Feature engineering strategies based on GIS and the SCS-CN method for improving hydrological forecasting in a complex mountain basin

Hydrological modeling and forecasting are important tools for adequate water resources management, especially in complex systems (basins) characterized by high spatio-temporal variability of runoff driving forces, landscape heterogeneity, and insufficient hydrometeorological monitoring. Yet, during the last decades, the use of machine learning (ML) techniques has become popular for runoff forecasting, and the current research trend focuses on performing feature engineering (FE) strategies aimed both at improving forecasting efficiencies and allowing model interpretation. Here, we employed three ML techniques, the Random Forest (RF) algorithm, traditional Artificial Neural Networks (ANN), and specialized Long-Short Term Memory (LSTM) networks, assisted by FE strategies for developing short-term runoff forecasting models for a 3300-km² complex basin representative of the tropical Andes of Ecuador. We exploited the information of two readily-available satellite products, the IMERG and GSMaP to overcome the absence of ground precipitation data, and the FE strategies proposed were based on GIS and the Soil Conservation Service Curve Number (SCS-CN) method to synthesize the use of land use and land cover, soil types, slope, among other hydrological concepts. To assess the forecasting improvement, we contrasted a set of efficiency metrics calculated both for the developed specialized models and for referential models without the application of  FE strategies. In terms of results, we were first able to develop accurate forecasting models by exploiting precipitation satellite data powered by ML techniques with different complexity levels. Second, the referential forecasting models reached efficiencies (Nash-Sutcliffe efficiency, NSE) varying from 0.9 (1-hour lead time) to 0.5 (11-hours), which were comparable for the RF, ANN, and LSTM techniques. Whereas for the specialized models, we found an improvement of 5–20 % in NSE-values for all lead times. The proposed methodology and the insights of this study provide hydrologists with new tools for developing short-term runoff forecasting systems in complex basins otherwise limited by data scarcity and model complexity issues.