Predictive Modeling of US Corn Fields – A Machine Learning Approach

To plan out the sourcing part in a food supply chain, the problem is two-fold. One is logistics optimization, which is well-recognized by food vendors. The other, however, is very hard to predict and map into planning: the... [ view full abstract ]

To plan out the sourcing part in a food supply chain, the problem is two-fold. One is logistics optimization, which is well-recognized by food vendors. The other, however, is very hard to predict and map into planning: the actual agricultural production. While process-based agriculture models, like DNDC (De-Nitrification-De-Composition), are capable of simulating the daily activity of cropping systems, the embedded complexity of calculating ecological activities significantly slows down model performance. Currently those models are either used for coarse level (county or state) analysis, or a very limited number of cropping fields, whereas large food companies need to collect site-specific information for a broad landscape (national).

Model performance improvement is therefore necessary for such project. Our goal is to use DNDC to project activities of corn fields in United States on a fine scale. Specifically, we adopt Random Forest, a supervised learning strategy to retrieve DNDC results. We first test the methodology in representative states. Results showed that Random Forest is capable of producing corn yield and GWP results with 1000 times faster speed than DNDC. Total fertilization rate, clay fraction and irrigation rate are the three attributes that have the most impact on corn growth and emissions in general, but there are cross-state difference due to farming strategy change. We are going to apply the strategy to the year of 2099 using down-scaled daily weather data. Impact of different farming practices will also be tested, for example altering the amount/types of fertilizers applied to corn field.

By the end of this project, we will be able to spot areas more sensitive to climate change and certain agricultural management scenarios, based on which we can offer management suggestions to local farmers, deliver a message on agricultural policy making, and provide a platform for large food vendors on business plannin.