Artificial Intelligence and Satellite Based Remote Sensing to Predict Soybean (Glycine max) Yield

Authors: JOSHI, DEEPAK - South Dakota State University; CLAY, SHARON - South Dakota State University; SHARMA, PRAKRITI - South Dakota State University; MORADI, HOSSEIN - South Dakota State University; Kharel, Tulsi; THAPA, RESHAM - Tennessee State University; CLAY, DAVID - South Dakota State University

Submitted to: Agronomy Journal
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/25/2023
Publication Date: 10/12/2023
Citation: Joshi, D.R., Clay, S.A., Sharma, P., Moradi, H., Kharel, T.P., Thapa, R., Clay, D.E. 2023. Combining Artificial Intelligence and Satellite Based Remote Sensing to Predict Soybean (Glycine max) Yield. Agronomy Journal. 2023;1–14. http://dx.doi.org/10.1002/agj2.21473.
DOI: https://doi.org/10.1002/agj2.21473

Interpretive Summary: In the past, farmers and agronomists estimated yields based on counting the number of plants per hectare, pods per plant, and seeds per pod. This information, while useful, was highly variable, expensive, and may not be suitable for precision agriculture. Scientists from USDA-ARS, Crop Production Systems Research Unit, Stoneville, MS; South Dakota State University, Brookings, SD; and Tennessee State University, Nashville, TN explored the possibility of artificial intelligence (AI) techniques based on remote sensing to estimate yield . In 2019 and 2021, soybean yields were measured in 3 large production fields (>39 ha) located at 3 different counties in South Dakota. Combine harvester equipped with a yield monitor system was used to collect yield data from all three fields. Multispectral (Blue, Green, Red, NIR with 3.1m spatial resolution) space-based remote sensing (PlanetScope satellite) data was collected six times from emergence/cotyledon leaf (VE/VC) to full maturity (R7) growth stage. Yields and remote sensing data were aggregated into 10 by 10 m grid cells. Five artificial intelligence (AI) models [deep neural network (DNN), random forest (RF), support vector machine (SVM), LASSO, and ADABOOST] techniques were used to predict soybean yields using remote sensing data. Generally, the DNN model outperformed the other models, and as the crop matured, from VE/VC to R4/R5, the percent of yield variability explained by the model improved from less than 50% to over 60%. These findings indicate that AI techniques can be used to predict soybean yields at all growth stages, however estimates improve as the plant matures. Result from this study will be helpful to grower for numerous purposes including pest management decisions, crop marketing, price forecasting, insurance, and harvest plans.

Technical Abstract: The manual counting of a soybean (Glycine max) plants, pods, and seeds/pod is the classical approach to estimate yields. Since this method is labor-intensive, expensive, and highly variable, alternative methods that employs remote sensing and advanced data analytics are desired. Therefore, this study determined if artificial intelligence (AI) techniques, combined with remote sensing data at multiple growth stages could be used to capture within field spatial variability and predict yield in commercial soybean fields. For this we monitored soybean yields in three large production fields (> 39 hectares) using a combine harvester equipped with a yield monitor system during 2019 and 2021 growing seasons. Additionally, we calculated vegetation indices using multispectral (Blue, Green, Red, NIR) space-based remote sensing (PlanetScope satellite with 3.1 m spatial resolution) data six times during the soybean growing season. For this satellite images were acquired to encompass various growth stages from emergence (VE) to full maturity (R7): VE/VC, V1/V3, R1/R2, R2/R3, R4/R5 and R6/R7. Remote sensing and soybean yield monitor data were aggregated into 10 by 10 m grid cells. Utilizing remote sensing data, five AI models [deep neural network (DNN), random forest (RF), support vector machine (SVM), Least Absolute Shrinkage and Selection Operator (LASSO), and AdaBoost] techniques were used to predict soybean yields. We found that the DNN outperformed all other AI models under investigation. Moreover, as crop matured from VE/VC to R4/R5, the ability of the model to predict yield variability improved from less than 50% to over 60%. These findings indicate that remote sensing capabilities combined with advanced AI techniques can be used to predict soybean yields, and yield estimates improve as the crop matures.