Multi-modal feature integration from UAV-RGB imagery for high-precision cotton phenotyping: A paradigm shift toward cost-effective agricultural remote sensing

Abstract

Cost-effective remote sensing solutions are critically needed to democratize precision agriculture technologies. While hyperspectral and LiDAR systems deliver high accuracy, their prohibitive costs limit widespread adoption. This study demonstrates that systematic multi-modal feature integration transforms standard UAV-based RGB imagery into a powerful phenotyping instrument, achieving crop trait prediction accuracy comparable to systems costing 10–50 times more. We developed a comprehensive framework integrating spectral indices, geometric parameters, and texture metrics from commodity RGB sensors to predict five critical cotton traits: leaf area index (LAI), intercepted photosynthetically active radiation (IPAR), above-ground biomass, lint yield, and seed cotton yield. The progressive integration approach employed Random Forest regression with four feature configurations: baseline color indices (CI_base), refined color indices (CI_ref), geometric parameters (CI_ref + GP), and texture metrics (CI_ref + GP + T). Field experiments across three trials over two growing seasons (2022–2023) with varying genotypes, planting densities, and sowing dates provided 2,126 ground truth measurements for model development and validation. The optimal multi-modal model achieved R² = 0.97 for IPAR (rRMSE = 6 %), R² = 0.91 for LAI (rRMSE = 15 %), and R² = 0.85 for biomass (rRMSE = 32 %), with lint yield and seed cotton yield demonstrating R² values of 0.92 and 0.77, respectively. Variance partitioning analysis revealed texture features as the dominant contributor (16.2 % ± 7.1 %), followed by spectral indices (9.1 % ± 4.2 %) and geometric parameters (8.0 % ± 2.8 %), with substantial shared variance (45–65 %) indicating strong feature complementarity. Phenological analysis demonstrated that flowering-stage imagery outperformed boll opening stage measurements, while stage-general models showed superior robustness. Cross-temporal validation confirmed model generalizability, with trial-general models achieving R² values of 0.91–0.97 for IPAR across diverse environmental conditions. The framework enables sub-meter spatial resolution trait mapping while maintaining operational simplicity and cost-effectiveness, demonstrating that systematic feature engineering can democratize high-precision phenotyping technologies for broader agricultural applications.