A in-seasonal phenology monitoring approach for wheat breeding accessions with time-series RGB imagery by using a combination KNN-CNN-RF model

Abstract

Accurate and near real-time monitoring of wheat phenology is crucial for both cultivation and breeding. It plays a pivotal role in guiding planting management, optimizing variety selection, enhancing yield and quality, and providing a scientific foundation for wheat production. This study aims to develop a high-throughput approach for identifying the real-time phenological stages and estimating the initiation date of key stages for abundant breeding accessions using UAV-derived RGB imagery. A two-year field experiment was conducted across diverse wheat accessions worldwide, including the Watkins landraces and modern varieties at different ecological locations in three provinces of Guangdong, Jiangsu, and Hebei. The minimalist neural network model VanillaNet was employed to classify the five phenological stages: tillering stage (TS), jointing and booting stage (JBS), heading stage (HS), anthesis and filling stage (AFS), and maturity stage (MS), based on image features. Meanwhile, the K-nearest neighbors algorithm categorized the phenological stage into three class—TS, JBS and heading to maturity stage (HMS)—using UAV-derived three-dimensional height data. To improve the classification accuracy, the weight coefficient was introduced to integrate the prediction probability of two classifiers. Finally, a random forest (RF) model was developed to estimate the initiation dates of key phenological stage based on the integrated time-series prediction probabilities. The results showed that the integrated classifier exhibited accuracies of 0.96, 0.88, 0.66, 0.87, and 0.96 in the five stages, respectively. Compared to the classification results obtained solely using neural network models, the increase in the F1-scores for the first three phenology stages was 7.41 %, 5.81 %, and 13.16 %, respectively. After stage classification, the RF model demonstrated robust performance in predicting the initiation dates of jointing, heading, anthesis, and maturity stages, with a coefficient of determination of 0.61–0.91 and a root mean square error of 1.83–4.09 days. Furthermore, the accuracy of phenological monitoring was analyzed under different data collection frequencies, revealing that the optimal interval for data collection was within 5–13 days. The proposed methodology realized synchronously the classification and quantification of phenological stages, thereby serving as a high-throughput screening technology of fine variety in smart wheat breeding.