Application of artificial neural network algorithms to estimate spatial soil organic carbon stock in Prairie lands from remote sensing and soil data pool

Estimating soil organic carbon stock (SOCS), the largest terrestrial carbon reservoir at the ecosystem level, is essential for identifying land management practices that support climate resilience. Remote sensing (RS) enables affordable analysis of SOCS, but lack of high-resolution RS data on biotic factors across ecosystems has limited its effectiveness. Data gaps in covariates accurately representing depth-wise soil parameters remain challenging too. Artificial intelligence plays a great role in expressing multifaceted relationships using non-linear complex models. Our objective was to model SOCS using a data-driven approach that linked RS products with soil variables through preprocessing, Random Forest modeling, and deep learning using four Artificial Neural Networks (ANN)-based modules. The best-performing ANN was chosen based on performance metrics and Uncertainty Analyses. Feature Importance analysis identified silt and pH of 0−5 cm soil-layer as the prime independent variables. Our approach effectively monitors carbon turnover and stabilization across both cropland and grassland ecosystems.