UAV-derived dynamic stay-green modeling enhances yield prediction for screening drought-resistant and water-saving winter wheat varieties

Screening drought-resistant and water-saving winter wheat varieties is critical for enhancing grain yield (GY) and alleviating the water resource crisis in the North China Plain (NCP). Although unmanned aerial vehicles (UAVs) offer a powerful high-throughput screening platform, current assessments primarily rely on discrete temporal observations, failing to capture continuous senescence dynamics and limiting inter-annual transferability. To overcome this, this study proposes a framework targeting the stay-green trait—which is closely associated with GY—to dynamically characterize senescence and standardize phenological trajectories. Based on a field experiment involving 211 winter wheat varieties under three irrigation treatments (W2: 150 mm, W1: 75 mm, W0: rain-fed), a Stay-green-Senescence (Sg-S) model was established. By coupling a representative vegetation index (VI) with accumulated temperature (AT), this model simulates the dynamic leaf senescence curves of different varieties. From their mathematical features, 9 phenological parameters (e.g., critical senescence inflection points and the area under the curve) were extracted and integrated into machine learning algorithms for achieving improved inter-annual GY prediction and reliable screening of drought-resistant and water-saving varieties. The main findings were: 1) High-performing varieties maintained higher greenness intensity during the early-to-mid senescence phases and reached higher AT by maturity, providing a physiological basis for high-throughput screening; 2) to capture these dynamics, the Red-edge chlorophyll index (RECI) was selected as the representative VI to construct the Sg-S model. The extracted modeling parameters significantly enhanced GY predictive accuracy over raw VIs, improving R 2 from 0.60 to 0.66 (internal) and 0.68–0.74 (inter-annual); 3) The framework achieved 86.73%–87.04% screening accuracy and 78.57%–94.44% recall across two seasons. Screened elite varieties exhibited clear advantages in both GY (+8.26%–13.82%) and water use efficiency (+4.56%–10.40%) compared to the average of the evaluated population. This framework supports high-throughput screening of drought-resistant and water-saving varieties and sustainable breeding and cultivation strategies in the NCP and similar agroecological regions.