Machine Learning (ML) algorithms have been used as an alternative to conventional and geostatistical methods in digital mapping of soil attributes. An advantage of ML algorithms is their flexibility to use various layers of information as covariates. However, ML algorithms come in many variations that can make their application by end users difficult. To fill this gap, a Smart-Map plugin, which complements Geographic Information System QGIS Version 3, was developed using modern artificial intelligence (AI) tools. To generate interpolated maps, Ordinary Kriging (OK) and the Support Vector Machine (SVM) algorithm were implemented. The SVM model can use vector and raster layers available in QGIS as covariates at the time of interpolation. Covariates in the SVM model were selected based on spatial correlation measured by Moran’s Index (I’Moran). To evaluate the performance of the Smart-Map plugin, a case study was conducted with data of soil attributes collected in an area of 75 ha, located in the central region of the state of Goiás, Brazil. Performance comparisons between OK and SVM were performed for sampling grids with 38, 75, and 112 sampled points. R² and RMSE were used to evaluate the performance of the methods. SVM was found superior to OK in the prediction of soil chemical attributes at the three sample densities tested and was therefore recommended for prediction of soil attributes. In this case study, soil attributes with R² values ranging from 0.05 to 0.83 and RMSE ranging from 0.07 to 12.01 were predicted by the methods tested.

Precision agriculture techniques contribute to optimizing the use of agricultural inputs, as they consider the spatial and temporal variability in the production factors. Prescription maps of limestone and fertilizers at variable rates (VRA) can be generated using various soil sampling techniques, such as point grid sampling, cell sampling, and management zone sampling. However, low-density grid sampling often fails to capture the spatial variability in soil properties, leading to inaccurate fertilizer recommendations. Sampling techniques by cells or management zones can generate maps of better quality and at lower costs than the sampling system by degree of points with low sampling density. Thus, this study aimed to compare the accuracy of different sampling techniques for mapping soil attributes in precision agriculture. For this purpose, the following sampling techniques were used: high-density point grid sampling method, low-density point grid sampling method, cell sampling method, management zone sampling method, and conventional method (considering the mean). Six areas located in the Brazilian states of Bahia, Minas Gerais, Mato Grosso, Goias, Mato Grosso do Sul, and Sao Paulo were used. The Root-Mean-Square-Error (RMSE) method was determined for each method using cross-validation. It was concluded that the cell method generated the lowest error, followed by the high-density point grid sampling method. Management zone sampling showed a lower error compared to the low-density point grid sampling method. By comparing different sampling techniques, we demonstrate that management zone and cell grid sampling can reduce soil sampling while maintaining comparable or superior accuracy in soil attribute mapping.

Machine Learning (ML) algorithms have been used as an alternative to conventional and geostatistical methods in the digital mapping of soil attributes. However, ML algorithms have many variants that can make their application difficult for end users. To fill this gap, the objective of this work was to develop the Smart-Map plugin, to be used in QGIS. The Ordinary Kriging (OK) geostatistical method and the ML Support Vector Machine (SVM) model were implemented in the plugin to generate interpolated maps based on machine learning and OK. Performance comparisons between OK and SVM were performed for sample grids with 112 sampled points. Algoritmos de Aprendizado de Máquina (ML) têm sido utilizados como alternativa aos métodos convencionais e geoestatísticos no mapeamento digital de atributos do solo. No entanto, os algoritmos de ML apresentam muitas variantes que podem dificultar a sua aplicação por usuários finais. Para preencher essa lacuna, o objetivo desse trabalho foi desenvolver o plugin Smart-Map, para ser utilizado no QGIS. Foram implementados no plugin o método geoestatístico Krigagem Ordinária (OK) e o modelo de ML Support Vector Machine (SVM), para geração de mapas interpolados com base em aprendizado de máquina e OK. Comparações de performance entre OK e SVM foram realizadas em um grid amostral com 112 pontos amostrados.