Hafsa Mahmood

Temporal drain flow dynamics and understanding of their underlying controlling factors are important for water resource management in tile-drained agricultural areas. The use of physics-based water flow models to understand tile drained systems is common. These models are complex, with large parameter sets and require high computational effort. The primary goal of this study was to examine whether simpler, more efficient machine learning (ML) models can provide acceptable solutions. The specific aim of our study was to assess the potential of ML tools for predicting drain flow time series in multiple catchments subject to a range of climatic and landscape conditions. The investigation is based on unique data containing time series of daily drain flow in multiple field scale drain sites in Denmark. The data include: climate (precipitation, potential evapotranspiration, temperature); geological properties (clay fraction, first sand layer thickness, first clay layer thickness); and topographical indexes (curvature, Topographical wetness indexes, Topographical position index, elevation). Both static and dynamic variables are used in the prediction of drain flows. The ML algorithm extreme gradient boosting (XGBoost) and convolutional neural network (CNN) were examined, and the results were compared with a physics-based distributed model (MIKE-SHE). The results show that XGBoost performs similarly to the physics-based MIKE-SHE models, and both outperform CNN. Both ML models required significantly less effort to build, train, and run than MIKE-SHE. In addition, the ML models support efficient feature importance analysis. This showed that climatic variables were important for CNN models and XGBoost. The results support the use of ML models for hydrologic applications with sufficient data for training. Further, the insights offered by the feature importance analysis may support further data collection and developments of physics-based models when existing data are insufficient to support ML approaches.

Temporal drain flow dynamics and understanding of their underlying controlling factors are important for water resource management in tile-drained agricultural areas. The use of physics-based water flow models to understand tile drained systems is common. These models are complex, with large parameter sets and require high computational effort. The primary goal of this study was to examine whether simpler, more efficient machine learning (ML) models can provide acceptable solutions. The specific aim of our study was to assess the potential of ML tools for predicting drain flow time series in multiple catchments subject to a range of climatic and landscape conditions. The investigation is based on unique data containing time series of daily drain flow in multiple field scale drain sites in Denmark. The data include: climate (precipitation, potential evapotranspiration, temperature); geological properties (clay fraction, first sand layer thickness, first clay layer thickness); and topographical indexes (curvature, Topographical wetness indexes, Topographical position index, elevation). Both static and dynamic variables are used in the prediction of drain flows. The ML algorithm extreme gradient boosting (XGBoost) and convolutional neural network (CNN) were examined, and the results were compared with a physics-based distributed model (MIKE-SHE). The results show that XGBoost performs similarly to the physics-based MIKE-SHE models, and both outperform CNN. Both ML models required significantly less effort to build, train, and run than MIKE-SHE. In addition, the ML models support efficient feature importance analysis. This showed that climatic variables were important for CNN models and XGBoost. The results support the use of ML models for hydrologic applications with sufficient data for training. Further, the insights offered by the feature importance analysis may support further data collection and developments of physics-based models when existing data are insufficient to support ML approaches.

Temporal drain flow dynamics and understanding of their underlying controlling factors are important for water resource management in tile-drained agricultural areas. The use of physics-based water flow models to understand tile drained systems is common. These models are complex, with large parameter sets and require high computational effort. The primary goal of this study was to examine whether simpler, more efficient machine learning (ML) models can provide acceptable solutions. The specific aim of our study was to assess the potential of ML tools for predicting drain flow time series in multiple catchments subject to a range of climatic and landscape conditions. The investigation is based on unique data containing time series of daily drain flow in multiple field scale drain sites in Denmark. The data include: climate (precipitation, potential evapotranspiration, temperature); geological properties (clay fraction, first sand layer thickness, first clay layer thickness); and topographical indexes (curvature, Topographical wetness indexes, Topographical position index, elevation). Both static and dynamic variables are used in the prediction of drain flows. The ML algorithm extreme gradient boosting (XGBoost) and convolutional neural network (CNN) were examined, and the results were compared with a physics-based distributed model (MIKE-SHE). The results show that XGBoost performs similarly to the physics-based MIKE-SHE models, and both outperform CNN. Both ML models required significantly less effort to build, train, and run than MIKE-SHE. In addition, the ML models support efficient feature importance analysis. This showed that climatic variables were important for CNN models and XGBoost. The results support the use of ML models for hydrologic applications with sufficient data for training. Further, the insights offered by the feature importance analysis may support further data collection and developments of physics-based models when existing data are insufficient to support ML approaches.

Agricultural areas are often artificially drained, especially in temperate and flat landscapes. This also applies to Denmark, where approximately half of the agricultural area is artificially drained, mostly with tile drains. The generated drain flow has significant impacts on various aspects of the hydrologic cycle such as groundwater recharge, flow paths and transport times. Consequently, drain flow is a major control on the transport of nutrients such as nitrogen. Yet, detailed knowledge of spatial and temporal variability of drain flow is inadequate due to insufficient observations of drain flow, lacking knowledge of drain infrastructure and issues of scale and hydrogeologic heterogeneity.The objective was to improve the simulation of both the spatial and temporal variability of drain flow in a large-scale hydrological model used to map nitrate transport. This model is a physically-based, distributed groundwater-surface water model of all of Denmark. It is a major challenge to simulate drain flow distribution in space and time with the national model due to its coarse horizontal resolution (500m or 100m), and the lack of drain flow observations at relevant scale. Hence, to achieve the objective, we gathered existing field-scale drain flow observations from all over Denmark. For these drain catchments, fine-scale (10m) physically based hydrological models were setup and calibrated against the drain flow observations. After successful calibration, the resulting simulated distributions of drain fraction (drain flow relative to precipitation) were regionalized to applicable areas across all of Denmark. The regionalization was performed using decision tree machine learning algorithms, and a set of topographic and geologic covariates available nationally at fine resolution. An analysis of spatial transferability of the machine learning algorithm allowed to limit predictions to applicable areas. Finally, these estimates of drain fraction are used in the calibration of the large-scale national hydrologic model, amongst other objective functions such as streamflow and groundwater heads.

Temporal drain flow dynamics and their underlying controlling factors are important for understanding the needs for water resource management in tile drained agricultural areas. The use of physics-based water flow models to understand tile drained systems is quite common. These physics-based models are complex and have high computational demand due to the high spatial and temporal dimensionality of the problem. We examine whether machine learning (ML) models can offer a simpler tool for water management.   The main aim of our study is to assess the potential of ML tools for predicting drain flow with varying climate parameters and hydrogeological properties in different catchments in Denmark. We rely on unique data containing time series of daily drain flow in 26 field-scale tile drained catchments in Denmark: climate data (precipitation, potential evapotranspiration, temperature); geological properties (clay fraction, first sand layer thickness, first clay layer thickness); and topographical indexes (curvature, topographical wetness indexes, topographical position index, elevation etc.). The ML algorithm XGBoost is used to predict drain flow in the 26 drain catchments based on both static and dynamic variables. This algorithm also provides an independent measure of the value of information contained in variables related to climate, geology and topography for the prediction of tile drain flows.   The ML approach examined could provide a more transferable, faster, and less computationally expensive tool to predict drain flow dynamics. Simultaneously, the results of the study offer insight into the underlying factors that control drain flow, allowing for improved data collection and physics-based model development.