A comparison of country-scale subsoil predictions between a numeric and a taxonomic soil classification system

Traditional soil classification systems are designed to communicate information; however, surveyor biases and tacit knowledge can lead to subjective soil class designations. Consequently, different soil scientists may classify the same soil differently. This becomes a critical issue when mapping soil classes, as there could be multiple interpretations for the same observation. To address this problem, numerical soil classification systems have been developed. However, little is known about how well they compare to taxonomic systems when spatially predicted on a national scale. This study aimed to compare a previously developed, unsupervised numeric classification system and South Africa's taxonomic soil classification system in terms of their spatial predictions across the country. The taxonomic system of South Africa has 19 defined subsoil horizons, which were aggregated into eight horizons and compared to a nine horizon numeric classification as well as South Africa's profile (soil form) classification comprising of 73 different soil groupings, which was used as a control. The comparison was conducted from predictions through gradient tree boosting in Google Earth Engine at a 30 m resolution. The numerical system (kappa = 0.30, accuracy = 0.57) exhibited poor spatial predictions, with a kappa 22% lower and accuracy 2% lower than the control (kappa = 0.52, accuracy = 59%). On the other hand, the taxonomic system performed well, with a kappa of 0.57 and an accuracy of 67%, exhibiting a 5% increase in kappa and an 8% increase in accuracy compared to the control. It was hypothesized that the overpredictions of the predominant horizon contributed to the numeric system's poor performance. Nevertheless, both systems showed the highest maximum entropy in arid regions of the Karoo and savannah biomes, albeit in spatially distinct ecoregions. It was thought that the divergence in the two systems' maximum entropy was due to their association with precipitation differences (amount and seasonality) as well as vegetation type and cover (woodlands vs. shrublands). To map the country in more detail, further soil sampling should be conducted in arid regions and optimisation of the predictive algorithm for each soil category should be performed.