基于深度学习与融合地形特征的黄土陷穴面向对象提取方法

黄土陷穴作为黄土高原地区一种特殊的地貌类型及地质灾害，其研究对指导黄土地区水土保持与工程建设工作具有重要意义。现阶段对陷穴的研究多基于传统野外调查，该方式成本高、效率低。为此，该研究开展面向对象与卷积神经网络（Convolutional Neural Networks，CNN）相结合的黄土陷穴自动化提取方法研究，并讨论融合地形特征对CNN模型提取精度的影响。研究选取黄土陷穴发育的典型区域，基于WorldView3遥感数据与ALOS高程数据，通过莫兰指数与灰度共生矩阵熵值确定影像的分割尺度，以面向对象的方式提取黄土陷穴的光谱、形状、纹理以及地形特征，制作融合地形特征与未融合地形特征的两类训练样本，进而训练两种CNN模型对同一区域内黄土陷穴进行提取，根据精确率、召回率以及F1分数评价模型的提取精度、分析对比两种CNN模型的提取结果，并建立支持向量机（Support Vector Machine，SVM）模型与CNN模型进行比较。研究结果表明，融合地形特征进行训练的CNN模型精确率达94.62%，召回率达86.27%、F1分数达90.26%，综合提取性能最好，相较于未融合地形特征训练的CNN模型，黄土陷穴的错分量大大减少，精确率提升18.10个百分点，F1分数提升9.15个百分点；两种CNN模型F1分数均达80%以上，比SVM模型分别高出6.94个百分点，16.09个百分点，提取结果均优于SVM模型；综上，融合地形特征的CNN模型可快速、精确地提取黄土陷穴，从而为黄土地区陷穴防治工作提供支持。

Object-oriented extraction method for loess sinkholes based on deep learning and integrated terrain features

Abstract: A loess sinkhole is a typical geological hazard that is widely distributed in the Loess Plateau of China. However, some latent damages are difficult to detect in the natural hazard. Particularly, the serious damage caused by loess sinkholes is ever increasing, as the infrastructure rapidly developed in recent years. The resulting water and soil loss can aggravate and even trigger the occurrence, development of mudflows, collapses, and landslides. In addition, the sinkhole collapse can also seriously damage the farmland infrastructure, such as terraces, water storage facilities, and irrigation systems, as well as the infrastructures such as highways, railways, bridges, and oil and gas pipelines. Thus, it is a high demand to detect the loess sinkholes for the soil and water conservation and engineering construction in the loess regions. However, the traditional field surveys are costly and inefficient on the loess sinkholes. In this study, an automatic and object-oriented extraction of loess sinkholes was proposed to determine the influence of integrated terrain features on the extraction accuracy of the Convolutional Neural Networks (CNN) model. The study area was selected as Lanzhou City, Gansu Province in western China. The segmentation scale of the image was then determined by the Moran''s I and Gray Level Co-occurrence Matrix entropy using WorldView 3 remote sensing image and ALOS digital elevation model data. An object-oriented extraction was implemented to obtain the spectrum (i.e., Mean_R, Mean_G, Mean_B, Std_R, Std_G, Std_B, Max_diff, and Brightness), shape (i.e., Area, Asymmetry, Border_index, Border_Length, Compactness, Density, Elliptic_Fit, Length/Width, Length, Shape_index, Roundness and Width), texture (i.e., GLCM_Homogeneity, GLCM_Contrast, GLCM_Mean, GLCM_Dissimilarity, GLCM_Entropy, and GLCM_Std), and terrain (i.e., Mean_DEM, Mean_Slope, Mean_Hillshade, Std_DEM, Std_Slope, and Std_Hillshade) features of the loess sinkholes. Two types of training samples were constructed with/without the terrain features, and then to train two CNN models for the extraction of the loess sinkholes in the same area. After that, the extraction accuracy of the model was evaluated, according to the precision rate, recall rate, and F1 score. A Support Vector Machine (SVM) model was also established to compare with the two CNN models. The research results show that the CNN model trained by integrating terrain features presented an accuracy rate of 94.62%, a recall rate of 86.27%, and an F1 score of 90.26%. Specifically, the false positive was significantly reduced, while the accuracy rate and F1 score increased by 18.10 percentage points, and 9.15 percentage points, respectively. The F1 scores of the two CNN models were both over 80%, which were 6.94 percentage points, and 16.09 percentage points higher than that of the SVM model, respectively. Consequently, the integrated terrain features to train the CNN model can effectively reduce the number of False Positive in the loess sinkholes, thus improving the comprehensive performance of the model. The extraction performances of the CNN models were all better than those of the SVM model. Therefore, the CNN model can be widely expected to integrate with the terrain features from the satellite data and object-oriented image analysis for the accurate and efficient extraction of the loess sinkholes.