BloomSight: An ultra-high-frequency phenotyping framework for diurnal flowering dynamics in japonica and indica rice to enable genetic dissection and hybrid-breeding applications.

Abstract

Rice ( Oryza sativa ) production underpins food security in many rice-consuming nations. As a critical developmental transition that directly determines yield and grain quality, flowering dates and timing are genetically complex and highly sensitive to environmental fluctuations. This complexity requires new methods to quantify diurnal floral characteristics, which are essential to hybrid breeding in cereals. Here, we present BloomSight, an ultra-high-frequency and deep-learning (DL) powered framework for phenotyping and measuring minute-level flowering dynamics in japonica and indica rice. After monitoring 172 rice accessions selected from the Chinese Rice Mini-Core Collection using cost-effective time-lapse imaging platforms for 16 days, we acquired over 530,000 accession-level images and established the Open Rice Flowering Training (ORFT) dataset, with over 39,000 panicles and 350,000 anthers annotated. Next, a two-stage customised DL model (i.e. YOLACT-Panicle for panicle segmentation and UNet-Anther for anther identification) was trained using the ORFT set, enabling ultra-high-frequency measures of anther extrusion at the minute level. Based on trait analysis, we further fitted curves to dynamically identify diurnal flowering patterns, including key timepoints such as the initial flowering timepoint ( T Ini. ), quickest flowering timepoint ( T Qck. ), and peak flowering time ( T Peak ), and novel traits such as the duration of rapid flowering phase ( P Rpd. ) and flowering density across key phases. After validating BloomSight-derived traits against manual observations, we classified the japonica and indica accessions into three patterns: Slow, Moderate, and Fast, all of which had distinct flowering windows. These analyses helped us integrate phenotypic variations into a genome-wide association study (GWAS), revealing many significant single nucleotide polymorphisms (SNPs) associated with known (e.g. EMF1 , OsMYB8 , and PME42 ) and several repeatedly identified unknown loci (one of these loci has been recently verified by other groups), demonstrating the value of the BloomSight framework. Taken together, we believe that BloomSight provides an ultra-high-frequency framework for diurnal flowering phenotyping, enabling the measurement of biological meaningful floral traits with minute-level resolution that can enable flowering-related developmental studies and hybrid-breeding applications in rice and more broadly benefit the plant and crop research community.

Plant phenotyping relevance

イネの開花動態を高頻度画像と深層学習で抽出するフェノタイピング基盤を開発し、データセット構築と手動観測による検証も行っているため、方法が研究の中心である。

Code and data availability

The paper's Data and code availability statement explicitly provides public access to the ORFT annotated image dataset (BioStudies S-BSST2157), Python source code for floral trait analysis (GitHub The-Zhou-Lab/BloomSight), and trained DL models (GitHub releases). SRA accessions are molecular sequencing data, not phenot