Deep Learning-based Fracture Extraction In Inclined Photographic Coal Mining Area

Surface cracks caused by coal mining are one of the main forms of damage to mining areas,which have a serious impact on safe production and ecological environment quality.Low-altitude unmanned aerial vehicle(UAV)oblique photography technology,as a quick and low-cost way to obtain high spatial resolution geographic spatial information,can be an effective approach for monitoring cracks in mining areas.However,without relying on manual editing and automatic image processing technology,there are significant recognition difficulties and differences in precision for extracting cracks in coal mining areas based on UAV oblique photography images,which limit the practical application of this technology in mining research.In this paper,we conducted research to address these issues and achieved the following main results:(1)Based on the location and terrain characteristics of the crack research area in the Zhongmei Tower Mountain Coal Mine in Datong City,Shanxi Province,we analyzed the steps of UAV image acquisition,discussed the method for determining the UAV flight parameters and flight range planning under complex terrain conditions,as well as the data pre-processing process and construction of a three-dimensional surface model.The generated orthophoto image,digital surface model,and three-dimensional model have clear terrain and geological features,and the modeling results meet the expected goals.(2)We created a crack labeling dataset for the mining area using methods such as manual annotation,object-oriented classification,and multi-scale segmentation.We used data augmentation methods to increase the dataset size and annotated a total of690 real mining area images of size 4000x3000.We optimized the storage method of the dataset,and after testing,found that HDF5 effectively solved the problem of memory errors during training.We analyzed deep learning model training methods such as transfer learning,hyperparameter setting,commonly used optimization algorithms,and loss functions,and explained the significance of evaluation metrics for deep learning models,such as confusion matrix,precision,recall,accuracy,F1 score,mean intersection over union(m Io U),and mean pixel accuracy.(3)We conducted training and evaluation of a crack detection model based on the U-Net.We evaluated different types of loss functions using the dataset and found that the weighted cross-entropy(WCE)loss function was the most optimal.We chose WCE as the loss function and Adam as the optimizer,and determined different hyperparameter optimization methods.With a larger batch size,the convergence speed of the model increased,and the optimal loss function value decreased,while the accuracy,precision,recall,and F1 score of the model all improved.We evaluated the model using evaluation parameters,the test set,and real data sets,and achieved the best evaluation results for all three methods,which could improve the model’s fitting and generalization abilities.(4)Based on the DeepLab V3 Plus network structure,we constructed the Mobile Net and Xception two backbone networks on the basis of transfer learning and evaluated the crack extraction models.The pre-training model of Xception,which was the backbone network,outperformed that of Mobile Net in accuracy,mPrecision,mIoU,mPA and other evaluation metrics.(5)We compared and analyzed the crack extraction effects of the U-Net and Deep Lab V3 Plus models,and found that the Deeplab V3 Plus-Xception model 5 had better segmentation results.We used the Deeplab V3 Plus-Xception model 5 as the prediction model to build a web-based crack extraction system to provide a convenient and practical crack extraction tool.