Proximally sensed digital data library to predict topsoil clay across multiple sugarcane fields of Australia: Applicability of local and universal support vector machine

Abstract

Knowledge about topsoil (0–0.3 m) clay is required to maintain sugarcane profitability in Queensland, Australia. However, laboratory analysis to get this knowledge is tedious and time consuming. To add value to limited clay data, a digital soil map (DSM) can be created by using digital data and mathematical models. At the field level, site-specific linear regression (LR) models are often used along with gamma-ray (γ-ray) spectrometry and electromagnetic induction data (i.e. soil apparent electrical conductivity – ECa). But these LR might not perform well in site-independent calibrations across multiple sites. In this regard, support vector machine (SVM) might be useful. In this research, we first aimed to determine, using a stepwise SVM and calibration dataset, the optimal digital data (i.e. individual or combined) to develop local (for each individual site) and universal (for combined sites) SVM models. Using optimal digital data, our second aim was to predict clay for validation datasets by using local SVM in site-specific approach and universal SVM in site-independent, holdout and spiking approaches. Using these approaches, DSM of predicted clay and associated uncertainty were generated for a representative study site (i.e., Mossman). The third aim was to determine the suitable number of spiking samples and by varying the size of both spiking set and calibration model. Approaches were compared using prediction agreement (Lin’s concordance) and accuracy (ratio of performance to deviation – RPD). We concluded from stepwise SVM that combining digital data resulted in better accuracy (RPD = 2.17) than individual γ-ray (1.79) or ECa (1.49) data. In terms of independent validation, the results of Mossman reflected the general rank order of different approaches with site-specific (3.03) excellent, spiking (1.89) very good, site-independent (1.84) good and holdout (1.18) poor predictions. In case of DSM and uncertainty maps, under-predictions were problematic at field edges and where digital data changed as a function of soil type and or proximity to a prior stream channel. Considering the suitable number of spiking samples, size of spiking set showed linear relationship with improvement in predictions while that of calibration model showed no influence. These results suggest the usefulness of proximally sensed data and universal SVM to effectively predict clay across multiple sites.