Automated Underground Water Leakage Detection with Machine Learning Analysis of Satellite Imagery

Increasing water shortages, droughts, and global warming demand methods to rapidly detect underground water leaks. Conventional techniques are costly, time-consuming, and error-prone. However, remote sensing techniques can offer innovative solutions. Previous studies mostly used optical sensors. However, optical data has limitations, including noise interference, limited null subsurface penetration, and weather dependency. Therefore, the study in this paper aims at exploring the combination of radar satellite data and machine learning to automatically identify underground water leakages. Radar data offers sensitivity to soil moisture below the surface. Moreover, image texture features were leveraged from dual-polarized radar data to enhance prediction. Gray-level co-occurrence matrix texture features were combined with backscattering coefficients to create a feature space that could better train the random forest. Results indicate the ability to automatically detect 69% of underground leaks with subsurface moisture alone, which lists, tables, figures, display equations, footnotes, or references.