Deep Learning for Sorghum Yield Forecasting using Uncrewed Aerial Systems and Lab-Derived Imagery

Authors: BARI, MD ABDULLAH - Kansas State University; BAKSHI, ALIVA - Kansas State University; PRAMANIK, SWARAJ - Kansas State University; CHOTON, JAHID - Kansas State University; WITT, TREVOR - Kansas State University; CAREGEA, DOINA - Kansas State University; Bean, Scott; JAGADISH, KRISHNA - Texas Tech University; FELDERHOFF, TERRY - Kansas State University

Submitted to: Plant Phenomics
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/20/2025
Publication Date: 2/14/2026
Citation: Bari, M., Bakshi, A., Pramanik, S., Choton, J., Witt, T., Caregea, D., Bean, S.R., Jagadish, K., Felderhoff, T. 2026. Deep Learning for Sorghum Yield Forecasting using Uncrewed Aerial Systems and Lab-Derived Imagery. Plant Phenomics. Volume 8, Issue 1, March 2026, 100133. https://doi.org/10.1016/j.plaphe.2025.100133.
DOI: https://doi.org/10.1016/j.plaphe.2025.100133

Interpretive Summary: Sorghum is an important grain crop in the central plains of the United States and is used for feed, fuel, and food. Sorghum has a wide degree of genetic and phenotypic diversity which can be exploited to improve the agronomic performance and end-use quality and value of the crop. Grain yield is a primary trait that the sorghum breeding industry is working on improving as increased yields directly relate to the value and productivity. Therefore, to take advantage of the genetic diversity of sorghum and develop new lines with improved yield, methods for rapidly determining and predicting yield are necessary. This research evaluated the use of deep learning algorithms to predict yield from images of sorghum and found that yield could be forecast using deep learning processing of images. This research will benefit sorghum breeders and seed industry by facilitating higher throughput measurements of grain yield in breeding programs leading to the development of sorghum germplasm with improved agronomic performance and value.

Technical Abstract: With the current Artificial Intelligence revolution, advanced graphical processing units, and open-source platforms for machine learning, and deep learning algorithms offer new opportunities for rapid, accurate phenotypic feature extractions using Uncrewed Aerial Systems derived field and manually obtained lab imagery, leading to sorghum yield forecasting. Yield forecasting analytics are critical for breeding and research programs to efficiently evaluate genetics and breeding materials in order to enhance cultivar development efficiently. The field trial, including 36 sorghum genotypes, was commenced following an RCBD with three replicates at Ashland Bottoms, Kansas, in 2023 to extract yield attributes from images and integrate them into machine learning predictive models to forecast sorghum yield. The field images were captured at 6 meters Above Ground Level from the nadir and oblique angles (45°) using a DJI M300, and in the laboratory setup, we manually collected and imaged four panicles. This research deployed two state-of-the-art deep learning based computer vision models named YOLO and Detectron2 to harness UAS field images augmented with manually collected laboratory images. YOLO outperformed Detectron2 in detecting sorghum panicle heads, achieving a higher accuracy with a mAP@0.50 ranging from 0.92 to 0.98, compared to Detectron2’s range of 0.61 to 0.89 across training datasets. By training appropriate models, we detected and counted field panicles, determined panicle size dimensions, counted seeds, and extracted seed area with correlation coefficients of 0.86, 0.98, 0.94, and 0.25 with the ground truth observations, respectively. For the predictive models, Support Vector Regression, Decision Tree Regression, and Random Forest Regression were used to predict yield with a correlation coefficient of 0.65, 0.60, and 0.97 with actual yield. These algorithms are poised to extract yield features from images to integrate into robust predictive analytics to address the evolving need of digitizing plant phenotyping.