Research And Application Of Railway Environmental Change Detection Method Based On Multi-source Remote Sensing Images

China’s railway construction is in a stage of rapid development.With the continuous expansion of railway construction scale,a series of environmental problems such as illegal land occupation and ecological damage inevitably exist in the construction process.Traditional ground survey methods are difficult to reflect the construction status in a timely and comprehensive manner.Change detection algorithms based on remote sensing images can be used to analyze and determine changes in the monitoring area,making up for the shortcomings of ground surveys.Compared with traditional methods,change detection methods based on deep learning can better analyze complex surface conditions and improve the accuracy of change detection.However,existing change detection models have problems such as fuzzy detection result boundaries and loss of small targets.Moreover,large deep learning network structures are complex,and the number of parameters is huge,requiring significant training resources.To address these issues,this article focuses on the following research content:(1)Designing an end-to-end change detection model based on auxiliary information.To reduce the influence of color feature differences of the same type of ground objects in multi-source images on detection results,straight features that conform to artificial building characteristics are first extracted from high-resolution images as prior knowledge and fused with the original images,followed by change detection.Secondly,the network backbone adopts a Siamese UNet++ network to fuse multi-scale features and introduce the difference between dual-time images into the skip connection process to guide the model’s attention to the change information between images.Finally,the self-attention mechanism is used to improve the multi-output fusion strategy of UNet++ to solve the semantic differences between different scale outputs,refine detection boundaries,reduce the loss of small targets,and achieve an F1 score accuracy of 90.4%.(2)Designing a lightweight change detection model based on attention mechanism.To improve the real-time and effectiveness of the detection model,the first four layers of the lightweight and efficient Efficient Net B4 are used as the feature extraction backbone network to design a UNet-style change detection model.In the decoder stage,the use of depthwise separable convolution improves the decoding block,reducing the number of parameters and computations.In the skip connection process,group convolution is used to improve the feature fusion method and introduce CBAM attention to refine the feature representation.The final model reduces the parameter amount to 1.06 M and the computation amount to 2.16 G while ensuring a certain accuracy,and its effectiveness is verified in public datasets and railway datasets.(3)Designing and implementing a railway environmental monitoring system.Using an open-source platform for secondary development,the improved algorithm is encapsulated to implement railway construction supervision.The environmental monitoring system includes functions such as data preprocessing,information extraction,multi-source image change detection,railway construction progress assessment,temporary engineering land occupation status and vegetation cover recovery detection,thematic map generation,and report output.