Soil mapping for precision agriculture using support vector machines combined with inverse distance weighting

Kriging has been shown to be the best interpolator to interpolate maps in precision agriculture. However, Kriging requires a high number of sampling points to generate accurate maps. Recently, machine learning (ML) techniques have shown the potential to produce maps with a lower number of sampling points. In addition, using ML map generation can be automated and use much more feature information to improve map quality. Therefore, the objective of this study was to implement a ML technique and compare it to IDW and to Ordinary Kriging (OK). The ML algorithm used was the Support Vector Machine (SVM). Software based on the SVM method was developed (Smart-Map) using the Python language. This software was tested in an area of 204 ha cultivated with soybeans. The performance of the SVM method was compared to traditional interpolation methods, IDW and Ordinary Kriging (OK). Based on the analysis of 10 soil attributes, OK had better performance than IDW and the ML method when the Moran's I (Index) values were significant and higher than 0.67. With a low density of points and low degrees of spatial autocorrelation, the ML method performed better than IDW and OK.

Automated summary

Kriging is the optimal interpolator for precision agriculture, but it requires a high number of sampling points for accurate map generation. Machine learning techniques have shown potential in generating maps with fewer sampling points. In this study, a Support Vector Machine (SVM) algorithm was implemented and compared to IDW and Ordinary Kriging (OK). Results showed that OK outperformed IDW and the ML method when Moran's I values were significant and higher than 0.67. However, the ML method performed better than IDW and OK in situations with low density of points and low degrees of spatial autocorrelation.