Blasthole Location Detection Using Support Vector Machine and Convolutional Neural Networks on UAV Images and Photogrammetry Models

Abstract

Identifying the as-drilled location of blastholes is crucial for achieving optimal blasting results. This research proposes a novel integrated methodology to control drilling accuracy in open-pit mines. This approach is developed by combining aerial drone images with machine learning techniques. The study investigates the viability of photogrammetry combined with machine learning techniques, particularly Support Vector Machine (SVM) and Convolutional Neural Networks (CNN), for automatically detecting blastholes in photogrammetry representations of blast patterns. To verify the hypothesis that machine learning can detect blastholes in images as effectively as humans, various datasets (drone images) were obtained from different mine sites in Nevada, USA. The images were processed to create photogrammetry mapping of the drill patterns. In this process, thousands of patches were extracted and augmented from the photogrammetry representations. Those patches were then used to train and test different CNN architectures optimized to locate blastholes. After reaching an acceptable level of accuracy during the training process, the model was tested using a piece of completely unknown data (testing dataset). The high recall, precision, and percentage of detected blastholes prove that the combination of SVM, CNN, and photogrammetry (PHG) is an effective methodology for detecting blastholes on photogrammetry maps.