High-throughput phenotyping and deep learning to analyze dynamic panicle growth and dissect the genetic architecture of yield formation

Abstract

The dynamic growth of shoots and panicles determines the final agronomic traits and yield. However, it is difficult to quantify such dynamics manually for large populations. In this study, based on the high-throughput rice automatic phenotyping platform and deep learning, we developed a novel image analysis pipeline (Panicle-iAnalyzer) to extract image-based traits (i-traits) including 52 panicle and 35 shoot i-traits and tested the system using a recombinant inbred line population derived from a cross between Zhenshan 97 and Minghui 63. At the maturity stage, image recognition using a deep learning network (SegFormer) was applied to separate the panicles from the shoot in the image. Eventually, with these obtained i-traits, the yield could be well predicted, and the R2 was 0.862. Quantitative trait loci (QTL) mapping was performed using an extra-high density single nucleotide polymorphism (SNP) bin map. A total of 3,586 time-specific QTLs were identified for the traits and parameters at various time points. Many of the QTLs were repeatedly detected at different time points. We identified the presence of cloned genes, such as TAC1, Ghd7.1, Ghd7, and Hd1, at QTL hotspots and evaluated the magnitude of their effects at different developmental stages. Additionally, this study identified numerous new QTL loci worthy of further investigation.

Keywords

• Deep learning
• Growth prediction
• High-throughput phenotyping
• Panicle growth
• QTL
• Rice