Monitoring soil cadmium (Cd) at farm-scales (average 3 km 2 ) can potentially be rapid and cost-efficient by implementing proximal sensing techniques benefiting from a leveraged regional-scale (≥ 40,000 km 2 ) soil spectral library (RSSL). However, prediction models based on RSSL are often of limited use when applied at farm-scales because the coarseness of the RSSL. In this study, a New Zealand RSSL was used to assess the Cd concentration in a farm-scale sample set. For all samples, total Cd was determined, and visible-near-infrared (vis-NIR), mid-infrared (MIR), and portable X-ray fluorescence (pXRF) spectra were collected. A localisation technique to predict farm-scale Cd using RSSL spectral data was developed, based on spectral similarity or land use similarity relative to the farm-scale samples, and/or supplemented with selected farm-scale samples, as input for partial least squares regression and LOCAL algorithms. A model using MIR data from a RSSL pastoral samples subset ( n = 283) spiked with 12 extra weighted (×4) farm-scale samples as an input for a LOCAL algorithm, quantified Cd optimally (root mean square error = 0.22 mg Cd/kg; concordance correlation coefficient = 0.78; ratio of performance to interquartile distance = 1.93). Spiking the RSSL subset with farm-scale samples, including otherwise under-represented attributes such as soil order and Cd concentration range, improved the performance of models predicting farm-scale total Cd concentrations. A hybrid technique of localisation approach considered in this study may reduce compliance costs for Cd surveying and management, benefiting farmers. • A localisation technique, leveraging regional-scale SSL, quantified farm-scale Cd. • Localised models quantifying Cd based on MIR data outperformed others. • Similarity based deterministic search method and spiking optimised outputs. • PLS loadings captured the relevance of SSL subsets and spiking on localised models.
Shrestha et al. (Sun,) studied this question.