Abstract Detail

Crops and Wild Relatives

Majumder, Sambadi [1], Mason, Chase [1].

A Machine Learning approach to gain insight into crop sunflower cultivation under climate change and forecast future trends in the United States.

Crop sunflower (Helianthus annus) is an important oilseed crop in the U.S. alongside corn, soybean, and canola. The average yield in 2021 was 1530 pounds per acre, and total sunflower production was 36% lower compared to 2020, the lowest level since 1989. Adverse effects to global food production are expected to occur due to a changing climate, with the possibility of large-scale crop losses due to extreme weather conditions at local and regional scales. Accurate forecasting of crop yield and crop failure at a fine spatial scale would enable farmers and policy makers to evaluate current practices and facilitate future planning to ensure food security. The primary aim of this study is to use publicly available data in developing a model that can be used to reliably predict future crop yields and crop failure at a county-level across eight U.S. states under varying weather conditions. A descriptive modelling approach was used to identify which weather variables, soil conditions, and geographical factors were the most predictive of historical crop yield and crop failure. Feature selection approaches such as recursive feature elimination and the Boruta algorithm was used to identify the most influential factors. The two algorithms overlapped in which factors were deemed important. Longitude of counties, maximum and minimum monthly temperatures during the growing season as well as minimum temperatures outside of the growing season were deemed to be the most influential variables for crop yield. For crop failure, the most influential factors seemed to be maximum and minimum monthly temperatures during the growing season as well as for several months outside of the growing season. Random Forest (RF), Gradient Boosting Machines (GBM) and an artificial neural network was used to build predictive models and they were validated on historical data. The RF model outperformed all other predictive models when forecasting crop yield (root mean square error: 0.40) while the GBM model was the most superior for crop failure prediction (root mean square error: 0.70). The best performing models in relation to yield and crop failure were then used to forecast future predictions under four shared socio-economic pathways using future climate projections from the CNRM-ESM2-1 climate model as inputs along with geographical coordinates and soil variables. The forecasts were then spatially projected onto the map of the eight states of the study area. Results indicate that crop yield in several counties might decline quite significantly (by about 20 to 30 percent) between the years 2041-2060 while the crop failure rate is expected to rise within the same timeframe. The findings of this study might be used to inform public policy about maintaining yield preserving food security.

1 - University Of Central Florida, Department Of Biology, 4110 Libra Drive, Orlando, FL, 32816, United States

Machine Learning
climate change
Interpretable Machine Learning
Food Security
seed yield
crop failure.

Presentation Type: Oral Paper
Number: CWR1010
Abstract ID:392
Candidate for Awards:None

Copyright © 2000-2022, Botanical Society of America. All rights reserved