Leveraging remotely sensed data and machine learning for quantifying soil organic carbon stocks in woody-encroached areas of Bisley Nature Reserve.

Abstract

Woody encroachment has emerged as a significant driver of land cover change in grasslands, with profound effects on Soil Organic Carbon (SOC). SOC is an important indicator of soil fertility, and thus crucial for grassland productivity. Previously, woody encroachment has been reported as a primary source of SOC alteration in grasslands. However, there are still debates and uncertainties on whether this phenomenon amplifies or reduces SOC sequestration. Therefore, it is necessary to further evaluate SOC accumulation in grasslands affected by proliferation of woody plants. Remote sensing offers freely available and cost-effective data with improved spatial and spectral resolution to quantify SOC. In this regard, the current study aimed to evaluate the role of remote sensing in quantifying the spatial variability of SOC across a woody encroached Bisley Nature Reserve. The first objective focused on quantifying the spatial variability of SOC stocks in both pristine and woody-encroached grasslands using PlanetScope spectral bands and vegetation indices. At a depth of 0-30 cm, the study found that landscapes dominated by woody encroachment exhibited higher SOC values compared to pristine grasslands. Using Deep Neural Networks, a combination of PlanetScope spectral bands and vegetation indices model achieved acceptable accuracy (R² = 0.64) for quantifying SOC stocks at this depth. Interestingly, NDVI was the most important variable for estimating SOC within woody encroached grassland. However, to fully understand the dynamics of SOC accumulation and its vertical distribution across different soil depths, it was necessary to expand the analysis. Hence the second objective extended the investigation by utilizing a Random Forest algorithm and integrating additional remotely sensed data to model SOC stocks at multiple soil depths (0-30 cm, 30-60 cm, and 60-100 cm). This approach provided a more comprehensive view of SOC variability, revealing a higher concentration of SOC in the top 30 cm compared to deeper layers. By incorporating topographic variables, Synthetic Aperture Radar (SAR), Sentinel-2, and PlanetScope data, the model produced higher accuracy for deeper soil layers, with R² values of 0.76 at 60 cm and 0.79 at 100 cm. SAR data enhanced the model by offering insights into subsurface conditions. These findings underscore the necessity of investigating SOC at different depths to fully capture its spatial distribution and highlight the potential of remote sensing and machine learning to improve SOC mapping accuracy across woody-encroached grasslands.