Objective In recent years,the use of deep learning to extract building information from images has be-come one of the research hotspots in the field of remote sensing.In order to accurately and efficiently extract the building information in the remote sensing imagery.Methods The backbone network of U-Net was replaced by ResNet,and a new building extraction network model was constructed by fusing Self-Corrected Convolution(SC-Cnov)and Efficient Channel Attention(ECA).The Massachusetts building dataset was used to design an abla-tion experiment,the extraction accuracy and practical effect of the constructed network model were compared and analyzed,and the universality of the network was verified by using the WHU dataset,and the migration experi-ment was designedby using the UAV image dataset of a mining area in Anhui Province to verify the migration a-bility of the constructed network.Results The constructed network reached 82.89％,92.26％and 88.32％on the three indexes of mIoU,mPrecision and mRecall respectively,which were 1.70％,1.08％and 1.19％higher than those before the improvement.In addition,in the migration experiment,the network reached 88.66％,94.37％and 93.19％on the three indicators of mIoU,mPrecision and mRecall,respectively.Conclusion The SCEC-Unet pro-posed in theresearch has good results in building extraction and performs well in the extraction of independent small buildings,special-shaped buildings and edge buildings,and the network has good migration ability,which can be used for transfer learning of building extraction tasks in mining areas.
self-correcting convolutional(SCConv)efficient channel attention(ECA)convolutional neural network(CNN)deep learningremote sensing information extraction
维普
万方数据
