Statistical and machine learning models for early cotton yield prediction under resource gradient

Early cotton yield prediction is critical for optimizing the productivity of smallholder farmers. UAV-based yield modeling faces challenges due to limited ground-truth data and spectral multicollinearity. Although data-intensive machine learning shows promise, smallholder systems require models optimized for limited datasets and multicollinearity. This study assessed agronomic data-only, remote sensing data-only, and hybrid data modeling approaches to define an optimized framework that integrates remote sensing imagery and non-destructive field measurements. The approach compares multiple linear regression, Stepwise selection, LASSO, Elastic Net, and Support Vector Machines (SVMs) to evaluate predictive performance and identify the earliest growth stage for yield prediction. Multispectral UAV imagery and extensive agronomic data were collected at six critical phenological stages across two experimental plots with varying water and nitrogen fertilization treatments. The hybrid models outperformed the single-data-source models, with LASSO regression achieving the highest precision (R 2 = 0.80) by combining multispectral data with plant height at 90 DAS. The phases between 50% flowering and boll growth were identified as optimal windows for accurate yield estimation using remote sensing observations. The results highlight the strengths of regularized regression in limited-data situations by addressing the overfitting caused by multicollinearity. Our hybrid approach enables yield prediction for small farms by combining UAV imagery and field measurements with regularized machine learning, mitigating multicollinearity, and requiring minimal ground data. This study suggests that yield may be predicted several months before harvest using remote sensing-only models at 90–115 DAS, providing a possible route for scaled precision agriculture in data-poor regions.