Search

This study aimed to evaluate the use of multiple covariates in robust geostatistical modeling of soil chemical properties characterized by the presence of outliers. Different spatial prediction methods were compared using data from two agricultural areas located in Brazil´s Southeast: one with rotational grazing and one cultivated with sugarcane. Considering the variable-rate fertilizer prescription in the context of precision agriculture, the use of multiple covariates for the prediction of four chemical soil properties (phosphorus (P), potassium (K), cation exchange capacity (CEC) and base saturation (V)) was evaluated. The covariates data set was divided into five categories representing soil, vegetation, relief, management of the area and geographic. Five methods were used: inverse distance weighting (IDW), robust multiple linear regression (RMLR), robust ordinary kriging (ROK), robust universal kriging with spatial co-ordinates in the trend (RUKcoord) and robust universal kriging with environmental and management covariates in the trend (RUKcovars). The model based on the mean was used as a null reference. In general, the use of covariates in robust prediction methods improves the accuracy of spatial prediction of soil properties in the presence of outliers. However this effect was not observed in all situations, depending on the dataset characteristics and the spatial variability of the fields. The management practices are important information for modeling the trend in digital soil mapping for fertilizer prescription purposes. RMLR produces prediction results that are, at least, equivalent to that of robust geoestatistics. [ABSTRACT FROM AUTHOR]

• Machine learning and geostatistical models for soil mapping. • Relationship between soil fertility and environmental/management covariates. • A trade off between spatial cover grids and sample planning optimization. • The benefits of using covariates in sample optimization depend on field management. • As higher the knowledge about field variability, the more accurate is the optimization. The study aimed to evaluate whether the use of covariates in sample optimization helps characterize the spatial variability of soil chemical properties. To this end, data from two agricultural areas located in southeastern Brazil were used. The covariates were used at two stages: sample planning optimization and building predictive methods. The covariates were divided into five categories: relief, vegetation, soil, management, and spatial coordinates. We compared two criteria in sample planning optimization: one aims to ensure spatial distribution by minimizing the mean squared shortest distance (MSSD); the other integrated four criteria through a multipurpose function (SPAN), aiming to represent the distribution and correlation between the covariates, the minimization of the mean squared shortest distance and the distribution of pairs of points for each lag size of the semivariogram. We evaluated three predictive methods (Robust Ordinary Kriging (ROK), Random Forest (RF), and Random Forest-Robust Kriging (RF-RK)) and a null model (mean-based) to predict four soil properties (Cation exchange capacity (CEC), Base saturation (V), Potassium (K) and Phosphorus (P)). We observed that not all areas would benefit from the inclusion of covariates in the sample planning. When our previous knowledge of variability is limited or unknown, we need to evaluate the management in the areas. When the cropping system is more complex and diversified, spatial coverage sampling is more feasible, using the covariates only on the prediction models. [ABSTRACT FROM AUTHOR]