Projecting long-term soil organic dynamics in Australia's rangelands

Rangelands in Australia are vast and occupy more than 80% of the continental land area. They extend across arid, semi-arid, and the tropical regions with seasonal, variable rainfall in the north. They include diverse, relatively undisturbed grasslands, shrublands, woodlands and tropical savanna ecosystems. They represent Australia’s largest terrestrial carbon sink as they account for almost 70% of Australia's total soil organic carbon stock (Viscarra Rossel et al., 2014), more than all above-ground sources of carbon (native grasses, trees and shrubs) in these regions (Gifford et al., 1992). Here we have developed a novel space-time approach for projecting the long-term C dynamics of rangelands soils using Long Short-Term Memory (LSTM) deep learning neural networks. We further demonstrate how the networks might be interpreted and quantified the influence of explanatory variables on the spatiotemporal dynamics of soil C in these regions. Our results provide an improved ability to accurately model long-term carbon dynamics, which is needed to confidently predict changes in soil C from change in climate or anthropogenic disturbance. The information is critical for improving our understanding of soil C in these regions and for understanding the potential for sequestering C in the rangelands.