Tesis (Magíster en Ingeniería Civil con énfasis en Recursos Hidráulicos y Medio Ambiente)

Drought is a natural, erratic phenomenon that has a widespread and significant impact on
socioeconomic and environmental development. The early monitoring and evaluation of
drought through forecasting models would allow the articulation of early control and mitigation
strategies, thus achieving an optimal development in planning and preparation for climate
change. Therefore, this research developed a methodology for spatiotemporal analysis of
drought patterns using automatic learning tools in the dry corridor of Central America. To this
end, some specific milestones were defined. These include: (i) To assess temporal and spatial
meteorological and agricultural droughts events, (ii) Identifying and validate results of the
spatiotemporal events using computer vision techniques and finally (iii) Implementing machine
learning drought forecasting models.
ERA 5 monthly land average dataset was used as input for index estimation, spatiotemporal
analysis and forecasting models. The frequency of drought events was calculated using
standardized SPI and SPEI indices for accumulation periods of 1,3,6,9. However, 3,6 allowed
a more realistic analysis of the seasonal change conditions in the hydrological regime of the
area and the identification of the existing teleconnection between drought events and scale
propagation. Regarding the spatiotemporal dynamics, 97 drought events of greater extension
were identified, which are generally originated in countries such as Guatemala, Nicaragua, and
El Salvador between seasonal periods not longer than 7 months. Additionally, the suitability of
automatic learning models such as SVR, ANN and deep learning such as LSTM for index
forecasting (r2=0.80) and drought dynamics in a temporal window of 1 to 6 months ahead was
verified with considerable performance.
The presented methodology provides an important basis for drought characterization and
forecasting through the integration of spatiotemporal tracking models and machine learning
techniques. Therefore, the methodological development can be adapted as an instrument for
monitoring and forecasting, articulated to management and early mitigation policies. Finally,
we suggest adapting variables related to the orographic context, relief, land use and land cover
change, for instance, to improve the forecasting performance of the exposed forecasting models.