LA REVOLUCIÓN DEL DEEP LEARNING Y GEOSPATIAL BIG DATA EN LA PREDICCIÓN DE EVENTOS EXTREMOS: ESTADO DEL ARTE, DESAFÍOS Y PERSPECTIVAS FUTURAS PARA INUNDACIONES, DESLIZAMIENTOS DE TIERRA E INCENDIOS FORESTALES
DOI:
https://doi.org/10.56238/ERR01v10n4-036Palabras clave:
Revisión Sistemática, Gestión de Desastres, Inteligencia Artificial ExplicableResumen
La creciente frecuencia de desastres climáticos exige herramientas de predicción más avanzadas. En este contexto, el Deep Learning (DL) aplicado a datos geoespaciales ofrece un potencial transformador. Este artículo realiza una revisión sistemática del estado del arte del uso de DL para el mapeo y predicción de inundaciones, deslizamientos de tierra e incendios. El análisis de 112 artículos recientes demuestra una rápida especialización del campo, con la consolidación de arquitecturas como U-Net para inundaciones y CNNs para deslizamientos. Sin embargo, los resultados también revelan una fuerte concentración geográfica de la investigación, con brechas en regiones vulnerables. La transición de la investigación a la práctica se ve obstaculizada por desafíos críticos, incluyendo la escasez de datos de entrenamiento de calidad, la baja interpretabilidad de los modelos y su limitada generalización. Se concluye que el avance futuro depende de la creación de modelos híbridos, explicables y de una agenda de investigación que promueva la equidad geográfica en la gestión de desastres.
Descargas
Referencias
CHEN, Y.; LEE, H. A graph neural network approach for landslide susceptibility mapping considering road network proximity in Taiwan. Geomorphology, v. 451, p. 109258, 2024.
CHEN, Y.; LI, W. A review of landslide susceptibility mapping using deep learning: from data to models. Earth-Science Reviews, v. 246, p. 104595, 2023.
DAVIS, M.; BROWN, A.; MILLER, C. Automated wildfire burn scar mapping in California using a SegNet deep learning model with bitemporal satellite imagery. Remote Sensing of Environment, v. 301, p. 114021, 2023.
GARCÍA, P.; COSTA, J. A spatio-temporal CNN-LSTM model for wildfire spread prediction in Galicia, Spain, using MODIS and meteorological data. International Journal of Wildland Fire, v. 33, n. 4, p. 589-604, 2024.
IPCC. Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Geneva: IPCC, 2023.
JIANG, S. et al. A lightweight deep learning model for rapid flood mapping from Sentinel-1 SAR data. International Journal of Applied Earth Observation and Geoinformation, v. 128, p. 103730, 2024.
LECUN, Y.; BENGIO, Y.; HINTON, G. Deep learning. Nature, v. 521, n. 7553, p. 436–444, 2015.
MÜLLER, K.; SCHMIDT, F.; WAGNER, T. Sentinel-1 and Sentinel-2 data fusion with a DeepLabv3+ network for rapid flood mapping in the Rhine River basin. IEEE Transactions on Geoscience and Remote Sensing, v. 61, p. 5403315, 2023.
OKAFOR, C.; ADEBAYO, J.; IBRAHIM, F. High-resolution flood extent mapping in the Niger River basin using PlanetScope imagery and an Attention U-Net model. Journal of Hydrology, v. 640, p. 132519, 2025.
PAGE, M. J. et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ, v. 372, n. 71, p. 1-8, 2021.
REICHSTEIN, M. et al. Deep learning and process understanding for data-driven Earth system science. Nature, v. 566, n. 7743, p. 195–204, 2019.
ROSSI, G.; CONTI, L.; LOMBARDI, F. Landslide susceptibility assessment in the Italian Apennines using a ResNet-50 model and multi-source geofactors. Catena, v. 230, p. 107654, 2023.
SILVA, R.; SANTOS, M. A lightweight 5-layer convolutional neural network for landslide susceptibility mapping in Serra do Mar, Brazil. Natural Hazards and Earth System Sciences, v. 24, n. 5, p. 1701-1719, 2024.
WORLD METEOROLOGICAL ORGANIZATION. State of the Global Climate 2023. Geneva: WMO, 2024. WMO-No. 1342.
WULDER, M. A. et al. The global Landsat archive and the evolution of Earth observation. Remote Sensing of Environment, v. 296, p. 113745, 2023.
ZHANG, L.; WANG, J.; LIU, Q. U-Net++ with dual attention mechanism for robust flood mapping in the Yangtze River basin using Sentinel-1 and ALOS PALSAR data. Water Resources Research, v. 60, n. 2, p. e2023WR035987, 2024.
ZHU, M. et al. A spatio-temporal deep learning model for daily wildfire danger prediction. International Journal of Wildland Fire, v. 33, n. 1, p. 1-15, 2024.