Machine learning produces higher prediction accuracy than the Jarvis-type model of climatic control on stomatal conductance in a dryland wheat agro-ecosystem

• Published in: Agricultural and forest meteorology Vol. 304-305; p. 108423
• Main Authors: Houshmandfar, Alireza, O'Leary, Garry, Fitzgerald, Glenn J, Chen, Yang, Tausz-Posch, Sabine, Benke, Kurt, Uddin, Shihab, Tausz, Michael
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15-07-2021
• Subjects: Random forest; Support victor machine; Transpiration; Water cycle; Transpiration; Random forest; Support victor machine; Water cycle
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2021.108423

•Machine learning (ML) produces high prediction accuracy of gs in wheat.•The ML models are especially important for interpolating datasets with good accuracy.•The ML models require large datasets for training to achieve statistical significance.
We compared Support Vector Machine (SVM) and Random Forest (RF) machine learning approaches with the widely used Jarvis-type phenomenological model for predicting stomatal conductance (gs) in wheat (Triticum aestivum L.) using historical measurements collected in the Australian Grains Free-Air CO2 Enrichment (AGFACE) facility. The machine learning-based methods produced greater accuracy than the Jarvis-type model in predicting gs from leaf age, atmospheric [CO2], photosynthetically active radiation, vapour pressure deficit, temperature, time of day, and soil water availability (i.e. phenological and environmental variables determining gs). The R2 was 0.76 for the Jarvis-type but 0.92 for SVM and 0.97 for RF machine learning-based models, with a calculated RMSE of 0.292 mol m−2 s−1 in the Jarvis-type compared to 0.129 mol m−2 s−1 in SVM and 0.081 mol m−2 s−1 in RF. The machine learning models, however, needed large datasets for training to achieve statistical significance, and do not offer the same opportunity to provide physiological insights through a statistically testable hypothesis. These results show that using the machine-learning based methods can achieve high prediction accuracy of gs that is especially important when incorporated into larger models, but their ability to extrapolate beyond observed data ranges will need to be assessed before they could be considered in place of the physical model.