A Low-Cost RGB-Based Image Processing Method for High-Throughput Assessment of Rice Grain Chalkiness

doi:10.1016/j.rsci.2025.12.003

Abstract

Abstract:

Although numerous rice genotypes have been developed worldwide, post-harvest evaluation of chalkiness, a key grain trait, remains a significant challenge in breeding programs. Conventional phenotyping methods rely on manual grain separation and analysis, which limit the speed and performance of decision-making. This study aimed to assess the efficiency of a low-cost, image-based phenotyping method for characterizing rice grain chalkiness and morphological traits (grain length and width) in comparison with traditional evaluation methods. Grains from 270 rice samples were imaged using a hyperspectral camera (visible to near-infrared, 400-1 000 nm) and a Nikon digital single-lens reflex (DSLR) camera. Only RGB information was used for analysis, including RGB channels extracted from hyperspectral imagery to simulate low-cost setups. Python scripts were used to segment grains, estimate morphological parameters, and calculate chalkiness degree. Results from both imaging systems were compared with reference data obtained from the SeedCount platform. Strong correlations with SeedCount reference data were observed for chalkiness degree, reaching 93% when using hyperspectral-derived RGB data and 76% when using DSLR-acquired RGB images. Binary classification metrics showed high discriminative performance, with area under the curve (AUC) values above 0.90 for most traits. The proposed method enabled image acquisition and processing in approximately 21 s per sample, compared to 1.5 min required by the conventional platform. The findings demonstrate the feasibility of a rapid and low-cost image-based phenotyping strategy to support rice breeding programs, particularly for chalkiness quantification and grain morphology assessment. The complete image-processing pipeline is provided as supplementary material, reinforcing the transparency and reproducibility of the method.

Key words: Oryza sativa, grain trait, chalkiness degree, hyperspectral imagery, high-throughput phenotyping

Jardel da Silva Souza, Sandra Helena Uneda-Trevisoli, Felipe Dalla Lana, Roberto Fritsche-Neto. A Low-Cost RGB-Based Image Processing Method for High-Throughput Assessment of Rice Grain Chalkiness[J]. Rice Science, 2026, 33(3): 381-391.

Figures/Tables 9

Fig. 1. Correlations between hyperspectral image-based estimates and SeedCount reference values for rice grain traits. A‒C, Correlations of chalkiness degree (A), grain length (B), and grain width (C) between hyperspectral image-based estimates (New) and SeedCount reference values (Old). Each point represents one sample. Pearson’s correlation coefficient (r) is shown in each panel.

Fig. 2. Correlations between SeedCount reference data and Nikon digital single-lens reflex (DSLR)-acquired RGB image-based estimates for rice grain traits. A‒C, Correlations of chalkiness degree (A), grain length (B), and grain width (C) between values obtained by Nikon DSLR/RGB camera analysis method (New) and SeedCount (Old). Each point represents one sample. Pearson’s correlation coefficient (r) is shown in each panel.

Fig. 3. Mean spectral reflectance profiles of three rice grain samples obtained with a hyperspectral imaging system. Each curve represents average reflectance across the full wavelength range (400-1 000 nm), illustrating spectral variability among different genotypes.

Fig. 4. Performance curves for digital single-lens reflex (DSLR)-based phenotyping system. ROC, Receiver operating characteristic; PR, Precision-recall; AUC, Area under the curve.

Fig. 5. Performance curves for hyperspectral-based phenotyping system. ROC, Receiver operating characteristic; PR, Precision-recall; AUC, Area under the curve.

Table 1. Comparative performance metrics of digital single-lens reflex (DSLR) and hyperspectral imaging for rice grain quality assessment.

Attribute	Metric	DSLR	Hyperspectral	Best
Chalkiness degree	Accuracy	0.52	0.65	Hyperspectral
	Correlation	0.76	0.93	Hyperspectral
	F1-score	0.07	0.45	Hyperspectral
	PR AUC	0.80	0.91	Hyperspectral
	ROC AUC	0.81	0.92	Hyperspectral
Grain length	Accuracy	0.50	0.50	Tie
	Correlation	0.93	0.89	DSLR
	F1-score	0.67	0.67	Tie
	PR AUC	0.97	0.95	DSLR
	ROC AUC	0.97	0.95	DSLR
Grain width	Accuracy	0.52	0.52	Tie
	Correlation	0.94	0.83	DSLR
	F1-score	0.67	0.00	DSLR
	PR AUC	0.97	0.90	DSLR
	ROC AUC	0.97	0.90	DSLR

Fig. 6. Workflow of proposed image-based phenotyping platform. Both RGB and hyperspectral images were acquired and processed through the same Python-based pipeline. In the case of hyperspectral imaging, only RGB channels were extracted from the spectral cube. The workflow includes image acquisition, segmentation, trait extraction, validation against SeedCount data, and performance evaluation.

Fig. 7. Image-processing steps applied to hyperspectral images. A, Raw hyperspectral image. B, Thresholded grayscale image used for segmentation. C, Binary mask after morphological operations. D, Final segmented grains used for feature extraction.

Fig. 8. Processing of RGB camera images.

References 24

[1]	Araus J L, Cairns J E. 2014. Field high-throughput phenotyping: The new crop breeding frontier. Trends Plant Sci, 19(1): 52-61.
[2]	Ashida K, Iida S, Yasui T. 2009. Morphological, physical, and chemical properties of grain and flour from chalky rice mutants. Cereal Chem, 86(2): 225-231.
[3]	Aznan A, Gonzalez Viejo C, Pang A, et al. 2021. Computer vision and machine learning analysis of commercial rice grains: A potential digital approach for consumer perception studies. Sensors, 21(19): 6354.
[4]	Cervantes Wilches A P, Silva Carreño M O, Aroca Araújo A A. 2023. Mathematics in the preparation and sale of Colombian rice pies and its potential contribution to mathematics education. Acta Sci, 25(6): 272-298.
[5]	Chicco D, Jurman G. 2020. The advantages of the Matthews correlation coefficient (MCC) over F 1 score and accuracy in binary classification evaluation. BMC Genomics, 21(1): 6.
[6]	Crossa J, Pérez-Rodríguez P, Cuevas J, et al. 2017. Genomic selection in plant breeding: Methods, models, and perspectives. Trends Plant Sci, 22(11): 961-975.
[7]	da Silva Souza J, Pedrosa L M, de Almeida Moreira B R, et al. 2022. The more fractal the architecture the more intensive the color of flower: A superpixel-wise analysis towards high-throughput phenotyping. Agronomy, 12(6): 1342.
[8]	Dozal L, Silván Cárdenas J L, Moctezuma D, et al. 2018. Evolutionary approach for detection of buried remains using hyperspectral images. Photogramm Eng Remote Sens, 84(7): 435-450.
[9]	Fitzgerald M A, McCouch S R, Hall R D. 2009. Not just a grain of rice: The quest for quality. Trends Plant Sci, 14(3): 133-139.
[10]	Hlisnikovský L, Menšík L, Barłóg P, et al. 2023. How weather and fertilization affected grain yield and stability of winter wheat in a long-term trial in the South Moravian region, Czech Republic. Agronomy, 13(9): 2293.
[11]	Hori K, Sun J. 2022. Rice grain size and quality preface. Rice, 15(1): 33.
[12]	Liu F, Yang R, Chen R Q, et al. 2024. Digital techniques and trends for seed phenotyping using optical sensors. J Adv Res, 63: 1-16.
[13]	Mishra P, Lohumi S, Khan H A, et al. 2020. Close-range hyperspectral imaging of whole plants for digital phenotyping: Recent applications and illumination correction approaches. Comput Electron Agric, 178: 105780.
[14]	Ni B, He Y, Chen Q. 2019. Automatic morphological characterization of paddy rice using computer vision and deep learning techniques. Food Analytical Methods, 12:1851-1862.
[15]	Powers D M W. 2011. Evaluation: From precision, recall and F-measure to ROC, informedness, markedness & correlation. J Machine Learning Technol, 2(1): 37-63.
[16]	Smiderle O J, Dias J R M. 2008. Harvest time and physiological quality of seeds in irrigated rice (Oryza sativa cv. Brs Roraima). Res Agron Trop, 38(3): 123-131. (in Portuguese)
[17]	Sokolova M, Lapalme G. 2009. A systematic analysis of performance measures for classification tasks. Inf Process Manag, 45(4): 427-437.
[18]	Sreenivasulu N, Butardo V M Jr, Misra G, et al. 2015. Designing climate-resilient rice with ideal grain quality suited for high-temperature stress. J Exp Bot, 66(7): 1737-1748.
[19]	Su Y, Xiao L T. 2020. 3D visualization and volume-based quantification of rice chalkiness in vivo by using high resolution micro-CT. Rice, 13(1): 74.
[20]	Tam B P, Tu P T B, Pha N T, et al. 2023. Correlation among quality characteristics in medium-grain rice. Vietnam J Agric Sci, 6(2): 1765-1777.
[21]	Tashiro T, Wardlaw I F. 1991. The effect of high temperature on kernel dimensions and the type and occurrence of kernel damage in rice. Aust J Agric Res, 42(3): 485-496.
[22]	Wang C X, Caragea D, Narayana N K, et al. 2022. Deep learning-based high-throughput phenotyping of chalkiness in rice exposed to high night temperature. Plant Methods, 18(1): 839.
[23]	Wang K, Li Y G, Wang Y, et al. 2017. On the asymmetry of the urban daily air temperature cycle. J Geophys Res Atmos, 122(11): 5625-5635.
[24]	Yu F H, Feng S, Yao W X, et al. 2020. BAS-ELM based UAV hyperspectral remote sensing inversion modeling of rice canopy nitrogen content. Int J Precis Agric Aviat, 3(3): 59-64.