Abstract:
The Geological Strength Index is a vital tool for determining the strength of rock mass including two main factors, structure of the rock mass & weathering conditions. This allows to classify the rock mass between 1-100 from the table developed by Hoek. Earlier the methods used were inefficient & incompatible since they were focused entirely on visual observation & practical expertise of the individual which was prone to many discontinuities in the results. Keeping in mind this gap, our prime objective was to develop an image processing framework for rock mass images with help of CNN, to train machine learning to predict GSI values & to validate the accuracy & reliability of the predictive models. Finally, a user-friendly tool for GSI prediction was created based on the image inputs. Following an efficient methodology in which we acquired our rock mass data from Abbottabad Motorway with help of drone DJI Mavic 3, then images taken were pre-processed to grey-scale images to control the shading & lightning conditions of rock mass. These pre-processed images were divided into two datasets, training & testing datasets. Afterwards, a CNN based framework was established in which the layers were included to teach the model only about the structure & weathering conditions which in turn gave us the GSI value from the input images. The training set was now used to train the model & remove the discrepancies. Cross validation was done to improve the accuracy & reliability of model. This tool is ready to be used on site for predicting GSI of rock mass. Nexus to above, some of the future directions regarding this research to improve the practicality & efficiency of this model will be to further refining the CNN model architecture to improve prediction accuracy. Exploring additional features or data sources that could enhance the model's performance & investigating the transferability of the model to different geological settings or rock types. Finally, carrying out detailed validation experiments to assess the real-world applicability of the model.
