Análisis multitemporal del espejo de agua en el sistema Chingaza mediante redes neuronales aplicadas a imágenes de radar

Abstract

According to the UN (UN, 2021), approximately 36% of the world's population, or 2.4 billion people, live in water-scarce regions, and 52% will experience severe water scarcity by 2050. This scenario reflects the difficult situation cities face in accessing and managing drinking water. In Colombia, the Chingaza system, which supplies about 70% of Bogotá's drinking water, has experienced an alarming lack of rainfall, reaching only 16.97% of its total capacity (Granadas, 2024). Despite the importance of this water system, there is no constant monitoring of water levels, highlighting the vulnerability of the city's water resource management. In this context, it is necessary to develop innovative solutions for observing and monitoring the evolution of water bodies. One promising alternative is the use of convolutional neural networks, which allow for the reliable classification of water surface area. Furthermore, synthetic aperture radar (SAR) images represent a valuable resource, as they eliminate the optical and climatological limitations present in conventional optical image-based methods, reducing the loss of temporal information. In this project, a multi-temporal analysis of the water surface area was performed using the U-Net convolutional neural network architecture applied to SAR images in the Chingaza system reservoirs during the rationing period implemented by the district mayor's office. This approach demonstrated the effectiveness of the U-Net architecture and its potential for monitoring various water bodies in the country