Determination of dry bean physiological maturity from UAS RGB time series using a stacking ensemble of color‐derived spectral features

Abstract Evaluating physiological maturity is an important trait in dry bean breeding ( Phaseolus vulgaris L.), but field scoring can be inaccurate, subjective, and a persistent bottleneck in multi‐environment trials. To address this, we developed a low‐cost, high‐throughput pipeline that predicts plot‐level days after planting at maturity from time‐series unmanned aerial system (UAS) red–green–blue imagery. The dataset used in this study was comprised of 22 trials across four site‐years and two Michigan locations (Montcalm 2021; Saginaw Valley Research and Extension Center 2022, 2024, and 2025), spanning six market classes (Black, Navy, Yellow, Great Northern and Pinto, Red and Pink, and Mixed). For initial maturity estimation, a 36‐feature predictor matrix was developed by combining 23 color‐based maturity methods, including chromatic coordinates, visible‐spectrum vegetation indices, and a newly introduced hue maturity index, with five time‐series‐derived dynamic features, four flight‐metadata descriptors, and four cross‐method consensus statistics. To ensure generalizability on completely unseen trials, six base regression models and a random forest stacking meta‐learner were evaluated under leave‐one‐trial‐out (LOTO) cross‐validation. Among base learners under LOTO, gradient boosting was the strongest single predictor (root mean squared error [RMSE] = 2.74 days), followed closely by random forest (2.77 days). The stacking ensemble outperformed all base models, achieving an RMSE of 2.68 days ( r = 0.62) under LOTO cross‐validation; error rose to 3.5–3.6 days under stricter cross‐year (leave‐one‐year‐out) and cross‐environment (leave‐one‐environment‐out) evaluation, while a more permissive random 10% holdout yielded 2.40 days ( r = 0.90). This tool enables breeders and agronomists to estimate maturity directly from UAS imaging and keep prediction accuracy at the cross‐validation level while reducing flight count.