A machine learning based modelling framework to predict nitrate leaching from agricultural soils across the netherlands

Abstract

Throughout recent decades, the excessive use of animal manure and fertiliser put a threat on the quality of ground and surface waters in main agricultural production areas in Europe and other parts of the world. Finding a balance between agricultural production and environmental protection is a prerequisite for sustainable development of ground and surface waters and soil quality. To protect groundwater quality, the European Commission has stipulated a limit value for NO3− of 50 mg l−1. Member states are obliged to monitor and regulate nitrate concentrations in groundwater. In the Netherlands, this monitoring is carried out by sampling nitrate concentrations in water leaching from the root zone at farm level within the national Minerals Policy Monitoring Program. However, due to the costly procedure, only a limited number of about 450 farms can be sampled each year. While this is sufficient for providing a national overview of nitrate leaching, as a result of current and future challenges regarding the sustainable development of the agricultural system, Dutch policymakers need to gain insight into the spatial distribution of nitrate at smaller spatial scales. This study aimed to develop a predictive modelling framework to create annual maps with full national coverage of nitrate concentrations leaching from the root zone of Dutch agricultural soils, and to test this model for the year 2017. We used nitrate data from a national monitoring program and combined them with a large set of auxiliary spatial data, such as soil types, groundwater levels and crop types. We used the Random Forest (RF) algorithm as a prediction and interpolation method. Using the model, we could explain 58% of variance, and statistical errors indicate that the interpolation and map visualisation is suitable for interpretation of the spatial variability of nitrate concentrations in the Netherlands. We used the variable importance from the RF and the partial dependency of the most important variables to get more insight into the major factors explaining the spatial variability. Our study also shows the caveats of data-driven algorithms such as RF. For some areas where no training data was available, the model’s predictions are unexpected and might indicate a model bias. The combination of visualisation of the spatial variability and the interpretation of variable importance and partial dependence results in understanding which areas are more vulnerable to NO3− leaching, in terms of land use and geomorphology. Our modelling framework can be used to target specific areas and to take more targeted regional policy measurements for the balance between agricultural production and protecting the environment.