Deep Learning for Improved Agricultural Risk Management
- Nathaniel Newlands, Science and Technology Branch, Agriculture and Agri-Food Canada, Government of Canada, Summerland, British Columbia, Canada
- Azar Ghahari, Mathematical Sciences, University of Texas at Dallas, Dallas, Texas, United States
- Yulia R. Gel, Mathematical Sciences, University of Texas at Dallas, Dallas, Texas, United States
- Vyacheslav Lyubchich, Chesapeake Biological Laboratory, University of Maryland Center for Environmental Science, Solomons, Maryland, United States
- Tahir Mahdi, Actuarial and Forecasting Unit (AFU-SAP), Agriculture and Agri-Food Canada, Ottawa, Ontario, Canada
AbstractDeep learning provides many benefits, including automation, speed, accuracy, and intelligence, and it is delivering competitive performance now across a wide range of real-world operational applications - from credit card fraud detection to recommender systems and customer segmentation. Its potential in actuarial sciences and agricultural insurance/risk management, however, remains largely untapped. In this pilot study, we investigate deep learning in predicting agricultural yield in time and space under weather/climate uncertainty. We evaluate the predictive power of deep learning, benchmarking its performance against more conventional approaches alongside both weather station and climate. Our findings reveal that deep learning offers the highest predictive accuracy, outperforming all the other approaches. We infer that it also has great potential to reduce underwriting inefficiencies and insurance coverage costs associated with using more imprecise yield-based metrics of real risk exposure. Future work aims to further evaluate its performance, from municipal area-yield, to finer-scale crop-specific producer-scale yield.
Return to previous page