Proper management of soil, including its organic carbon component, can mitigate shortages in food, water, and energy. Measuring and monitoring soil organic carbon is therefore critical to enhancing our ability to make farming practices sustainable. Changes in soil organic carbon are not only dictated by weather events and management practices, but also by other soil processes such as plant nutrient uptake, soil organisms, soil texture, micro-nutrient content, and soil disturbance. Due to the complex nature of soil science, the exact relations among all soil processes are not yet known. Further, there is no accepted universal method for studying soil organic carbon and the relations among soil processes. Moreover, current models of soil organic carbon (RothC-26.3, Century and DNDC) are not consistent with the latest advances in understanding of soil processes and do not generalize to different soil conditions found globally as they are limited by their modeling assumptions. Our goal is to create a method that can generalize well across regions, soil types and soil management practices that can aid in standardization efforts for soil organic carbon measuring, reporting and verification.  Although ML methods have demonstrated an improved ability to predict soil organic carbon, the reliance of conventional ML on the i.i.d. assumption that training data represents the deployed environment limits their out-of-distribution generalizability. Using causal discovery methods is a way to overcome this limitation as these frameworks help explicitly model cause-and-effect relations among the soil processes and management practices to improve the out-of-distribution generalizability. In this work, we present an approach based on causal graphs to estimate soil organic carbon stocks. We conclude that causal approaches produce better results for soil organic carbon estimation on unseen fields (from different locations, with different soil properties, farm management practices and land use) as compared to non-causal approaches.

Using causal machine learning to model soil processes improves generalization

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.