Knowledge-guided assimilation to bridge the gap between sensing and modeling with indirect labels for global-scale carbon monitoring

Shuo Xu, Zhihao Wang, Ruichen Wang, Xiaowei Jia, Lei Ma, George C. Hurtt, Yiqun Xie

Keywords: Knowledge-guided Learning, Indirect Labels, Sensing, Assimilation

Abstract: Advanced air-borne and in-situ sensing platforms have generated invaluable observations of the Earth systems and offer exciting opportunities in enhancing the monitoring and forecasting capabilities to tackle challenges such as global warming. While process-based models have been developed for decades, they have limited ability to incorporate real-world observations to further enhance the prediction ability, especially to correct simplified sub-processes that tend to cause deviations from the observations. In particular, many process-based models rely on itemized lower-level processes, whereas the sensors very often can only collect aggregated mixed-up information, constraining the use of these observations to improve the modeling. Existing works on knowledge-guided learning mainly focus on connecting process-based and data-driven methods via directly matched variables, using physical rules and simulations to constrain the training process. We propose a knowledge-guided assimilation approach to integrate process-based and learning models to improve the utilization of large-scale simulations with aggregated indirect observations. To evaluate approach, we carry out a global-scale case study with ecosystem models that are widely used in carbon monitoring. The results on global-scale benchmark data show that knowledge-guided integration of indirect labels can significantly enhance prediction skills compared to existing learning methods.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 8105