Research And Application Of Regional Remote Sensing Object Supervision Algorithm

In today’s complex international landscape,marked by tense relations and frequent geopolitical conflicts,it is critical to establish a regional remote sensing object supervision system to effectively monitor regional strategic situations.With the advancement of satellite technology and aerospace development,remote sensing images can be leveraged to supervise regional vital objects.This thesis proposes remote sensing object detection and remote sensing change detection methods,which aim to achieve regional object supervision.Remote sensing object detection aims at locating movable objects such as vehicles and ships,and fixed facilities such as bridges,through remote sensing images.Remote sensing change detection extracts change information of buildings by comparing remote sensing images taken at two or more different time points in the same region.However,the unique characteristics of remote sensing images pose significant challenges,including their bird’s-eye view perspective,complex environmental background,small and densely distributed objects,and large time intervals.To overcome these challenges,this thesis proposes corresponding solutions.Firstly,for remote sensing object detection,a method based on center-ness and repulsion constraints is introduced to locate oriented remote sensing objects.This method locates oriented remote sensing objects with adaptive point sets that can capture key points with rich semantics and geometric information,thereby improving the accuracy of remote sensing object localization.A sample quality assessment module is introduced to assess and assign samples under constraints such as center-ness,which improves the quality of predicted objects.The oriented repulsion regression function is proposed to deeply analyze the spatial interactions among objects and enhance the robustness of small and densely distributed object detection.Secondly,for remote sensing change detection,this thesis proposes a change detection method based on a Siamese network with feature interaction.The method adopts a weight-sharing Siamese network architecture to extract bi-temporal features which can reduce the network parameters and avoid the difference in feature distribution caused by different extractors.Utilizing the multi-head self-attention mechanism of the Swin Transformer,the model can adaptively focus on changed and non-changed areas.Two feature interaction strategies i.e.spatial interaction and channel interaction,are proposed to realize communication and information interaction between bi-temporal features and further enhance the network’s perception of changed areas.Finally,this thesis verifies the effectiveness of the remote sensing object detection method in the representation of oriented objects,sample assessment and assignment,and small and densely distributed object detection on four remote sensing object detection datasets.The effectiveness of the change detection method in change feature extraction,change information perception,and changed area discrimination is verified on three remote sensing change detection datasets.