Research And Application Of Satellite Image Object Extraction Method For Railway Construction Project Supervision

The improvement of satellite image resolution and the rapid development of deep learning have led to the gradual application of deep learning-based satellite image object extraction algorithms in various domains such as land use classification,urban planning,and environmental monitoring.Leveraging the advantages of satellite imagery,we construct a railway construction project dataset based on satellite images to address the challenges of difficulty,low efficiency,and incomplete coverage in railway construction project supervision.We propose an object detection and segmentation model algorithm tailored for railway engineering construction supervision tasks,and integrate it with practical requirements to design and develop a remote sensing interpretation system that serves as a paradigm for the combination of satellite imagery and deep learning applications.This thesis presents a study conducted on the preprocessing of satellite imagery,dataset construction,railway construction project signage extraction algorithm,and remote sensing interpretation system,based on Gaofen-2 satellite image data.The main contributions and research aspects are outlined as follows:(1)Construction of a remote sensing image dataset for railway construction project supervision.Firstly,utilizing radiometric calibration and atmospheric correction preprocessing methods,a satellite image sample library containing 100 scenes with a resolution of 0.8 meters is constructed,ensuring image quality.Secondly,a railway construction project dataset named UJN-Traffic is established using visual interpretation methods for object detection and segmentation in high-resolution imagery.This dataset includes 100,000 key process markers of four types: bridge piers,roadbeds,subgrades,and bridge decks.Taking bridge pier samples as an example,the effectiveness of this dataset is validated through training and detection verification using three mainstream object detection algorithms.(2)Development of small object detection algorithms for railway construction projects.To address the detection of small-sized samples in the dataset,two algorithms are proposed: a remote sensing small object detection algorithm based on context information and a satellite image small object detection algorithm based on multiscale features and a cascade structure.The remote sensing small object detection algorithm based on context information first utilizes image preprocessing with median filtering to solve noise interference issues,and then employs a feature pyramid based on context information to enhance detection performance.The satellite image small object detection algorithm based on multiscale features and a cascade structure first extracts fine-grained features using a multiscale feature extraction network,and then selects a spatial information fusion feature pyramid and a cascade structure detection network to mitigate overfitting issues.The effectiveness of the algorithms is demonstrated through experimental comparisons.(3)Development of an object segmentation algorithm for railway construction projects.To address the issue of texture feature extraction for target information in the dataset,a satellite image object segmentation algorithm based on sparse self-attention is proposed.Firstly,improvements are made to the backbone network to handle cross-modal and multiscale information.Secondly,sparse matrices are utilized to map features to a high-dimensional feature space,refining spatial features.The proposed algorithm is compared and analyzed against mainstream object segmentation algorithms in the field,demonstrating good detection and segmentation performance on the proposed dataset.(4)In response to the satellite image application requirements for railway construction project supervision,a remote sensing interpretation system is developed.Based on satellite image management and statistical analysis,the proposed object detection and segmentation algorithms are integrated into the system.Through visual analysis,the system provides current construction progress of railway construction projects,assisting in the detection of engineering progress in conjunction with manual inspection.