基于多光谱图像的小麦种子形态和成分性状的自动化检测算法开发

The development of automated analysis algorithms for characterizing wheat seeds'morphological traits and internal components using multispectral imagery

在全球气候变化加剧的背景下,小麦(Triticum aestivum)产量与中国粮食安全密切相关,因此,基于小麦种子外部形态与内部成分的快速、无损检测对高通量鉴定其品质和活力意义重大。针对当前种子检测通量及准确性的限制,本研究通过多学科融合,结合多光谱成像、计算机视觉、自动化图像处理等技术,开发了一种通过多光谱图像从近百粒种子中快速提取单粒、并在单粒尺度完成外部形态(如种子面积、长度、宽度与圆度等)和内部物质成分(如植物色素、淀粉、植物油脂与水分含量等)量化分析的算法。针对所选的15个小麦品种,算法对1 347粒种子形态性状的分析结果与人工测量结果间的决定系数(R2)为面积0.900 (RMSE=1.504)、长度0.981 (RMSE=0.188)、宽度0.911 (RMSE=0.795);对513粒种子的六个关键光谱波段的R2为0.973、0.970、0.983、0.953、0.891、0.893;以上P值均小于0.005。在此基础上,通过形态与光谱性状的聚类和主成分分析,本研究还构建了区分不同品种小麦种子的分类方法,进而探索了重要种子农艺性状(如种子破损区域及种胚检测等)的自动鉴选算法,为种子活力和品质的高通量检测研究提供了新思路和新方法。

The increasing global climate change is impacting wheat(Triticum aestivum)production,a sta-ple crop that is key to ensure China's food security.In order to safeguard wheat seed quality and vigor for better crop performance,it is important to assess seed morphological features and internal contents reli-ably and at a large scale,resulting in the importance of rapid and non-destructive seed analysis and the necessity of advancing analytical methods in this research domain.In the study presented here,we have combined multispectral seed imaging,computer vision and automated image processing techniques to address methodological problems in seed phenotyping and seed-based phenotypic analysis in terms of throughput and accuracy.We have developed an automated algorithm that can segment individual seed from hundreds of wheat seeds acquired by multispectral imaging device,through which morphological traits(e.g.seed area,length,width,and roundness)and internal components(e.g.plant pigments,starch,vegetable oil and water content,etc.)can be quantified.We have selected 15 wheat varieties and performed correlation analysis between computational results of morphological traits and manual measures for 1347 wheat seeds,resulting in coefficient of determination(R)0.900 for area(RMSE=1.504),0.981 for length(RMSE=0.188),and 0.911 for width(RMSE=0.795),P<0.001;also,the R2 of six key spectral bandwidths between computational and manual scoring for 513 seeds were 0.973,0.970,0.983,0.953,0.891,0.893,with P<0.005.Then,we utilized morphological and spectral traits in clustering and principal component analysis,establishing a classification method to differentiate wheat seed varieties;finally,we explored several important agronomic traits at the seed level,including seed coating damage and embryo.Hence,we believe that the methods described here provide new approaches for high-throughput seed phenotyping,enabling biological discoveries with regards to seed quality and vigor studies.