Unravelling drivers of field-scale digital mapping of topsoil organic carbon and its implications for nitrogen practices.

• Investigate the most statistically significant covariates to predict topsoil (0–0.3 m) organic carbon (SOC). • Assess proximal and remote sensors versatility either in combination or alone. • Examine geostatistical, machine learning and hybrid models. • Identify optimal calibration sample size. • Determine nitrogen (N) fertiliser practices considering mineralisation from SOC. Nitrogen (N) is essential for sugarcane growth, with one source mineralisation of soil organic carbon (SOC, %). To assist farmers determine suitable N fertiliser rates, the Six-Easy-Steps nutrient guidelines were devised by Sugar Research Australia and based on topsoil (0–0.3 m) SOC. Given the heterogeneous nature of SOC, a digital soil mapping (DSM) approach may be useful to map SOC at the field scale. This study aims to examine various models to predict SOC, including ordinary kriging (OK), regression kriging (RK), machine learning (ML) such as random forest (RF) and support vector machine (SVM), and hybrid models that combine ML with RK (i.e., RFRK and SVMRK). Various digital data sources, including proximal (i.e., gamma-ray [γ-ray] and electromagnetic [EM] induction data) and remote (i.e., land surface temperature [LST, oC]) sensing data were explored in combination and alone. Minimum calibration sample size (i.e., 110–10) was also investigated. The comparisons were validated using a hold-out dataset, with prediction agreement (Lin's concordance correlation coefficient [LCCC]) and accuracy (ratio of performance to deviation [RPD]) assessed. The most statistically significant digital data used in combination included γ-ray (i.e., total count [TC]), EM (i.e., 1mHcon and 2mHcon), remote (i.e., LST) and northing. The SVMRK model performed best when using all calibration data, producing moderate agreement (LCCC = 0.73) and fair (RPD = 1.65) accuracy. This was also the case when sample size was >= 50, but when it was < 50, RFRF returned equivalent agreement (0.70) and accuracy (1.55). Similar results were achieved when only γ-ray (i.e., TC) was used, however, RFRK produced moderate agreement and fair accuracy using as few as 20 and as many as 70 samples, with the optimal being 40 (i.e., LCCC = 0.79 and RPD = 1.7). By comparison, OK and RK generally produced poor agreement (<= 0.65) and accuracy (<= 1.4) regardless of sample size. The final DSM using RFRK provided an indication of soil condition that limits soil capability to store N as a function of SOC and acts a means to replenish soil capital with N application rates according to the Six-Easy-Steps. Specifically, Tenosols and Rudosols require 130 kg/ha to 120 kg/ha, whereas Kandosols require 110 and 100 kg/ha. The DSM equates to a potential ∼ 22 % decrease in the cost of N fertiliser application, relative to a one-size fits all approach. [ABSTRACT FROM AUTHOR]