How will the next-generation of sensor-based decision systems look in the context of intelligent agriculture? A case-study
André Freitas Colaço, Jonathan Richetti, R. G. V. Bramley, Roger Lawes
Abstract
The development of cost-effective, digitally based decision support systems is a key challenge in the optimization of farm management. Yet, the majority of sensor-based decision tools which support fertiliser management have relied on simplistic mechanistic frameworks normally informed by a single sensor. This study used a 20-year nitrogen (N) experiment on winter wheat (Triticum aestivum L.) to test a range of approaches for N decision support systems, including commercial sensor-based options and a novel, multivariate, data-driven approach. The latter was based on a non-mechanistic framework in which various digital variables were trained directly against optimum N application rates using machine learning. It was hypothesized that such a method would enhance our ability to handle system complexity, resulting in higher accuracy for the decision, as compared to current farm management or to available sensor-based options, both of which are normally underpinned by mechanistic methods. Results showed that the proposed approach was able to predict the optimal N rate with an RMSE of 16.5 kg N ha–1 (R2 = 0.79). This method was also the only one that was statistically superior (p < 0.05) to the control scenario (the application of the historical average optimal N rate; RMSE =38.0 kg N ha–1). This proposed approach used a multivariate digital input including a spectral vegetation index (normalized difference vegetation index, NDVI), weather and soil moisture data and information from on-farm experimentation (the in-situ N response using a ‘N-rich’ strip) to guide the decision. When similar data input and modelling techniques were used to predict yield potential to then derive an N recommendation through a mechanistic decision framework – a nutrient mass balance – the recommendation error (RMSE) increased to 26.0 kg N ha−1 (R2 = 0.51). In summary, by forcing the input data through the mechanistic framework, the decision error increased. This study challenges the ideas that farm decisions should follow pre-established agronomic mechanistic frameworks and that digital technologies must necessarily be used to estimate specific crop and soil attributes so as to enable deployment of current decision systems at scale and site-specifically.