Deep learning combined with UAV to assist in leaf color breeding of Chinese cabbage

Abstract

Leaf color is an important trait affecting vegetable quality, yield, and market value. However, traditional methods for leaf color assessment are often subjective or destructive, which limits accurate and high-throughput phenotyping. In this study, an unmanned aerial vehicle (UAV)–based multispectral imaging platform was used to collect phenotypic data from 214 Chinese cabbage inbred lines at the rosette stage. A multispectral UNet model was applied to segment individual plants, and a membership function was used to quantify leaf color on a continuous scale. Based on these high-throughput phenotypic data, a genome-wide association study was used to identify two candidate genes, BrEMB976 and BrGSH2, on chromosome A06. Subsequent virus-induced gene silencing analysis showed that silencing these genes altered leaf color. In addition, a deep learning-based genomic selection model, BrDeepGS, was developed for leaf color prediction, which achieved a Pearson correlation coefficient of 0.853. These results demonstrate the potential of integrating UAV-based high-throughput phenotyping, candidate gene analysis, and genomic prediction for leaf color evaluation and selection in Chinese cabbage breeding.