Machine Learning predictions of cropland ecosystem respiration

Abstract:

Accurate estimation of cropland ecosystem respiration is essential to predicting variations of Soil Organic Carbon (SOC), which not only exerts first-order control on crop yields, but serves as one of the most important pools in the global carbon cycle. Recent advances in process-based models have provided us a unique means to reproduce ecosystem respiration rate (R_eco, consisting both autotrophic and heterotrophic respiration) obtained by Remote Sensing (RS) techniques, and a powerful tool to interpret future SOC changes. However, the computational costs for such simulations are extremely expensive. This high cost has prevented the application of those process-based models, and thus limited our prediction capability. In this work, we build a surrogate model for a novel process-based model – ECOSYS using a state-of-the-art artificial Neural Network (NN) framework, aiming at reducing the runtime of each simulation while preserving its accuracy. This NN framework uses RS observations as inputs (e.g. GPP, LAI, air temperature, humidity, etc.), to reproduce R_eco at field-scale.

Results from the NN framework show close agreements compared to the original ECOSYS simulations (R² higher than 0.99), while reducing the runtime by over three orders of magnitude. After calibrating and benchmarking the simulated R_eco with data collected from three Eddy Covariance towers, we apply the surrogate NN framework to every crop field in the United State corn belt. This application provides the first high-resolution R_eco map for corn fields, establishing a foundation for future climate-change predictions and decision making.