Use this link to cite:
https://hdl.handle.net/2183/49645 Desarrollo de un sistema de Business Intelligence y analítica predictiva para los incendios forestales en Galicia a nivel municipal
Loading...
Identifiers
Publication date
Authors
Castro Doce, Paula Hong
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] Este trabajo aborda el diseño, implementación y evaluación de un sistema de inteligencia de negocio y analítica predictiva para la ocurrencia de incendios forestales a escala municipio día en Galicia. A partir del conjunto de datos IberFire, integrado con los registros históricos de la Estadística General de Incendios Forestales (EGIF), se construyó un almacén de datos en Oracle siguiendo la metodología de Kimball, sobre el que se realizó un análisis exploratorio en Power BI para caracterizar el fenómeno y orientar las decisiones de modelado. Mediante técnicas de importancia de variables (SHAP) se seleccionó un subconjunto reducido de diez predictores, sobre el que se entrenó y ajustó un modelo de clasificación basado en XGBoost con validación cruzada de tipo walk-forward, tomando como referencia comparativa el índice IRDI (Índice de Riesgo Diario de Incendio Forestal) empleado actualmente por la Xunta de Galicia. La evaluación sobre un conjunto de test prospectivo (2021-2024) muestra que el modelo final supera notablemente a un clasificador aleatorio y que la configuración de diez variables mantiene un rendimiento muy próximo al del modelo enriquecido, con una caída de rendimiento respecto a validación concentrada principalmente en el año 2022, coherente con el cambio de régimen de siniestralidad observado en dicho periodo. El estudio adicional de distintos horizontes temporales de predicción confirma que la ventana de siete días ofrece el mejor equilibrio entre utilidad operativa y capacidad discriminativa. El sistema desarrollado se plantea como una herramienta de apoyo complementaria a mecanismos de referencia como el IRDI.
[Abstract] This work addresses the design, implementation and evaluation of a business intelligence and predictive analytics system for wildfire occurrence at the municipality-day level in Galicia. Building on the IberFire dataset, integrated with historical records from the Estadística General de Incendios Forestales (EGIF), a data warehouse was built in Oracle following the Kimball methodology, supporting an exploratory analysis in Power BI aimed at characterizing the phenomenon and guiding subsequent modeling decisions. Using SHAP-based variable im portance techniques, a reduced subset of ten predictors was selected and used to train and tune an XGBoost classification model with walk-forward cross-validation, benchmarked against the Daily Forest Fire Risk Index (DFFRI), currently used by the Xunta de Galicia. Evaluation on a prospective test set (2021-2024) shows that the final model substantially outperforms a random classifier, and that the ten-variable configuration remains close in performance to the enriched model, with the performance drop relative to validation concentrated mainly in 2022, consistent with the shift in fire-regime observed that year. A further study of different prediction horizons confirms that the seven-day window offers the best balance between operational usefulness and discriminative power. The resulting system is proposed as a support tool that could complement reference mechanisms such as the DFFRI.
[Abstract] This work addresses the design, implementation and evaluation of a business intelligence and predictive analytics system for wildfire occurrence at the municipality-day level in Galicia. Building on the IberFire dataset, integrated with historical records from the Estadística General de Incendios Forestales (EGIF), a data warehouse was built in Oracle following the Kimball methodology, supporting an exploratory analysis in Power BI aimed at characterizing the phenomenon and guiding subsequent modeling decisions. Using SHAP-based variable im portance techniques, a reduced subset of ten predictors was selected and used to train and tune an XGBoost classification model with walk-forward cross-validation, benchmarked against the Daily Forest Fire Risk Index (DFFRI), currently used by the Xunta de Galicia. Evaluation on a prospective test set (2021-2024) shows that the final model substantially outperforms a random classifier, and that the ten-variable configuration remains close in performance to the enriched model, with the performance drop relative to validation concentrated mainly in 2022, consistent with the shift in fire-regime observed that year. A further study of different prediction horizons confirms that the seven-day window offers the best balance between operational usefulness and discriminative power. The resulting system is proposed as a support tool that could complement reference mechanisms such as the DFFRI.
Description
Keywords
Incendios forestales Predicción de ocurrencia Aprendizaje automático XGBoost Almacén de datos Selección de variables (SHAP) Índice de Riesgo Diario de Incendios Forestales (IRDI) Wildfire prediction Occurrence forecasting Machine learning Data warehouse Feature selection (SHAP) Daily Forest Fire Risk Index (DFFRI)
Editor version
Rights
Todos os dereitos reservados. Prohíbese a reprodución, transformación, distribución e comunicación pública da obra por terceiros. Permítese a súa visualización e a descarga dunha copia privada para o uso persoal.







