The use of satellite and radar products and machine learning to improve hydrological modelling in Reola watershed, Estonia

Código

I-EBHE0070

Autores

felipe bortolletto civitate, EMILIO GRACILIANO FERREIRA MERCURI

Tema

WG 1.12: Development & application of river basin simulators

Resumo

Almost half of Estonia's land area is covered with hemi-boreal forests, which play a significant role in the country's economy and ecology. Estonia is the fourth nation in the EU in terms of area per capita, with a value of 1.70 ha/capita. Over the past century, Estonian forests have undergone substantial changes, with the forest area tripling and becoming a cornerstone of the national economy. Given the economic dependence on forestry and the ecological importance of these forests, accurate hydrological forecasting models are vital for informed decision-making in water resource management. This study focuses on the Reola Basin, a 237 km² catchment in southeast Estonia, draining into the Emajõgi River. The main objective of the research is to develop ways to improve hydrological modeling with time and space varying hydrological big data, such as radar precipitation data and satellite evapotranspiration. Data from two external meteorological stations, Tartu-Tõravere and Võru, along with the EURADICLIM radar dataset, were utilized in conjunction with the Reola hydrometric station used as the basin's outlet. Evapotranspiration data was collected from the MODIS sensor of the Terra satellite and from de Station for Measuring Ecosystem-Atmosphere Relations (SMEAR Estonia). Modis evapotranspiration was corrected with SMEAR Estonia H20 fluxes made by Eddy Covariance technique. The year 2019 was used for model warm-up, model calibration was done in 2012-2018 and the year 2020 was used for validation. The research compares the performance of the GR4J-Cemaneige conceptual hydrological model and the Long Short Term Memory (LSTM) machine learning architecture using precipitation, evapotranspiration, and temperature inputs. Performance was assessed using the Nash-Sutcliffe efficiency metric for three scenarios of precipitation: pluviometer data, uncorrected radar data, and radar data corrected with the Double Mass Curve method. Results for the GR4J-Cemaneige model showed Nash values of 0.8025 (pluviometers), 0.3715 (uncorrected radar), and 0.5788 (corrected radar). For the LSTM model, the values were 0.6952 (pluviometers), 0.4086 (uncorrected radar), and 0.7240 (corrected radar). Additionally, incorporating the previous day's flow data into the LSTM model yielded improved Nash values of 0.9425 (pluviometers), 0.9217 (uncorrected radar), and 0.9450 (corrected radar). These results indicate that while radar precipitation data align with observed phenomena, they do not yet surpass the performance of pluviometer data. The Double Mass Curve corrections significantly improved the quality of radar-derived precipitation data. Comparing the models, the GR4J model performed better with pluviometer data, while the LSTM model excelled with radar data. However, the LSTM model's ability to utilize short- and long-term memory with previous day's flow data demonstrated exceptional performance, outstanding the other two applications. We corroborate that machine learning empirical methods such as LSTM have become more accurate tools for hydrological modeling compared to process-based models. Also, the physical interpretability of the results from these black boxes are becoming accessible, including other parts of the hydrological cycle like soil moisture and snowpack variations. This research allows a better understanding of the water balance dynamics in hemi-boreal ecosystems, an area that is changing its hydrological response due to climate change.