针对遥感图像山体滑坡分割提取难度大、精度低等问题，提出了 一种基于改进U-Net网络的遥感图像山体滑坡分割提取方法。首先将原始网络中的特征提取模块用残差网络ResNet进行替换，加深网络防止梯度消失，可以学习到更深层的特征;其次，融入(multi-scale features fusion module)多尺度特征融合注意力模块增强发现山体滑坡区域的能力;最后，采用广义的损失函数FTL(Focal Tversky Loss)替换带权重的交叉熵损失函数以平衡准确率和召回率之间的关系。实验结果表明，改进后算法mIoU为65。92％，比改进前提升了 2。5个百分点，mPA为73。93％，比改进前提升了 3。56个百分点，F1-score综合得分指标为60。08％，比改进前提升5。09个百分点。改进后模型算法能有效提高山体滑坡分割性能。
Landslide Segmentation Extraction from Remote Sensing Images Based on Improved U-Net Network
A remote sensing image landslide segmentation and extraction method based on an improved U-Net network is proposed to address the difficulties and low accuracy of remote sensing image landslide segmentation and extraction.Firstly,replacing the feature extraction module in the original network with the residual network ResNet and deepening the network to prevent gradient vanishing,we can learn deeper features.Secondly,incorporating the MSF(multi-scale features fusion module)can enhance the ability to discover landslide areas in mountains.Finally,a generalized loss function FTL(Focal Tversky Loss)is used to replace the weighted cross entropy loss function to balance the relationship between accuracy and recall.The experimental results show that the improved algorithm has an mIoU of 65.92％,which is 2.5 percentage points higher than before,an mPA of 73.93％,which is 3.56 percentage points higher than be-fore,and an Fl score index of 60.08％,which is 5.09 percentage points higher than before.The improved model algorithm can effectively improve the segmentation performance of mountain landslides.
remote sensing imagemountain landslide segmentationU-Net networkResNetloss function
NETL
NSTL
万方数据
