Comparative Analysis of Spatial Statistical Methods for Modelling Rice Yield Variability in Uniformity Trials

Spatial heterogeneity is a fundamental characteristic of agricultural landscapes, and accurately modelling its impact on crop yield is critical for precision agriculture. This study presents a comparative analysis of spatial statistical and machine learning approaches to model yield variability in a field exhibiting strong spatial dependence. Initial geostatistical analysis confirmed significant positive spatial autocorrelation (Moran's I = 0.7355, p < 2.2e-16). A series of models were evaluated, progressing from linear and polynomial trend surfaces to spatial regression (Lag and Error), mixed effects, and a Random Forest ensemble. The results demonstrate that while traditional spatial models effectively captured and explained the spatial autocorrelation, a Random Forest model significantly outperformed all others in predictive accuracy, achieving a 31% lower Root Mean Square Error (36.05) than the best spatial regression model and explaining 71.02% of the variance. Among the spatial models, a quadratic trend surface with a row × column interaction provided the most parsimonious fit, revealing an accelerating yield increase along rows and a U-shaped pattern across columns. The Random Forest model's residuals showed no statistically significant remaining spatial autocorrelation (Moran's I p-value = 0.97); however, we caution that failure to reject spatial randomness does not prove its absence. Equivalence testing (TOST) and power analysis support the conclusion that spatial dependence was adequately captured. We conclude that machine learning ensembles like Random Forest offer superior predictive power for yield mapping, though spatial regression models remain valuable for providing parsimonious, interpretable insights into the underlying structural gradients driving yield variability.