基于深度学习的黄丘区正负地形分割  

Positive and Negative Terrain Segmentation in the Loess Hilly and Gully Region Based on Deep Learning

作　　者：赵子林 韩磊[2,3] 陈芮 赵永华 李亚北[1] 康宏亮 ZHAO Zilin;HAN Lei;CHEN Rui;ZHAO Yonghua;LI Yabei;KANG Hongliang(School of Earth Sciences and Resources,Chang'an University,Xi'an 710054,China;School of Land Engineering,Shaanxi Key Laboratory of Land consolidation,Chang'an University,Xi'an 710054,China;State Key Laboratory of Loess and Quaternary Geology,Institute of Earth Environment,Chinese Academy of Sciences,Xi'an 710061,China)

机构地区：[1]长安大学地球科学与资源学院,西安710054 [2]长安大学土地工程学院、陕西省土地整治重点实验室,西安710054 [3]中国科学院地球环境研究所黄土与第四纪地质国家重点实验室,西安710061

出　　处：《水土保持研究》2023年第5期21-30,共10页Research of Soil and Water Conservation

基　　金：国家自然科学基金项目(41871190);陕西省重点研发计划(2021SF-440);黄土与第四纪地质国家重点实验室开放基金(SKLLQG2002);长安大学中央高校基本科研业务费专项资金(300102353201)。

摘　　要：[目的]基于深度学习,实现大范围、高精度的正负地形分割,正负地形的有效分割对黄土高原水土流失治理及生态恢复重建具有重要的理论价值和指导意义。[方法]在黄土高原丘陵区选取典型样区,采用中分辨率DEM数据制作地形分割数据集,构建了改进Unet的黄丘区正负地形分割模型,该模型以Unet模型结构为基础,引入残差模块替换卷积模块加深网络结构,增加了地形信息的提取;结合卷积注意力模块,排除无用信息增加了模型抗干扰性;优化激活函数与损失函数,增强了模型鲁棒性与精度。[结果]坡面畸变邻域判断法地形分割总体精度为70.3%,在深度学习模型中,改进型Unet深度学习模型效果最优,相较于Unet模型与Res-Unet模型都有一定的提升,总体精度达到了86.2%。[结论]与传统的坡面畸变邻域判断法比较,基于深度学习的网络模型分割结果精度评价指标均较优,并验证了改进Unet的黄丘区正负地形分割模型的有效性。[Objective]Based on deep learning,a large-scale and high-precision positive and negative terrains segmentation is realized,and the effective segmentation of positive and negative terrains has important theoretical value and guiding significance for soil erosion control and ecological restoration and reconstruction in the Loess Plateau.[Methods]The typical sample area was selected in the hilly area of the Loess Plateau,and the terrain segmentation data set was made by using the medium resolution DEM data.The positive and negative terrains segmentation model of the improved Unet was constructed.Based on the Unet model structure,the residual module was introduced to replace the convolution module to deepen the network structure and increase the extraction of terrain information.Combined with the convolution attention module,eliminating useless information increased the anti-interference of the model;the activation function and loss function were optimized to enhance the robustness and accuracy of the model.[Results]The overall accuracy of the terrain segmentation of the slope deformation neighborhood judgment method is 70.3%.Among the deep learning models,the improved Unet model has the best effect,with a certain improvement compared with both the Unet model and the Res-Unet model,and the overall accuracy reaches 86.2%.[Conclusion]Compared with the traditional slope distortion neighborhood judgment method,the accuracy evaluation index of the network model segmentation results based on deep learning is better,which verifies the effectiveness of the improved Unet model in terrain segmentation.

关 键 词：正负地形分割 深度学习 Unet 残差模块 卷积注意力 

分 类 号：P208[天文地球—地图制图学与地理信息工程]