Water leak detection through satellite imagery and deep learning

The society is aware of the critical lack of water problem that is facing currently. However, approximately 40% of the extracted water is lost due to water leaks, which is still a more critical problem. Most methods employed to detect such leaks are expensive and ineffective because they rely on fieldwork, which is very time-consuming and barely covers the entire pipeline network. An interesting optional data source to deal with this problem is Satellite imagery, they are an alternative solution to water leak detection due to their capacity to capture more electromagnetic spectrum data. This data enables the measurement of specific soil conditions, including soil moisture level and temperature, for instance. In this work, a Convolutional Neural Network trained with satellite imagery is proposed to deal with water leak detection. The satellite imagery employed was acquired according to ground truth data of real water leaks located in urban areas in Aguascalientes, Mexico. Hence, the coordinates corresponding to the identified and subsequently repaired leaks during the 2020-2022 time frame were utilized to download images from the Sentinel-2 (S2) satellite. Furthermore, the Land Surface Temperature (LST) of each image was estimated using Landsat 8 data and added as a spectral band to all S2 images; that means, adding one layer to the S2 images. Several experiments were done with combinations of data sources, and the best-obtained result on the test dataset was 81% accuracy for the S2 + LST model. The obtained results underscore the feasibility of effectuating water leak detection through the synergistic application of satellite imagery and deep learning methodologies.