The BELT and phenoSEED platforms: shape and colour phenotyping of seed samples

Abstract

Background: Quantitative and qualitative assessment of visual and morphological traits of seed is slow and imprecise with potential for bias to be introduced when gathered with handheld tools. Colour, size and shape traits can be acquired from properly calibrated seed images. New automated tools were requested to improve data acquisition efficacy with an emphasis on developing research workflows.

Results: A portable imaging system (BELT) supported by image acquisition and analysis software (phenoSEED) was created for small-seed optical analysis. Lentil (Lens culinaris L.) phenotyping was used as the primary test case. Seeds were loaded into the system and all seeds in a sample were automatically individually imaged to acquire top and side views as they passed through an imaging chamber. A Python analysis script applied a colour calibration and extracted quantifiable traits of seed colour, size and shape. Extraction of lentil seed coat patterning was implemented to further describe the seed coat. The use of this device was forecasted to eliminate operator biases, increase the rate of acquisition of traits, and capture qualitative information about traits that have been historically analyzed by eye.

Conclusions: Increased precision and higher rates of data acquisition compared to traditional techniques will help to extract larger datasets and explore more research questions. The system presented is available as an open-source project for academic and non-commercial use.

Keywords: Camera; Colour; Computer vision; Lentil; Open source; Phenotyping; Seed coat.

Fig. 1

BELT Cross section render of the physical BELT system

Fig. 2

Seed image processing flowchart. Preprocessing is conditional and can be skipped

Fig. 3

Sample images of several seeds.a lentil with a dark seed coat b canola c yellow pea d green pea e oat f wheat

Fig. 4

Sample height calibration images. a A top view, b, c two side views with different distances from the prism. Note the depth of focus is shallow and the side view of the ruler when it is far from the prism is not nearly as sharp as the other views

Fig. 5

Height calibration process. a a cropped image of a ruler b the results of summing the sobel edge filter image along the direction of the ruler rick marks c peaks in the signal are plotted against the distance between ruler tick marks

Fig. 6

Colour calibration in the L*a*b* colour space. Plotted relationships between measured and calibrated data to the supplied L*a*b* values for a subset of a ColorChecker Digital SG

Fig. 7

A sample lentil with a dark seed coat after colour calibration. This is the same lentil image showcased in Fig. 3

Fig. 8

The results of clustering lentil seed colour data. Clusters are highlighted in blue and green. Background pixels are red. Row (1) original lentil images cropped to the bounding boxes Row (2) the results of K-means clustering Row (3) the results of Gaussian mixture model clustering

Fig. 9

phenoSEED information for wild and cultivated lentils. Boxplots of shape, size and colour information for cultivated (C) and wild (W) lentil varieties

Fig. 10

Steps of segmenting the views of a lentil.a colour calibrated image b a merged image of the scaled values of a*, b* and -L*/2 added together c results of segmentation, with the top mask overlaid in blue and side mask overlaid in yellow with size measurements overlaid. Not shown are the intermediate steps of cropping to the top and side views