基于图像的水稻穗粒相关性状智能检测算法研究

Dong Wang, Chen Jiawei, Liyan Shen, A-hong Wang, Ji Zhou (Lead Author)

Research output: Contribution to journalArticlepeer-review

1 Downloads (Pure)

Abstract

Rice (Oryza sativa) and its production is of great importance to China’s food security. To assess its yield potential and grain quality, spikelet-related traits such as the number of spikelets per spike, seed coat color, spikelet length, width and length, and width ratio (W/L ratio) can be employed as key phenotypic evidence. Nevertheless, the present methods for spikelet-related trait analysis are still largely relying on manual counting or various post-harvest grain testing equipment, which is time-consuming, prone to error, and costly. Here, we present an original phenotypic analysis algorithm that can analyze key spikelet-based
traits based on images collected by high-definition cameras, photo scanners, or smartphones, under both indoor and outdoor conditions. The vision-based algorithm combines automated image processing and deep learning techniques with domain knowledge in rice spikelet measurement, which does not require any prerequisites such as threshing or post-harvest processing. Besides the number of spikelets per spike, the algorithm can quickly detect and measure complete rice spikelets from the acquired images, quantifying spikelet-related traits such as grain length, grain width, W/L ratio, circularity, and grain coat color. We chose 18 rice varieties to verify the algorithm using images collected from indoor and in-field settings. Based on the trait analysis results, we examined correlations (i.e. the coefficient of determination, R2) between the computational and manual scoring for traits such as the number of spikelets per spike and concluded significant positive correlations: R2 =0.95 (P<0.001; n=4 930 spikelets, scored from indoor images), R2=0.88 (P<0.001; n=4 191, scored from indoor rice spikelets), R2=0.86 (P<0.001; n=2 490, scored from infield images), and R2=0.85 (P<0.001; n=2 645, scored from in-field rice spikelets). Additionally, we performed correlation analysis using the W/L ratio trait and concluded strong positive correlations, R2=0.84 (P<0.001; n=1 454, scored from indoor images) and R2=0.71 (P<0.001; n=726, scored from in-field images). Both correlations indicate a mechanistic link between the algorithm-derived traits and plant specialists’ scores. The above results suggest that the work presented here is reliable for automated phenotypic analysis of spikelet-related traits in rice. Furthermore, through cluster analysis, the 18 tested rice varieties can be reliably divided into four categories based on the spikelet-related trait analysis produced by our algorithm, similar to the clustering analysis results using manual scoring. This indicates that our work can be employed to enable the identification of phenotypic variation of these rice varieties, effectively. Hence, we trust our work can have a wide range of applications to provide an original, low-cost, accurate and generalized algorithmic solution for spikelet-related trait analysis in rice, enabling biological studies in this important research domain.
Translated title of the contribution基于图像的水稻穗粒相关性状智能检测算法研究
Original languageMultiple languages
Pages (from-to)957–971
Number of pages15
JournalPlant Physiology Journal
Volume58
Issue number5
DOIs
Publication statusPublished - 8 May 2022

Keywords

  • Rice
  • spikelet-related traits
  • automated image processing
  • deep learning
  • cost-effective phenotypic analysis of spikelets

Cite this