Predicción de inundaciones en el municipio de Guasca Cundinamarca, mediante inteligencia artificial e imágenes satelitales

Abstract

This document develops a proposal for flood prediction in the municipality of Guasca, Cundinamarca, using multispectral satellite imagery and supervised learning computational models. The model involves downloading and preprocessing Sentinel-2 images, calculating spectral indices such as NDVI and NDWI, constructing a georeferenced training database, and implementing classification algorithms. The purpose of this entire process is to enable the system to learn spectral patterns and identify the presence of water in a land cover. Using a time series, the model can project this behavior into future scenarios. The methodology used is based on supervised learning, employing algorithms such as Support Vector Classifier, Random Forest, and Gradient Boosting. These models were trained with a previously labeled dataset, validated using performance metrics, and applied to satellite imagery to generate predicted water cover maps. The combination of data preprocessing, variable normalization, and hyperparameter tuning ensures the model's robustness and generalization capabilities. In conclusion, it is demonstrated that computational prediction models applied to Sentinel-2 imagery are an effective tool for anticipating the spatial variation of water bodies in the municipality of Guasca. Future studies recommend expanding the training base and exploring advanced techniques, such as the inclusion of precipitation data, to strengthen flood prediction as a key input in territorial planning and risk management at the local level.