Research On Change Detection Method Of Remote Sensing Images Based On Transformer

Remote sensing image change detection has always been one of the research hotspots in remote sensing.Its main task is to identify the state difference by detecting ground objects in the same area at different times.This is very important for urban planning,environmental monitoring,agricultural survey,disaster assessment and map revision,etc.This article analyzes the research background and current situation of remote sensing image change detection,summarizes the difficulties existing in the current research and makes improvements,the work is summarized as follows:1.In order to solve the problem of "pseudo changes"(that is,the same object or area will show different colors or features due to different lighting conditions in remote sensing images collected at different times),the article combines convolution and Transformer to propose a multi-scale difference feature enhancement network.The difference enhancement module in this network includes a multi-scale difference enhancement encoder and a Transformer decoder.They are applied to features at different scales to establish the long-range relationship of pixels,enhance the representation of semantic changes and the information transfer between features at different scales,and reduce "pseudo changes".To alleviate the sample imbalance problem,a combination of multiple losses is introduced.Finally,this paper verifies the effectiveness of the proposed model through exhaustive experiments and it achieves better change detection results in remote sensing images than the state-of-the-art models.2.In order to more accurately identify the types of features in the change area in the semantic change detection task,the correlation between the two tasks of change detection and semantic classification is established.This paper proposes a dual-branch information aggregation network that combines the two branches of change detection and semantic segmentation to identify changed regions of images and accurately identify their detailed semantic categories.Based on Transformer,a dual-branch information aggregation module is designed,which obtains richer semantic features with local and global information from bitemporal images and establishes the correlation of bi-temporal features,effectively enhancing the representation of changing features to identify complex The feature type of the changing area in the scene.Finally,a multi-task loss function is designed to jointly supervise model training to improve network performance.