FRACTURE-FAULT DETECTION USING DEEP LEARNING WITH STEPWISE ELIMINATION FROM SATELLITE IMAGES IN DJIBOUTI

Abstract

Accurate estimation of groundwater flow is crucial in arid regions where permanent surface water is absent. In several groundwater simulation models, an important parameter for identifying areas with high potential for groundwater resources is the accurate fracture-fault detection. In the present study we propose a deep learning approach to detect fracture-fault structures in the Ali Faren sub-catchment of Ambouli Wadi in Djibouti. Our deep convolutional neural network (Deep-CNN) model is trained on high-spatial resolution multispectral satellite images using wadi streamline as labels. Fracture-fault structures are extracted using stepwise elimination based on geological characteristics observed in relief images derived from PALSAR-1/2 data. Our results demonstrate that the proposed Deep-CNN model accurately detects fracture-fault lines, achieving a validation accuracy of 0.9684, precision of 0.9124, recall of 0.9701, and F1 of 0.8997. The proposed model has the potential to identify potential areas for groundwater resources across the country, contributing to sustainable water management and improving Djibouti's water security. Our study highlights the potential of deep learning techniques in addressing challenges related to sustainable water management in arid regions.