A comparison of performance of SWAT and machine learning models for predicting sediment load in a forested Basin, Northern Spain

Abstract

In water bodies, sediment transport is a potential source of numerous negative effects on water resource projects and can damage environmental services. Two machine learning (ML) algorithms, the M5P and random forest (RF) models, have been explored for the first time as alternatives to the Soil and Water Assessment Tool (SWAT) model to estimate suspended sediment load (SSL) in the Oskotz river basin, a forested experimental basin in Navarra, northern Spain. In the ML models, streamflow and precipitation data were used to estimate daily SSL, testing different combinations of these inputs. The ML models were more accurate than the physically based hydrological SWAT model for all input scenarios tested at the daily scale. Moreover, although the SWAT results improved considerably at the monthly scale, the statistics obtained were generally inferior compared to the ML models. For the best combination of inputs, M5P demonstrated a superior ability to estimate SSL (R2 = 0.73, MAE = 135.04, RSR = 0.54, NSE = 0.71 and PBIAS = 5.19), compared to RF (R2 = 0.72, MAE = 143.39, RSR = 0.57, NSE = 0.67 and PBIAS = 11.60) and SWAT (R2 = 0.57, MAE = 181.24, RSR = 0.65, NSE = 0.57 and PBIAS = -1.27). The average sediment loads in winter, the season with the highest sediment generation in the Oskotz basin, were 2,094.04, 1,831.08 and 2,242.67 tonnes for M5P, RF and SWAT, respectively, compared to an observed SSL of 1,878.16 tonnes. These results indicate that M5P and RF are suitable models for simulating fluvial sediment production since they improved the results of the SWAT model, which also requires more time and data to set up and calibrate. However, since SWAT does not require observed streamflow as an input, it remains a useful model, achieving acceptable results in basins with limited streamflow data.