Automatic Detection of Deformation Cracks in Slopes Fused with Point Cloud and Digital Image

  Objectives  As one of the precursor features of slope deformation, surface cracks can provide forecast information for the early identification of geological hazards and the determination of motion instability characteristics. Affected by terrain conditions, the efficiency of manual investigation is low, and it is difficult to solve the problem of size effect and noise filtering of cracks in complex backgrounds using a single remote sensing data identification. In order to efficiently collect the location and geometry information of surface fractures on deformed slopes, the main purpose is to develop a method system for automatic crack identification and information statistics based on remote sensing data, which is suitable for large-scale slope surfaces.

  Methods  Using the high-resolution three-dimensional point cloud and orthophoto image generated by autonomous terrain-following flight technology as the data source, the automatic extraction of cracks and the statistical research on the deformed slope of the reservoir bank of Baihetan Hydropower Station are carried out. The initial automatic identification of slope cracks is completed by using six algorithm models with point cloud's roughness, slope, dispersion and digital image's pixel gradient, gray value and RGB (red green blue)value as identification features respectively. Completed receiver operating characteristic curve (ROC) tests of different models and determined segmentation thresholds. The initial extraction results are processed by morphological repair and three filtering algorithms of crack direction, length and frequency based on density-based spatial clustering of applications with noise (DBSCAN) algorithm index to deal with background noise. The two-category evaluation index is used to analyze the pros and cons of the 6 filtered crack extraction results, and for the effect of crack size, the optimal detection model after data fusion is obtained. Based on the crack skeleton and outline, 6 quantitative characteristic indicators of quantity, length, width, direction, dispersion, and crack density are automatically calculated.

  Results  Starting from the fracture image characteristics of different scales, 6 different fracture identification models are used and ROC test is carried out, and the area under the curve values are all between 0.6 and 0.85. The highest F₁ value of the model after fusion is 0.835 0, which can better meet the actual engineering needs. The morphological repair and filtering algorithm proposed can effectively reduce the background noise of cracks, improve the overall accuracy of the model, and can remove up to 82.7% of the noise with minimal crack distortion. The density clustering algorithm of DBSCAN is used to complete the fracture pixel classification and quantity statistics, and according to the ratio of pixel size and actual distance, the algorithm automatically completes the information acquisition of fractures, which can provide technical support for quantitative description and evaluation of fracture characteristics. Various types of crack identification algorithms used in this paper have different advantages and disadvantages and applicable situations, and the optimal combination relationship can be selected according to the field conditions with different characteristics.

  Conclusions  In this study, a method for automatic crack extraction and information statistics is constructed in a large-scale, multi-scale complex deformation slope scene, which can meet the needs of actual production operations. At the same time, it has important practical significance to further promote the early identification of geological disasters and intelligent monitoring and early warning.