IAHR Document Library

« Back to Library Homepage « Proceedings of the 40th IAHR World Congress (Vienna, 2023)

A Deep-Analysis of Soil Moisture, with Remotely Sensed Exogenous Variables, Case Study, Application over Quebec, Canada

Author(s): Mohammad Zeynoddin; Silvio Jose Gumiere; Hossein Bonakdari

Linked Author(s): Hossein Bonakdari

Keywords: LSTM; Precipitation; Temperature; Deep learning; GEE

Abstract: Soil moisture is one of the critical parameters in hydrological studies. This vital parameter affects soil erosion resilience as an external environment characteristic. Runoff and sediment transportation increase as soil moisture increases, though other soil characteristics also affect erosion. Hydraulic structures and watershedscaled studies depend on runoff, sediment, and erosion variables. In a vast country like Canada, soil moisture measurement limitations constitute a significant issue. Thus, a method should be investigated to obtain this valuable parameter in data-scarce regions and model it with potent methods. This study examines the possibility of forecasting the SSMC dataset using air temperature (ATemp) and precipitation (PrpT) time series as exogenous input variables and the long-short-term memory (LSTM) model. These models are examined for a Quebec site. The ACF tool and target series structure information tune the model structure. Using PrpT and Atemp in LSTM with five hidden states (h) yields correlation coefficient (R) = 0.894, root mean square error (RMSE) = 3.233, unbiased RMSE (uRMSE) = 3.220, and Nash-Sutcliffe efficiency (NSE) = 0.763. This model is more parsimonious than others. LSTM (PrpT+ATemp, h52) outperforms the previous model slightly by complicating the model structure. Using historical SSMC data as mere inputs of the LSTM instead of exogenous variables reduces input measurement errors, increases model accuracy by 7% in R, 2% in uRMSE, and maintains model efficiency. The Box and distribution plots of the forecasts versus real values show that this model overestimates the SSMC minimums, but it is more accurate than the other two in forecasting the fluctuations of the peaks. The PrpT+ATemp combination is also validated by input sensitivity analysis.


Year: 2023

Copyright © 2024 International Association for Hydro-Environment Engineering and Research. All rights reserved. | Terms and Conditions