Data-driven low-complexity nitrate loss model utilizing sensor information — Towards collaborative farm management with wireless sensor networks

Abstract

Excessive or poorly timed application of irrigation and fertilizers, coupled with the inherent inefficiency of nutrient uptake by crops result in nutrient fluxes into the water system. The ability to predict nutrient-rich discharges, in real time, can be very valuable to enable reuse mechanisms within farm systems. Wireless Sensor Networks (WSNs) offer an opportunity to monitor environmental systems with unprecedented temporal and spatial resolution. As part of our previous work, we proposed a novel framework (WQMCM) to combine increasingly common local farm-scale sensor networks across a catchment to learn and predict (using predictive models) the impact of catchment events on their downstream environments, allowing dynamic decision. Existing models use complex parameters which are difficult to extract and this, coupled with constraints on network nodes (battery life, computing power etc., availability of sensors) makes it necessary to develop simplified models for deployment within the networks. The paper investigates data-driven model for predicting daily total oxidized nitrate (TON) fluxes by seeking simplification in model parameters and using only a yearlong training data set. Data from a catchment in Ireland is used for training the model. Model simplification is investigated by abstracting details from an existing nitrate loss model. By using M5 decision tree model on the training samples of the proposed parameters, results give R2 as 0.92 and RRMSE as 0.26. The proposed novel model gives better results with fewer samples and simple parameters when compared to the traditional model. This shows promise for enabling real time nutrient control and management within the collaborative networked farm system.