基于深度学习与多源遥感数据的新增建设用地自动检测

Automatic detection of newly increased construction land based on deep learning and multi-source remote sensing data

新增建设用地自动检测可以为自然资源保护提供一种新型有效的技术支持,本文以广西贵港市为研究区域,提出一种基于深度学习与多源遥感数据的新增建设用地自动检测方法。首先通过在训练区中对高分辨率遥感影像进行影像预处理、数据增广与差值处理得到训练数据,然后利用深度学习语义分割模型(DeepLabv3+)进行训练、调优,接着在测试区中结合遥感影像(Sentinel-2A)的变化区域提取结果对可能出现新增建设用地的区域进行筛选,最后对不同裁剪重叠率下的新增建设用地的自动检测结果进行验证。结果表明:1)在测试区中裁剪重叠率越高,图斑检测正确率越高,但同时也增加了检测计算量与图斑错分率,裁剪重叠率为70%时能在检测正确率、计算量和错分率之间取得较好的平衡。2)在70%的裁剪重叠率下,新增建设用地图斑检测正确率85.16%,错分率36.57%,图斑平均IoU为57.23%,检测面积率74.52%。

The automatic detection of new construction land can provide a new and effective technical support for the protection of natural resources.Taking Guigang city of Guangxi Province as the re‐search area,this paper proposes an automatic detection method for new construction land based on deep learning and multi-source remote sensing data.Firstly,in training area the high-resolution remote sensing image is performed preprocessing,data augmentation and difference processing to obtain train‐ing data.Secondly,deep learning semantic segmentation model(DeepLabv3+)is used for image train‐ing,tuning;and then according to the extraction results of the area where remote sensing image(Senti‐nel2A)changed,the regions where newly increased construction land might appear are screened in the test area.Finally,the automatic detection results of newly increased construction land under differ‐ent cropping overlap rates are verified.The results show that:1)In the test area,the higher the crop‐ping overlap rate is,the higher the patches detection accuracy will be;whereas at the same time,the calculation amount of detection and the patches error rate increase.When the cropping overlap rate is 70%,a good balance can be achieved between the detection accuracy,calculation amount and the patches error rate.2)A cropping overlap rate of 70%delivers 85.16%detection accuracy of the newly increased construction land patches,36.57%misclassification rate,57.23%IoU of the average patches and 74.52%detection area ratio.