Application of a Machine Learning Technique for Developing Short-Term Flood and Drought Forecasting Models in Tropical Mountainous Catchments

Abstract

AbstractFloods and droughts are among the most common natural hazards worldwide. They produce major impacts on society, economy, and ecosystems. Even worst, the frequency and severity of hydrological extremes are expected to increase with climate change and land-use alteration. As a countermeasure, during last decades, implementation of flood and drought forecasting models have globally become an emerging field of research for water management and risk assessment. In mountainous areas, hydrological extremes forecasting is unfortunately more challenging considering that information other than precipitation and runoff is not commonly available due to budget constraints, remoteness of the study areas and extreme spatio-temporal variability of additional driving forces. This is especially true for the tropical Andes in South America, which is the longest and widest cool region in the tropics. Recent advances in computational science coupled with long-term data availability have boosted Machine Learning (ML) applications. Among the variety of ML techniques, there is a potential to use the Random Forest (RF) algorithm due to its simplicity, robustness and capacity to deal with complex data structures. We used a step-wise methodology to developed short-term flood and drought forecasting models for several lead times (4, 8, 12 and 24 h) for two catchment representative of the Ecuadorian Andes. We found that derived models can reach maximum validation performances (Nash–Sutcliffe efficiency, NSE) from 0.860 (4-h) to 0.545 (24-h) for optimal inputs composed only by features accounting for 80% of the model’s outcome variance. Moreover, we found that a set of RF hyper-parameters can be transferred to a comparable catchment with a maximum model performance reduction of 0.10 (NSE). Overall, the forecasting of hydrological extremes (especially floods) is challenging mainly due to lack of relevant data (driving forces) and sufficient extreme events from which RF models learn. The applicability of this study is to assist authorities in flood and drought management to evaluate hazard risks and to found the basis for developing integrated action plans from a local and regional perspective.KeywordsNatural hazardsEarly warning systemFloodDroughtTimeseries forecastingHydroinformaticsMachine learningRandom forest