Predicting spatiotemporal yield variability to aid arable precision agriculture in New Zealand : a case study of maize-grain crop production in the Waikato region : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Agriculture and Horticulture at Massey University, Palmerston North, New Zealand

Abstract

Precision agriculture attempts to manage within-field spatial variability by applying suitable inputs at the appropriate time, place, and amount. To achieve this, delineation of field-specific management zones (MZs), representing significantly different yield potentials are required. To date, the effectiveness of utilising MZs in New Zealand has potentially been limited due to a lack of emphasis on the interactions between spatiotemporal factors such as soil texture, crop yield, and rainfall. To fill this research gap, this thesis aims to improve the process of delineating MZs by modelling spatiotemporal interactions between spatial crop yield and other complementary factors.

Data was collected from five non-irrigated field sites in the Waikato region, based on the availability of several years of maize harvest data. To remove potential yield measurement errors and improve the accuracy of spatial interpolation for yield mapping, a customised filtering algorithm was developed. A supervised machine-learning approach for predicting spatial yield was then developed using several prediction models (stepwise multiple linear regression, feedforward neural network, CART decision tree, random forest, Cubist regression, and XGBoost). To provide insights into managing spatiotemporal yield variability, predictor importance analysis was conducted to identify important yield predictors.

The spatial filtering method reduced the root mean squared errors of kriging interpolation for all available years (2014, 2015, 2017 and 2018) in a tested site, suggesting that the method developed in R programme was effective for improving the accuracy of the yield maps. For predicting spatial yield, random forest produced the highest prediction accuracies (R² = 0.08 - 0.50), followed by XGBoost (R² = 0.06 - 0.39). Temporal variables (solar radiation, growing degree days (GDD) and rainfall) were proven to be salient yield predictors. This research demonstrates the viability of these models to predict subfield spatial yield, using input data that is inexpensive and readily available to arable farms in New Zealand. The novel approach employed by this thesis may provide opportunities to improve arable farming input-use efficiency and reduce its environmental impact.