A systematic evaluation of multisensor data and multivariate prediction methods for digitally mapping exchangeable cations: A case study in Australian sugarcane field

The acidic soil conditions in Australian sugarcane fields limit soil's capability to provide exchangeable (exch.) calcium (Ca) and magnesium (Mg). To replenish nutrient loss, the Six-Easy-Steps nutrient management guidelines were introduced to improve soil condition. However, laboratory analysis is required, and this is onerous. To value-add to soil data, digital soil maps (DSM) were developed by using two sources of digital data in three modelling approaches of geostatistical, statistical and hybrid. The effectiveness of digital data, either individually or combined, and varying number of calibration samples (n = 10-120) was tested for obtaining moderate agreement (i.e. Lin's = 0.65-0.80). The influence of varying number of calibration samples, was determined by calculating nugget to sill ratio (NSR) and used with ordinary kriging (OK) to make predictions. Moreover, four statistical approaches including one linear (i.e., multiple linear regression [MLR] and three nonlinear (i.e., Cubist, support vector machine [SVM], and random forest [RF]) were compared. Hybrid models were generated by adding regression residuals (RR) to the statistical models. In addition, we compare the final DSM's by calculating the mean square prediction error (MSPE) and determining whether a traditional soil map, was as accurate. Results showed that while n = 50 could compute a moderate variogram (i.e. NSR = 25-75%), n >_ 90 were required for stable results. Considering all data (i.e. n = 120), OK predicted Ca (0.69) and Mg (0.75) with agreement comparable to that of RF (0.64) and SVM (0.79). Overall, hybrid models of regression kriging (RK) and CubistRR (Cubist after addition of RR) had had largest agreement for predicting Ca (0.76) and Mg (0.81). This was also reflected in minimum number of calibration samples whereby geostatistical and statistical models needed n >_ 80 while hybrid models could still achieve moderate agreement with as low as n >_ 20. The MSPE showed that irrespective of DSM approach, the traditional soil map was not as accurate for predicting either Ca (2.64) and Mg (0.26). The results show, that while the soil condition in the study fields was beyond the range of the Six-Easy-Seps nutrient management guidelines, the potential of DSM to improve soil capability was demonstrated. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

The study utilized digital soil maps (DSM) to improve acidic soil conditions in Australian sugarcane fields and achieved good results. The findings showed that hybrid models had the best performance in predicting calcium and magnesium.