Digital soil mapping of coarse fragments in southwest Australia: Targeting simple features yields detailed maps

The spatial distribution of soil coarse fragments (CF) is important for a variety of agricultural and environmental applications because it directly impacts soil processes including hydrology and nutrient cycling. However, there is often insufficient measured data to reliably model and map CF using a digital soil mapping approach. By targeting CF layer occurrence rather than CF as a continuous soil property, we increased the number of data sites available for spatial modelling which improved predicted patterns. We define CF layers as hard CF and segregations size > 2 mm, > 20% by volume, and > 10 cm thick, the definition of a ferric diagnostic horizon in the Australian Soil Classification system. Highly variable legacy data yielded nearly 40,000 georeferenced sites over the 1 M square kilometre study area. The binary classification models evaluated were random forest using regression trees (probability machines) and classification trees, with and without class balancing. The best performing models were regression forests, followed by classification using the threshold that maximised the Kappa coefficient. Prediction accuracy was determined by validating with a subset of legacy data randomly selected on an unaligned grid and withheld from model training; validation was performed on all modelling methods, plus other available CF digital soil maps. Incorporating low quality legacy observations improved CF predictions significantly. The final maps depict CF layer presence or absence in four depth slices (0-5, 5-15, 15-30, and 30-80 cm) and anywhere within the top 80 cm, for both CF of any composition and specifically for ironstone gravel (sesquioxide nodules). These new maps have high predictive power: ironstone gravel layers had AUC ranging from 0.86 to 0.89, Kappa between 0.48 and 0.52, and overall accuracy from 0.82 to 0.92; for CF layers of mixed composition AUC ranged from 0.79 to 0.84, with Kappa 0.43 to 0.46, and overall accuracy from 0.74 to 0.88. The maps are plausible representations of local variation in soil properties across the landscape. They reflect the expert knowledge encoded in conventional soil maps, and are more locally credible than other available modelled CF maps. Modelling CF as layers rather than continuous properties led to high accuracy spatial representation of simple but still useful soil features for our study area. These soil feature maps complement the quantitative soil property surfaces common to digital soil mapping studies that are frequently constrained by data availability.

Automated summary

The study focused on spatial distribution of soil coarse fragments and utilized a new approach to increase accuracy of prediction maps, providing high predictive power for various soil features within the study area.