H038-0022
Use of machine learning algorithms and remote sensing data to identify shallow groundwater occurrences in the Western Desert of Egypt

Tuesday, 8 December 2020
Poster
Hossein Sahour1, Mohamed Sultan1, Mehdi Vazifedan2, Mustafa Kemal Emil1, Abotalib Z. A Farag3, Bassam Abdellatif4 and Mohammed El Bastawesy5, (1)Western Michigan University, Department of Geological and Environmental Sciences, Kalamazoo, MI, United States, (2)Western Michigan University, Department of Statistics, Kalamazoo, United States, (3)Western Michigan Univiversity, Department of Geological and Environmental Sciences, Kalamazoo, MI, United States, (4)National Authority for Remote Sensing and Space Sciences, Cairo, Egypt, (5)Kingston, ON, Canada
Abstract:
Groundwater is the primary source of water in arid and semi-arid regions of the world, yet identifying the potential areas for groundwater extraction requires costly and time-consuming field exercises. This research aims to utilize readily available remote sensing data and machine learning algorithms to delineate areas of shallow groundwater (SGW) occurrences across the Western Desert of Egypt (area: 680,000 km2). A Four-fold exercise was conducted: (1) SGW locations as the target variable was identified based on the presence of natural water discharge (springs), and watermelon fields. (2) Variables contributing to, or correlating with, water table variation (elevation, slope, distance to faults, distance to sapping features, rainfall, NDVI, brightness temperature, soil moisture, and radar backscattering coefficient) were identified and extracted from multi-sensor remote sensing data, and the dataset was randomly divided into two subsets of training (75% ), and validation (25%). (3) The relationships between SGW occurrence and its controlling factors were extracted using three machine learning techniques, namely extreme gradient boosting (EGB), support vector machine (SVM), and logistic regression (LR), and the results were tested on the validation subset using kappa coefficient (K) and percent accuracy (P). (4) The relationships were applied to the set of known variables to classify the SGW areas across the Western Desert of Egypt. Findings show (1) the three applied models yielded favorable and competitive results with a slight outperformance by the EGB (k= 0.96, P= 98%) over the SVM (k= 0.94, P= 95%)) and LR (K= 0.91, and P= 96%). (2) Analysis of variables in the modeling process show elevation is the most important variable in all three applied models. (3) Results can be used for identification of groundwater withdrawal sites in the study area and similar settings elsewhere
<< Previous Abstract | Next Abstract >>