Learning vs. understanding: When does artificial intelligence outperform process-based modeling in soil organic...

Area of Science:

• Ecological modeling
• Soil science
• Machine learning applications

Background:

• Machine learning (ML) algorithms are increasingly used for ecological modeling.
• Limited evaluation exists for ML in predicting soil organic carbon (SOC) using small datasets typical of long-term studies.
• ML performance for SOC prediction has not been compared against traditional process-based models.

Purpose of the Study:

• To compare the performance of ML algorithms, process-based models, and ensembles for SOC prediction.
• To evaluate model performance using data from five long-term experimental sites in Austria.
• To assess the impact of dataset size and cross-validation strategies on ML performance.

Main Methods:

• Comparison of ML algorithms (Random Forest, Support Vector Machines with polynomial kernel), calibrated and uncalibrated process-based models, and ensembles.
• Utilized data from 256 independent data points across five long-term experimental sites in Austria.
• Applied leave-one-site-out cross-validation and reduced training sample sizes to test model robustness.

Main Results:

• ML approaches (Random Forest, SVM) outperformed process-based models when using all available data.
• ML algorithm performance decreased significantly with reduced training data or leave-one-site-out cross-validation.
• Process-based model accuracy was highly dependent on proper calibration and model combinations.

Conclusions:

• ML models are superior for SOC prediction when large datasets are available.
• Process-based models are more suitable for exploring underlying biophysical and biochemical mechanisms of SOC dynamics.
• Ensembles combining ML algorithms and process-based models are recommended to leverage the strengths of both approaches.