Accurate estimation of grain number per panicle in winter wheat by synergistic use of UAV imagery and meteorological data

Abstract

Rapid, accurate, and nondestructive estimation of grain number per panicle (GNPP) in winter wheat is crucial to accelerate smart breeding, improve precision crop management, and ensure food security. As two (panicle number per unit ground area and GNPP) of three commonly used yield components, GNPP was much less quantified with remotely sensed data than the former through visual counting. The limited research suffered from either low accuracies with ground canopy spectra or low efficiency with proximal panicle imaging systems. No studies have been reported on estimating GNPP with unmanned aerial vehicle (UAV) imagery, underscoring its strong advantages in high-resolution and efficient monitoring. To address these issues, this study proposed a practical approach for estimating GNPP in winter wheat by integrating UAV imagery and meteorological data with meta-learning ensemble regression. The potential contributions of different variables were examined for understanding the improvement in the spectral estimation of GNPP, including spectral indices (SIs), the optimal canopy height (CH) metric, and absorbed photosynthetic active radiation (APAR).

The results demonstrated that CH_P99 (CH at the 99th percentile in the region of interest) derived from red-green-blue (RGB) imagery exhibited the strongest correlation with measured plant height among all RGB- or multispectral (MS)-derived CH metrics. The incorporation of remotely sensed APAR and RGB-derived CH_P99 improved the accuracy of GNPP estimation over using merely color indices or SIs. Among all feature combinations, Comb. #6 (SIs + APAR_MS + CH_P99) yielded the highest overall accuracies in estimating GNPP for individual and multiple stages. Compared with the best anthesis models for Combs. #5–7 (R_val² = 0.52–0.64, RMSE = 2.85–2.47, RRMSE = 6.01–5.21 %), the multi-stage (heading + anthesis) models exhibited higher accuracies in independent validation (R_val² = 0.60–0.65, RMSE = 2.60–2.42, RRMSE = 5.48–5.10 %). The findings suggest this study has opened a new avenue for estimating GNPP with UAV remote sensing. The proposed method for the synergistic use of UAV imagery and meteorological data has great potential in the prediction of GNPP and grain yield over various regions with satellite imagery and climate datasets.