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

Area of Science:

• Agricultural Science
• Plant Breeding
• Remote Sensing

Background:

• Phenotyping is a major limitation in sorghum breeding, slowing genetic gain due to laborious yield measurements.
• Integrating remote sensing and machine learning (ML) offers a potential solution for efficient sorghum yield prediction.

Purpose of the Study:

• To evaluate the effectiveness of integrating unmanned aircraft systems (UAS)-based remote sensing data with ML models for predicting sorghum grain yield.
• To compare the performance of various ML models using different feature sets derived from UAS imagery.

Main Methods:

• UAS-based imagery was collected across multiple sorghum field trials.
• A YOLOv11 object detection model ('YOLO-SORG') extracted vegetation indices, canopy height, and panicle traits.
• Six ML models (RR, EN, LR, SVR, RF, XGB) were trained to predict plot-level yield using panicle traits, canopy traits, or both.

Main Results:

• Models utilizing canopy-derived features (R² ≈ 0.74-0.76) consistently outperformed those using only panicle traits (R² ≈ 0.28-0.42).
• Traditional regression models showed better variance partitioning and repeatability (R ≈ 0.59-0.60) compared to tree-based ensembles.
• Nadir imagery may indirectly capture panicle traits through canopy measurements, limiting the direct utility of panicle-specific metrics.

Conclusions:

• UAS-driven ML pipelines show significant promise for non-destructive sorghum yield prediction.
• Canopy traits are more reliable predictors of yield than nadir-derived panicle traits.
• Future research should incorporate multi-temporal and oblique imagery to improve predictive accuracy and panicle trait capture.