Research On Multi-task Learning Change Detection Method And Application Based On Spatio-temporal Features Fusion

As a core subject in Earth observation,change detection has garnered widespread attention due to its ability to identify alterations in objects or phenomena over time.The advent of deep learning has led to significant advancements in the performance and applications of change detection.However,existing algorithms still face numerous shortcomings due to factors such as image data sources,lighting conditions,and model generalization.This study establishes a deep change detection model that considers spatiotemporal feature information,based on the comprehensive utilization of multisource satellite images to construct a relatively complete change detection dataset.The specific research content is as follows:(1)A multi-task spatio-temporal matching change detection network is proposed.To address the impact of inconsistent lighting and image registration differences in change detection data,we employ a Siamese network encoding-decoding structure,Res Net architecture,and a feature pyramid fusion module to incorporate additional spatio-temporal information.By introducing a spatio-temporal feature matching task,and leveraging prior knowledge of geographic space,we achieve effective multi-task learning.Experimental results demonstrate significant performance improvements on both publicly available and self-built datasets,and effective differentiation of different categories in complex scenes.(2)To address the insufficient utilization of spatiotemporal feature information and the omission of small targets in deep learning models,a multitask spatiotemporal attention change detection network is further proposed.Experimental results indicate that,compared to methods without spatiotemporal attention,the proposed method achieves the highest F1 and Io U scores on both the public dataset and the custom dataset.The method exhibits marked improvements in building outline detection accuracy,identification of backfill and road change extents,recognition of building integrity,reduction of false positives,and detection of small target features.(3)Algorithm integration,application,analysis,and refinement: Based on the Cesium frontend framework,Mongo DB is used as a tile storage database,combined with the Flask backend to construct a frontend-backend-separated online change detection prototype system.To address issues such as incomplete feature detection in applications,a post-processing method based on morphological operations is proposed,providing a viable approach for algorithm implementation.The thesis has 47 pictures,9 tables,and 81 references.