APPLICATION OF MACHINE LEARNING TECHNIQUES TO PREDICT NITROGEN LEVELS IN THE MISSISSIPPI RIVER BASIN

Nutrient loss from agricultural activities is elevating in-stream nutrient levels throughout the Midwestern U.S., often to the point of risks to human health and with devastating ecological consequences. Nutrient losses are known to be mediated by weather patterns, wherein extended droughts result in the temporary storage of nutrients on the landscape which are mobilized by subsequent rainfall. Communities and industries that source water from rivers rely on estimates of nitrogen levels to plan water treatment operations or activate groundwater pumping to cope with high nutrient loads. Thus, a predictive model that can predict the range nitrogen levels in surface waters is desired. This study presents an Artificial Neural Network model (ANN) with back-propagation that can generate long range predictions of the range within which nitrogen levels will fall. The model leveraged quarter hour measurements collected by USGS and other organizations in watersheds in Iowa. Drainage areas for these watersheds in Iowa range from a few square miles to thousands of square miles. The study evaluated the influence exerted by different information inputs as well as the influence of past measurements on making predictions. Experimenting with different random initializers, regularizers, learning rates, and network models has resulted in a simple model that is accurate 98% of the time. Overall this study documents the workflow to develop an ANN predictor for environmental systems, evaluates ANN performance across orders of magnitude in spatial scale, and evaluates the key information required to make accurate predictions as well as the memory of the catchment.