基于改进ResU-Net的中分辨遥感影像滑坡检测方法

Landslide Detection Method Using Improved ResU-Net of Medium Resolution Remote Sensing Images

针对基于中分辨率遥感影像滑坡检测精度低的问题,提出了一种基于注意力机制的改进ResU-Net模型,并且基于多光谱遥感影像数据集得出了有益于滑坡检测的多特征模型输入组合。本研究所用的原始数据集共14个特征,首先剔除无效特征,并加入归一化植被指数和归一化水体指数,生成新数据集。然后将新数据集应用于改进的ResU-Net与U-Net,ResU-Net,Attention U-Net,BiSeNet,Semantic FPN,U-Net++的对比实验,结果表明改进的ResU-Net在测试集上可获得76.91%的F1分数,同时精确率和召回率分别为77.34%和76.49%,在该任务中优于其他对比模型,且比ResU-Net模型的F1分数高了0.43百分点,有效提高了中分辨率遥感影像的滑坡检测精度。最后,再向数据集中依次加入归一化湿度指数和坡向特征,对比不同特征组合数据集产生的检测效果,结果发现加入坡向特征可最大化提升滑坡检测精度,F1分数可达77.03%。

Aiming at the problem of low accuracy of landslide detection based on medium resolution remote sensing images,we propose an improved ResU-Net model based on attention mechanism,and multi-feature model input combination that is beneficial to landslide detection is obtained based on multispectral remote sensing imagery dataset.The original dataset used has a total of 14 features.Firstly,the invalid features are removed,and the normalized difference vegetation index and the normalized difference water index are added to generate a new dataset.Secondly,The new dataset is applied in the comparative experiments of the improved ResU-Net with U-Net,ResU-Net,Attention U-Net,BiSeNet,Semantic FPN,U-Net++.It is showed that the improved ResU-Net can obtain the F1 score of 76.91%on the test set,while the precision and recall are 77.34%and 76.49%,respectively,which are better than that of other comparison models in this task,and it is 0.43 percentage points higher than the F1 score of the ResU-Net model,which effectively improves the landslide detection accuracy based on medium resolution remote sensing images.Finally,the normalized difference moisture index and aspect features are added to the dataset in turn,and the detection accuracy of different feature combinations is compared.The results show that adding aspect features can maximize the accuracy of landslide detection,and the F1 score reaches 77.03%.