Remote sensing data and machine learning models estimate sorghum grain yield in a plant breeding program

Phenotyping remains a critical bottleneck in sorghum (Sorghum bicolor L. Moench) breeding programs, limiting rates of genetic gain due to labor-intensive yield estimation methods. To address this concern, this study investigates the potential of integrating remote sensing data with machine learning (ML) and deep learning (DL) models to improve sorghum grain yield predictions. Unmanned aircraft systems (UAS)-based imagery was collected across multiple field trials, extracting standard vegetation indices, canopy height features, and panicle traits using a YOLOv11-based object detection model, “YOLO-SORG.” Six ML models—including ridge regression (RR), elastic net (EN), LASSO regression (LR), support vector regression (SVR), random forest (RF), and XGBoost (XGB)—were trained to predict plot-level yield using three distinct feature sets: panicle traits, canopy traits, and a combination of both. Results indicate that models relying solely or partially on canopy-derived features provided the most consistent and accurate yield estimates (R² ≈ 0.74–0.76), whereas models relying solely on panicle traits performed poorly (R² ≈ 0.28–0.42), indicating nadir-derived panicle metrics were potentially being indirectly captured with the canopy traits. Traditional regression models outperformed tree-based ensemble methods in variance partitioning and repeatability (R ≈ 0.59-0.60), making them more suitable for many breeding applications. These findings highlight the promise of UAS-driven ML pipelines for non-destructive yield prediction but underscore potential limitations of nadir imagery for capturing panicle morphology and use in a robust yield prediction model. Future research should explore the inclusion of multi-temporal imaging, refined feature extraction approaches, and use of oblique, non-nadir imagery to enhance predictive accuracy in sorghum breeding programs.