Identificación de zonas inundables en Bogotá mediante análisis de imágenes satelitales sentinel y técnicas de clasificación basadas en inteligencia artificial

Abstract

Floods constitute one of the most recurring natural hazards in Colombia, causing significant impacts on infrastructure, the environment, and the population's quality of life. According to the National Water Study (2018) by the Institute of Hydrology, Meteorology and Environmental Studies (IDEAM), approximately 190,935 km² of the national territory exhibit conditions prone to flooding, which corresponds to 17% of the country's continental area (IDIGER, 2021). Focusing on the target city evaluated in this study, the identification of flood-prone areas is a critical component of urban planning and risk management. Climate variability, rapid urban growth, and deficiencies in drainage systems have contributed to frequent flooding across various localities within the city. To address this issue, this project proposes the use of Sentinel-1 and Sentinel-2 satellite imagery: Sentinel-1 for terrestrial observation under any weather condition, and Sentinel-2 for validation using optical imagery. These satellites provide radar imagery for numerous applications; Synthetic Aperture Radar (SAR) images represent the most effective method for monitoring land subsidence and structural damage. The ability of these satellites to deliver information under all weather conditions, both day and night, makes them ideal for maritime and Arctic monitoring. According to the official Copernicus program portal, Sentinel-1 is specifically designed to provide imagery for rapid response to natural disasters such as floods or earthquakes. Approximately 75% of natural disasters are flood-related. SAR satellites possess an inherent capacity to capture images without cloud interference. Furthermore, Sentinel provides frequent monitoring, making these satellites ideal for flood assessment. This monitoring can be used to estimate the extent of flooded areas, thereby evaluating human, economic, and environmental impacts. High-resolution Digital Elevation Models (DEMs), through interferometric techniques (InSAR) applied to Sentinel-1 image pairs, can be used to conduct runoff and flood analyses in data-scarce regions (Copernicus, n.d.). Sentinel imagery is freely available via the Copernicus Open Access Hub and can be processed using software such as QGIS or Google Earth Engine. This project aims to implement machine learning algorithms to automate the detection of flooded areas, comparing their accuracy with traditional spectral index-based methods (NDWI). Through this approach, it is expected to demonstrate that artificial intelligence classification optimizes flood risk management compared to traditional methods that rely on slower, manual processing (EOS, n.d.). Following image processing in Google Earth Engine, a multitemporal analysis of heavy rainfall and drought events in the water bodies of Bogotá is presented. Subsequently, a U-Net artificial intelligence model is trained using 46,376 georeferenced patches of 256 x 256 pixels as input data. Finally, model performance is evaluated using standard metrics: Accuracy, Precision, Recall, F1-Score, and the Kappa Index (Jamali et al., 2024; Lan & Wang, 2025). This evaluation is conducted on a pixel-by-pixel basis over a massive validation set to verify the network's ability to accurately discriminate dry land from water bodies. Through the confusion matrix, correct classifications and errors (false positives and false negatives) are quantified (Jamali et al., 2024), validating the hypothesis that AI identifies flood-prone areas with greater precision than traditional methods. The final results facilitate the recognition of spatial recurrence patterns, which are essential for generating risk projections toward the year 2030 and supporting decision-making processes in the land-use planning of Bogotá (Saleh et al., 2024; Lan & Wang, 2025).