Detection of Geological Features using Aerial Image Analysis and Machine Learning

Geologic structures are one of the crucial parameters in blast design. Structural geology and rock properties influence drilling patterns, blast layout, and initiation systems. A comprehensive understanding of structural characteristics is essential to achieve the desired blasting objectives. This paper aims to identify geologic features in a bench using images obtained from the drone. Images obtained from mine site are annotated for geological features such as faults and joints. Semantic segmentation of images is achieved using an encoder and decoder structure available in the machine-learning pipeline. The encoder consists of initial convolution and pooling layers to capture essential features of the geology and rock patterns from two-dimensional RGB images. Batch normalization is used to put all the features on the same scale. These images are then inputted as training data for the machine-learning algorithm. Machine learning model is trained on sub-section of interest to achieve reasonable accuracy. Initially, 1000 best annotations were used to train the model and additional samples were added later to achieve more accuracy and avoid iteration of machine learning model in wrong convergence. After the model achieves reasonable accuracy, it is used to recognize geological features on new data. Geological features identified on the new data can be used by the blasting crew in blast design and keeps the geological maps updated that can help the geotechnical team.