Canola yield prediction and genotype selection using uncrewed aerial system multispectral imagery and machine learning models

Evaluation and selection of canola breeding lines across multiple years and locations are vital for variety development. However, accurately measuring canola ( Brassica napus L.) yield under field conditions is challenging due to end‐of‐season events such as rain, windstorms, and hail, as well as variability in plant maturity and pod shattering. This study evaluated the potential of using uncrewed aerial systems (UAS)‐based multispectral imaging to enhance yield prediction and genotype selection in canola. Field trials were conducted across four environments using diverse genotypes in randomized complete block and augmented designs. Aerial images were collected using a UAS outfitted with 10‐band multispectral sensor. Most of the spectral band‐derived vegetation indices showed positive correlations (0.42–0.74) with yield, particularly during the end‐of‐flowering to near‐maturity stages (65–76 days after seeding). Model‐based feature selection selected nine features (ground cover, Green Normalized Difference Vegetation Index, Soil‐Adjusted Vegetation Index, Enhanced Normalized Difference Vegetation Index 2, Visible Atmospherically Resistant Index 2, Green–Red Ratio Index 2, Visible Difference Vegetation Index, Vegetation Index, and Green Chlorophyll Index) as yield predictors at this critical stage. Principal component analysis and K‐means clustering effectively analyzes summarized spectral diversity. Broad‐sense heritability of selected indices ranged from 19% to 74%, indicating genetic potential for improvement. Yield prediction across multiple validation scenarios showed that the ENET, least absolute shrinkage and selection operator, partial least squares regression, and Ridge regression models outperformed Multiple linear regression and random forest achieve strong predictive performance with coefficient of determination ranging between 0.51 and 0.76 and root mean squared error ranging from 188 to 461 kg ha −1 . In the validation test, vegetation index‐based selection shared 70%–88% of the selected genotypes across multiple validation scenarios when compared to the selection based on harvested yield, supporting its effectiveness. Future research should consider stage‐specific feature selection, starting from the early stage, to improve prediction accuracy. This UAS multispectral imagery‐based yield prediction and selection supports mitigating phenotyping challenges, thereby improving selection and genetic gain in canola breeding.