Keywords

Regression; Features Importance Assessment; Deep feed-forward; Neural Network; Water Quality

Start Date

5-7-2022 12:00 PM

End Date

8-7-2022 9:59 AM

Abstract

Machine Learning (ML) is receiving growing attention in hydrology and environmental sciences as well for the flexibility these techniques offer and the capability of learning the model structure directly from the data. However, in hydrology ML has not been fully exploited yet for the inherent complexity of the investigated natural processes and limited data availability that still characterizes some areas of the hydrological sciences, such as subsurface hydrology for example. In river water quality the capability of ML in identifying relevant drivers of change has been only partially explored. In particular, studies dealing with water quality variables not showing linear relationships with the drivers, for example, the case of Arsenic in surface and subsurface water bodies, are limited. To address these gaps, we investigate the impact of external drivers on water quality parameters in two contrasting European river basins: the Adige river, the second-longest Italian river, and the Ebro river, the largest Iberian river basin. The investigation involves extended seasonal time series (1994-2013) of quality parameters at 45 locations within the Adige river basin, and 42 within the Ebro river basin. We test a Dense feed-forward Neural Network model, selecting the hidden neurons and layers according to the Akaike criterion. We perform an Importance Features Assessment to highlight the relative weight of the drivers in the predictions. The model shows Nash Sutcliffe Efficiency (NSE) coefficients ranging from 0.68 to 0.78 for the dissolved minerals and the electrical conductivity and of 0.89 for the water temperature for the Adige, and with NSE values higher than 0.80 for the Ebro. From the sensitivity analysis of the stressors used in the prediction of the water quality targets, the most evident conclusion is that the spatial and temporal variability of the selected variables can be detected by seasonal and temporal information of the measurements. This indicates that the external forcing employed in the regression should be carefully evaluated and that, in some cases, with basic environmental information or information easy to achieve a good prediction of the variable can be obtained.

Stream and Session

false

Share

COinS
 
Jul 5th, 12:00 PM Jul 8th, 9:59 AM

River water quality modeling at catchment-scale using Deep feed-forward Neural Network

Machine Learning (ML) is receiving growing attention in hydrology and environmental sciences as well for the flexibility these techniques offer and the capability of learning the model structure directly from the data. However, in hydrology ML has not been fully exploited yet for the inherent complexity of the investigated natural processes and limited data availability that still characterizes some areas of the hydrological sciences, such as subsurface hydrology for example. In river water quality the capability of ML in identifying relevant drivers of change has been only partially explored. In particular, studies dealing with water quality variables not showing linear relationships with the drivers, for example, the case of Arsenic in surface and subsurface water bodies, are limited. To address these gaps, we investigate the impact of external drivers on water quality parameters in two contrasting European river basins: the Adige river, the second-longest Italian river, and the Ebro river, the largest Iberian river basin. The investigation involves extended seasonal time series (1994-2013) of quality parameters at 45 locations within the Adige river basin, and 42 within the Ebro river basin. We test a Dense feed-forward Neural Network model, selecting the hidden neurons and layers according to the Akaike criterion. We perform an Importance Features Assessment to highlight the relative weight of the drivers in the predictions. The model shows Nash Sutcliffe Efficiency (NSE) coefficients ranging from 0.68 to 0.78 for the dissolved minerals and the electrical conductivity and of 0.89 for the water temperature for the Adige, and with NSE values higher than 0.80 for the Ebro. From the sensitivity analysis of the stressors used in the prediction of the water quality targets, the most evident conclusion is that the spatial and temporal variability of the selected variables can be detected by seasonal and temporal information of the measurements. This indicates that the external forcing employed in the regression should be carefully evaluated and that, in some cases, with basic environmental information or information easy to achieve a good prediction of the variable can be obtained.