Detection of earthquake-damaged buildings via UAV high-resolution remote sensing images
Natural disasters occur frequently in Yunnan,China and cause enormous losses of life and property.An object detection technology based on the deep learning of remote sensing images can be used to rapidly locate damaged buildings caused by natural disasters and subsequently aid with disaster relief.However,several challenges affect the detection of damaged buildings,such as the lack of data on earthquake-damaged buildings and the weakness of the features of objects to be detected.Thus,a UAV remote-sensing image-based largescale high-resolution earthquake-damaged building database(UEDB)was constructed.A total of 4598 remote sensing images were collected in the disaster area of Yangbi Yi Autonomous County in Dali Bai Autonomous Prefecture,Yunnan Province,China.The dataset includes 76,012 building instances,with each instance labeled in three formats:an object location box label,an object segmentation label,and an object boundary label.Then,a novel Earthquake-Damaged Buildings Real-time Detection Model(EDBRDM)was constructed.This model includes three modules:object feature alignment(OFAM),feature difference calculation(FDCM),and object boundary constraint-based position box detection.The processing procedure of this model is as follows.Firstly,the OFAM correct the misalignment issues in images taken before and after a disaster,ensuring precise alignment of object features.This crucial step forms the foundation for subsequent feature analysis and difference calculation.Secondly,the FDCM is employed to compute the differences in features,highlighting the damage characteristics of buildings.By comparing the image features before and after the disaster,we can more clearly identify the damage of buildings,providing strong support for subsequent identification and analysis of damaged buildings.Lastly,the OBCPB introduces shallow boundary features into deep features,providing boundary constraints for the prediction of damaged building locations and categories.This step helps enhance detection accuracy,ensuring that we can accurately identify and locate damaged buildings.Through the collaborative effort of these three steps,we can achieve precise detection of damaged buildings.To validate the crucial role of the proposed modules,we delve into the internal operating principles of the model through the lens of feature visualization.Firstly,by comparing the feature changes after OFAM processing,we can clearly observe the significant improvements in the alignment of features across pre-and post-disaster images,demonstrating the effectiveness of OFAM in correcting image offsets.Secondly,by observing the enhancement of damaged building features by FDCM,we find that it effectively highlights the damaged areas of buildings,providing strong support for subsequent identification and analysis of damaged buildings.Finally,through the observation of the boundary constraint effect of OBCPB,we can see how it helps to improve the localization accuracy of the model,ensuring that damaged building objects can be accurately identified.It is noteworthy that our proposed model,EDBRDM,has achieved a remarkable accuracy of 86％on the UEDB test dataset,fully demonstrating its excellent performance.Furthermore,the application of EDBRDM to actual scenes in different locations has also yielded satisfactory results,further validating its effectiveness and reliability in practical applications.
deep learninghigh-resolution remote sensing imagesobject detectionchange detectionearthquake disaster
NETL
NSTL
万方数据
