Predicting ecosystem responses by data-driven reciprocal modelling

Area of Science:

• Environmental Science
• Agricultural Science
• Data Science

Background:

• Traditional experiments quantify treatment effects but often lack real-world representativeness.
• Field conditions present challenges for isolating individual environmental treatment impacts.
• Accurate quantification of environmental treatments is crucial for effective land management.

Purpose of the Study:

• To present a novel data-driven reciprocal modeling framework for quantifying environmental treatment effects under field conditions.
• To enable the estimation of individual treatment effects using representative survey data.
• To apply the framework to assess land-use change impacts on soil organic carbon stocks.

Main Methods:

• Utilized a representative survey dataset including treatment (A or B), target variable, and environmental properties.
• Trained a machine learning model to predict the target variable using observations from one treatment group (e.g., group A).
• Applied the trained model to the other treatment group (e.g., group B) within its applicability space, using residuals for effect estimation.

Main Results:

• The data-driven reciprocal modeling framework successfully quantifies individual treatment effects in field settings.
• Residuals from the applied model serve as case-specific effect size estimates.
• Demonstrated the framework's utility in estimating spatially explicit effects of land-use change on European agricultural soil organic carbon.

Conclusions:

• The proposed framework offers a more representative alternative to controlled experiments for estimating environmental treatment effects.
• This data-driven approach provides accurate and spatially explicit effect size estimates.
• The methodology is applicable to a wide range of environmental treatments and conditions.