Evaluating Digital Agriculture Recommendations with Causal Inference
Even though process-based crop models are widely used to simulate crop growth, the challenge of parameter calibration makes it difficult to use them in practice. Also, it is labor-consuming to improve the model by adjusting the model description. To address these issues, we propose a knowledge-guided machine learning model (DeepOryza) to directly learn the crop growth pattern from data. A synthetic dataset generated by a process-based model (ORYZA2000) was used to pretrain the DeepOryza. An observation dataset was used to finetune and evaluate the DeepOryza. The results showed that DeepOryza can perform equally or better than the well-calibrated ORYZA2000. To investigate the effect of the proposed knowledge-guided structure, we designed two DeepOryza models with different structures. Results showed that the knowledge-guided structure can improve the performance of DeepOryza when the synthetic dataset was generated by the uncalibrated ORYZA2000. This finding indicates that the knowledge-guided structure could potentially reduce the calibration requirement of the process-based model.
I. Tsoumas, G. Giannarakis, V. Sitokonstantinou, A. Koukos, D. Loka, N. Bartsotas, C. Kontoes, I.N. Athanasiadis, Evaluating Digital Agriculture Recommendations with Causal Inference, Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, pg. 14514-14522, 2023, doi:10.1609/aaai.v37i12.26697.
You might also enjoy (View all publications)
- Crop modeling frameworks interoperability through bidirectional source code transformation
- A weakly supervised framework for high-resolution crop yield forecasts
- How do countries frame climate change? A global comparison of adaptation and mitigation in UNFCCC National Communications