Deep Learning and Machine Learning Frameworks for Sorghum Yield Forecasting Using UAV and Laboratory Imagery

June 11, 2026, 9:30am
Md Abdullah Al Bari, Department of Agronomy, Kansas State University, and Department of Plant and Environmental Sciences, New Mexico State University
Department of Computer Science, Kansas State University
Trevor D. Witt, Department of Entomology, Kansas State University
Scott Bean, Center for Grain and Animal Health Research, ARS-USDA, Manhattan, KS
S.V. Krishna Jagadish, Department of Plant and Soil Science, Texas Tech University
Terry Felderhoff*, Department of Agronomy, Kansas State University

Recent advances in artificial intelligence (AI), graphics processing units (GPUs), and open-source computing platforms have accelerated the application of machine learning (ML) and deep learning (DL) approaches in agricultural research. These technologies enable rapid extraction of phenotypic information from imagery and offer new opportunities for digital phenotyping and yield forecasting in plant breeding programs. Accurate yield prediction can improve selection efficiency and accelerate cultivar development.

A field experiment consisting of 36 diverse sorghum genotypes was conducted in 2023 at Ashland Bottoms, Kansas, using a three-replicated randomized complete block design (RCBD). Field images were collected using a DJI Matrice 300 UAV at 6 m above ground level from both nadir (90°) and oblique (45°) viewing angles. Deep learning models, including YOLO and Faster R-CNN (Detectron2), were trained to extract yield-related traits from UAV field imagery and laboratory-acquired panicle images. YOLO consistently outperformed Faster R-CNN for panicle detection, achieving mAP@0.50 values ranging from 0.92 to 0.98 compared with 0.61 to 0.89 for Faster R-CNN. Predicted field panicle counts were strongly associated with manually measured counts (r = 0.86). Analysis of laboratory images enabled estimation of panicle area, seed number, and seed area, with correlations of 0.79, 0.94, and 0.25, respectively, relative to ground-truth measurements. Image-derived phenotypic features were subsequently integrated into machine learning models for grain yield prediction. Random Forest Regression (RFR), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost) achieved prediction accuracies (r) of 0.84, 0.88, and 0.90, respectively, with XGBoost providing the highest predictive performance. These results demonstrate that YOLO-based deep learning models can effectively extract biologically relevant and yield-predictive traits from UAV and laboratory imagery. The integration of deep learning-derived phenotypes with machine learning predictive analytics offers a scalable framework for digital phenotyping, yield forecasting, and data-driven selection in sorghum breeding programs, with the potential to enhance selection efficiency and accelerate genetic gain.