Gaussian Process Regression as an Alternative to Kriging and SVM for Spatial Yield Prediction

Abstract

Detecting spatial yield variability is essential for precision agriculture, as it reduces environmental impact and improves economic returns. This study evaluates Gaussian Process Regression (GPR), Ordinary Kriging (OK), and Support Vector Machine (SVM) under different sampling densities. GPR and OK perform similarly, with GPR showing a slight advantage in low-sampling conditions. With 322 samples, GPR achieves higher accuracy (RMSE = 0.64 t/ha, R² = 0.68) than OK (RMSE = 0.72 t/ha, R² = 0.60), while SVM performs worse (RMSE = 0.76 t/ha, R² = 0.55). Regardless of sample size, SVM-generated maps exhibit a smoothing effect, reducing sensitivity to local variations. OK remains effective but is more sensitive to sample density due to its reliance on the semivariogram model and the assumption of isotropy. These findings highlight GPR as a robust method for spatial yield prediction, particularly in sparse data conditions. The study was conducted in Patos de Minas, Brazil, using 795 georeferenced soybean yield samples over 3.7 hectares. From a practical perspective, GPR and OK remain strong candidates for yield interpolation, reinforcing the importance of model selection based on data availability and spatial variability.