北京测绘2025，Vol.39Issue(11)：1574-1579,6.DOI:10.19580/j.cnki.1007-3000.2025.11.003
基于孪生特征优化网络的耕地新增建筑提取
Extraction of new buildings on cultivated land based on Siamese feature optimization network
摘要
Abstract
In response to challenges in monitoring new buildings on cultivated land,such as missing sample datasets and low detection accuracy for small and medium-sized targets,this paper aims to construct a new building dataset suitable for multi-source,mixed-resolution remote sensing imagery and design a high-precision change detection model to enhance the automa-tion capabilities of monitoring new buildings on cultivated land.First,based on domestic satellite imagery with 0.5-2 m reso-lution,such as resources and high-resolution series,the paper constructed a mixed-resolution remote sensing image dataset for new building detection(MRNB-1.0),which contained 24,875 sample sets.Next,the paper improved the Siamese U network(U-Net)by introducing the V2 version of the residual network(ResNetV2)as the feature extraction backbone.Through shallow feature skip connections and multi-scale feature fusion optimization strategies,a residual network V2 Sia-mese shallow feature optimized U-Net(Res2-Siam-SFOUNet)was established.Finally,based on pre-trained weights from the ImageNet dataset,the MRNB-1.0 dataset was divided into training,validation,and test sets in a 6∶2∶2 ratio for model training.Experimental results demonstrate that the Res2-Siam-SFOUNet model achieves mean intersection-over-union(MIoU)of 0.591 on the test set,improving by 22.4%compared to the previous model.In application tests across four dis-tricts in Fujian Province,the average recall rate is 81.0%,and the average precision is 62.5%,representing increases of 11.5%and 8.1%,respectively,compared to the prior model.The proposed Res2-Siam-SFOUNet model effectively detects new buildings on cultivated land in mixed-resolution imagery,enhancing the automation level and efficiency of new building monitoring in cultivated land protection.关键词
耕地保护/高分辨率遥感/新增建筑物/卷积神经网络/孪生网络Key words
cultivated land protection/high-resolution remote sensing/new building/convolutional neural network/Siamese network分类
天文与地球科学引用本文复制引用
秦龙珍..基于孪生特征优化网络的耕地新增建筑提取[J].北京测绘,2025,39(11):1574-1579,6.基金项目
福建省测绘地理信息发展中心科技资助项目(202106) （202106）
