Remote Sensing Change Detection Method Based On DAUNet++ And Multi-Scale Fusion Model

Change detection technology has a wide range of applications in the field of remote sensing,with the primary goal of automatically detecting changes in remote sensing images taken at different times.These changes may arise from various causes,such as natural disasters,urban development,and land use.Traditional methods employ hand-crafted feature extractors and classifiers,while deep learning-based approaches can automatically extract features and possess strong generalization capabilities.However,existing research still has shortcomings in extracting differential features between bi-temporal remote sensing images.This thesis initially applies a series of data preprocessing techniques aimed at expanding the training samples and reducing memory requirements.These preprocessing techniques provide a high-quality and diverse data foundation for training and prediction of remote sensing change detection models.Subsequently,the existing issues in deep learning models are investigated and analyzed,and the models are improved and optimized.The main research work includes the following:(1)To achieve better change detection results and avoid setting unreasonable weight parameters manually,this thesis improves the traditional Focal Loss function by introducing batch class weight parameters and combining it with the Dice Loss function,proposing a dynamically adjusted hybrid loss function.Comparative experiments demonstrate that the improved Focal Loss and dynamically adjusted hybrid loss function effectively enhance the model’s detection accuracy and segmentation performance.(2)In order to more accurately extract the changed areas in bi-temporal images,this thesis improves the UNet++ network and proposes a remote sensing change detection algorithm combining difference-learning and gated attention mechanisms(Difference-learning Attention UNet++,DAUNet++).Firstly,a difference-learning module is constructed to address the UNet++ network’s insufficient ability to extract change information from remote sensing images,aiming to obtain multi-level,different scale difference feature information during the downsampling process.Considering the information loss during the downsampling process,a gated attention mechanism is constructed to strengthen the difference feature information.Finally,the encoder of the UNet++ network is improved by combining it with a twin network.Comparison experiments with UNet++ and other state-of-the-art algorithms demonstrate the effectiveness of the improvements,outperforming other algorithms and achieving better remote sensing change detection results.(3)To overcome the limitation of a single model in comprehensively analyzing differential information in complex scenes,this thesis proposes a multi-scale fusion model based on ensemble learning.Firstly,by combining convolutions with different dilation rates,an Atrous Spatial Pyramid Pooling(ASPP)is implemented to extract feature information at different scales in complex scenes.Furthermore,a channel attention mechanism is introduced to perform weighted learning of different base learners.Finally,a multi-scale fusion module is constructed using ASPP and the channel attention mechanism to learn the expression capabilities of each model at different scales.This module effectively enhances the performance of the model in their respective proficient scale feature information.Experimental results demonstrate that the multi-scale fusion model effectively addresses the deficiency of a single model in different scales and fully leverages the strengths of each model,thereby improving the accuracy of remote sensing change detection.