing objects.In addition,thesis introduces a class-balanced contrastive loss function to adaptively adjust the penalties for positive and negative samples,so as to obtain better change detection results.2.The method of remote sensing image change detection based on Transformer spatial-temporal information enhancement.Since the metric learning method does not consider the relationship with the surrounding pixel pairs when classifying,it is prone to a lot of noise.To solve the above problems,thesis uses a shallow neural network as a classifier to directly convert the pixel values in the feature map into category prediction probabilities to complete the dense classification of pixels.In order to make up for the defect that the convolution operation can only extract local feature dependencies,thesis introduces the Transformer spatial-temporal information enhancement module,which represents the pixels in the feature map with compact semantic tokens.It models the spatialtemporal contextual semantics of deep feature maps,and removes redundant informations to get more robust features.3.The method of change detection based on multi-scale differential feature fusion.Considering the change of target objects as a kind of semantic information,the change detection task can be regarded as a kind of semantic segmentation task.The encoder part of the U-Net network is changed to a siamese network structure,and then the difference operation is used to fuse the multi-scale features of the bi-temporal remote sensing images to obtain the multi-scale differential feature between different time-phase remote sensing images.Finally,the final change detection result is obtained by decoding the fused multiscale difference features.In thesis,an attentional spatial-temporal feature fusion operation is introduced to extract rich spatial-temporal contextual information from bi-temporal remote sensing image feature maps to improve change detection results.