Research On The Change Detection Method Of High-resolution Remote Sensing Images Based On Attention Mechanism

Over the past three decades,China’s transportation industry has made significant breakthroughs and achieved historic changes,laying a solid foundation for realizing the leap from a large transportation country to a strong transportation country.At present,the problems of insufficient supply of high-quality spatiotemporal information and insufficient capability of high-level spatiotemporal analysis are still the major challenges facing the high-quality development in transportation.Change detection technology is the process of determining the change of surface facilities at different times by analyzing the images captured at different times.The technology has obvious advantages in road discrimination,helps to obtain more accurate road spatial location information,has the characteristics of large scope,high accuracy,no manual field survey,can quickly realize the update and verification of road network,and provides important data support for road planning and operation maintenance.With the rapidly developing computer hardware and increasingly sufficient data reserves,the training of deep neural networks has become possible.Deep learning is an algorithm for learning representative features from a set of data in layers and is a special kind of machine learning method.In recent years,deep learning has received extensive attention in various research areas of computer vision,including the interpretation of remote-sensing images.Based on powerful deep-learning models,the interpretation of remote-sensing images has made rapid progress.However,to train large models,the amount of existing data for highresolution change detection is significantly insufficient,and there are more interference factors in change detection tasks than other computational vision tasks,such as illumination,cloud cover,and shooting angle differences.How to suppress irrelevant change type interference is a problem that must be solved for the change detection task.In addition,the remote sensing image scenes are complex and contain feature types with large scale variations,and the fixed perceptual fields in the deep learning model cannot meet the task needs.This paper focuses on the change detection method of dual-temporal high-resolution remote sensing images,based on the basic theoretical knowledge of deep learning,to realize the application of remote sensing interpretation technology in the field of transportation.The main works in this paper are as follow:(1)To address the problem of large changes in the size of change regions,this paper designs a change detection network with adaptive perceptual field features,which adaptively adjusts the perceptual field size according to the actual scene.Based on the Siamese U-Net structure,an efficient adaptive perceptual field convolution module is designed,which preserves the deep semantic information and ensures the computational efficiency of the network.After comparison,the method proposed in this chapter achieves significant results in terms of computational accuracy and efficiency,and shows excellent performance on both datasets.(2)To address the complex problem of interference factors for change detection.In this paper,a differential feature fusion module is designed for the effective suppression of interference factors.Based on the Siamese FCN connection method,Conv Ne Xt is used as the basic unit for feature extraction,and the difference fusion module is used for feature fusion in each layer.The training method of the Transformer is referred to during the training and compared with the Swin-Transformer structure.The superior performance of this paper’s method for the change detection task is verified on two datasets...