Krigagem Ordinária vs. Support Vector Machine no Smart-Map do QGIS: Qual método é mais acurado na descrição espacial da produtividade da soja?

Abstract

One of the major challenges in precision agriculture is detecting the spatial variability of the productive unit. The higher the accuracy in describing yield spatially, the lower the environmental impact and the greater the economic returns. To achieve this goal, dense sampling across the field is often conducted, which can increase the costs of implementing an agricultural management system. Machine learning (ML) algorithms and geostatistical methods can describe spatial variability even under reduced sampling conditions. However, these models exhibit various complexities that may hinder their application by end users. To bridge this gap, the Smart-Map plugin was developed to complement QGIS, simplifying the use of these methods and facilitating their practical application. However, it is still unclear in the literature which of these methods is more effective. In this study, we evaluated Ordinary Kriging (OK) compared to Support Vector Machine (SVM) in predicting soybean yield at different sampling levels: 100% (795 points), 80% (636 points), 60% (477 points), and 40% (318 points). The results indicate that OK generally outperformed SVM, with an average 𝑅2 of 0.6845 and RMSE of 0.609 tons/hectare, compared to an average 𝑅2 of 0.5315 and RMSE of 0.73 for SVM. At the full sampling level (100%), the OK showed the best results, with 𝑅2 = 0,819 and RMSE = 0,47 tons/hectare, whereas SVM achieved 𝑅2 = 0,704 and RMSE = 0,599 tons/hectare. Under reduced sampling conditions (40%), the OK maintained relatively consistent performance, while SVM experienced a significant drop in 𝑅2 to 0.394, with an increase in RMSE to 0.903 tons/hectare. These findings suggest that although OK requires more specialized knowledge, it is more robust under reduced sampling scenarios compared to SVM, which could benefit from additional input attributes beyond yield position. Moreover, SVM performance could improve if Smart-Map incorporated other covariance functions, such as the Rational Quadratic, providing a more accurate modeling of spatial variability.