Does using soil chemical variables in cokriging improve the spatial modelling of the commercial wood volume of Brazilian mahogany in an Amazonian agroforestry system?

•The commercial wood volumes of the individual trees were predicted by ANN.•Spatial modelling was performed by ordinary kriging and cokriging.•Correlation and principal component analyses were used to select the soil chemical variables.•Secondary variables selected by correlation analysis improve the precision of spatial modelling.•Calcium and pH in KCl are more correlated with commercial wood volume of Brazilian mahogany. Soil chemical variables are among the main factors that influence forest production; however, there is no consensus on which soil variables are most correlated with the individual volume spatial variability. Moreover, no studies that used the ordinary cokriging geostatistical technique to model the variability of the volume with soil chemical variables correlated as secondary variables in agroforestry systems were found. For these reasons, the objective of this study was to test whether the precision of the spatial modelling of the Brazilian-mahogany commercial wood volume (vc) by ordinary cokriging could be improved by adding soil chemical variables as secondary variables compared to ordinary kriging. Therefore, soil samples were collected at the centre of 36 georeferenced circular plots with approximate areas of 500 m2 to determine the soil chemical variables. In these plots, 108 standing trees were scaled to compute vc using Smalian’s formula. Subsequently, artificial neural networks were trained using the diameter at 1.3 m above the soil level and the commercial height of these sample trees to predict the vc of the remaining trees within the plots. Lastly, vc was spatially modelled by ordinary kriging (scenario one), ordinary cokriging using latent variables created from principal component analysis (scenario two), and soil chemical variables significantly correlated with vc (scenario three) as secondary variables. In all the scenarios, the exponential geostatistical model stood out as the best, as it presented better spatial dependence index and precision measures in the leave-one-out cross-validation process. However, none of the models applied in scenario two were good alternatives because of the lack of strong spatial dependence. The scenario-three approach proved to be the best alternative for interpolating vc, followed by scenario one.