Determining optimal digital soil mapping components for exchangeable calcium and magnesium across a sugarcane field

Abstract

Understanding the spatial distribution of exchangeable calcium (Exch. Ca) and magnesium (Exch. Mg) at field level is a fundamental component in managing fertilizer application for sugarcane farming. This information can potentially be created by digital soil mapping (DSM) protocols; using a mathematical model to couple soil and ancillary data (i.e., gamma-ray (γ-ray) and apparent electrical conductivity (ECa)). In this research, we aim to show which mathematical model (i.e. linear mixed model (LMM), regression kriging (RK), random forests (RF) and supportive vector machine (SVM)) was best, how many calibration samples were required, which ancillary data and what transect spacing was suitable. In terms of soil sampling, we collected 182 samples across a sugarcane field with 42 samples removed for validation. The remaining 140 samples were used along with a conditioned Latin Hypercube Sampling (cLHS) analysis of the ancillary data, to generate nine calibration datasets (i.e. n = 10–140). The ancillary data was collected on 7.5 m transect spacing and we compared 15, 30, 45 and 60 m spacing. The comparisons between models, sample size, ancillary data and transect spacing were made using Lin's concordance correlation coefficient (LCCC), root mean square error (RMSE) and mean error (ME). Results indicate that to predict Exch. Ca and Mg, the best model in terms of strongest LCCC and smallest RMSE was LMM, followed by RK, RF and SVM. A total of 60 samples were enough to permit the development of accurate predictions of Exch. Ca given all models had strong LCCC (> 0.8) and RMSE less than half the standard deviation (0.07). For Exch. Mg, 80 samples were required. However, from a farm management perspective, 30 samples with LMM was satisfactory to predict Exch. Ca (LCCC = 0.83) with 40 (0.84) sufficient for Exch. Mg. With respect to the ancillary data, ECa (with LMM and n = 60) was more accurate (RMSE = 0.06), less biased (ME = 0.003) and had stronger LCCC (0.85) compared to γ-ray data alone. However, using both ancillary data in combination was most accurate (0.05) and had the strongest LCCC (0.87). Similar findings were observed for Exch. Mg. The use of 7.5 m transect spacing was best, but 15 m and 30 m transect spacing also give accurate, unbiased and precise predictions.