A review of the opportunities for spectral‐based technologies in post‐harvest testing of pulse grains

Pulse grains are phenotypically diverse varying in colour, size, shape, and uniformity and have been integrated within many cultures and cuisines for several thousand years. Consumption of pulses within traditional dishes is still the dominant use for these grains, and therefore, the marketability is largely based on visual characteristics. There is also increasing interest into the utilisation of pulses in new processed food products because of their high protein content.


| INTRODUCTION: PULSE PRODUCTION AND CONSUMPTION
Pulse crops have been cultivated for several thousand years in many regions of the world (Caracuta et al., 2015;Erskine et al., 2016) and are an integral component of diverse cultures and cuisines (Sozer et al., 2017). Pulses are typically consumed as a minimally processed product, cooked either as whole grain or split grain to be included in dishes such as curries, soups and salads or sometimes ground to produce flour. Pulse consumption remains popular in regions with a strong history of traditional pulse-based dishes, such as the Indian subcontinent, the Middle East, North Africa and the Mediterranean (Singh et al., 2000). Among Western nations, pulse production is predominantly for export markets; however, because of the high protein content of some pulses, there is increasing local utilisation of these grains in new processed food products (Boukid et al., 2019;Joshi et al., 2017;Portman et al., 2018;Sozer et al., 2017).
With rising demand for high-quality pulse products, plant breeding, coupled with agronomic management strategies, aim to increase pulse-crop productivity through improved yield and disease-resistance, optimised plant stature, and adaptation across agroclimatic regions. Because of the establishment of crop improvement programmes, yields of pulse crops have grown by nearly 60% since the early 1960s (FAOSTAT, 2020), which can be attributed to using germplasm from the International Centre for Agricultural Research in Dry Areas (ICARDA). Within the typical 10-year breeding cycle to produce a new pulse variety, germplasm is exhaustively evaluated for adaptation across diverse agro-ecological regions which are subjected to abiotic effects that ultimately impact on grain yield and quality. To optimise plant management and maximise production, a range of agronomic practices such as time-of-sowing, plant row-spacing and fertiliser inputs are often superimposed on the evaluation of germplasm.
Assessment for grain quality traits occurs concomitantly with plant breeding and agronomic research to ensure newly released cultivars align with end-use quality and the specifications of markets. Testing for quality traits is based on the grains-industry standard methodology with the emphasis on quantifying the characteristics of the seed, such as size, colour and defects. Although these tests are not difficult to undertake, they are laborious, and therefore, testing is commonly withheld until later generations of the programme when the germplasm numbers derived from each inter-cross have been reduced because of other selection parameters such as grain yields or disease susceptibility.
The subjective and time-restrictive nature of standard qualitytesting protocols, and recent advances in computational and imaging technologies, has motivated research investment into the development of Machine Vision (MV) systems for product evaluation. There are opportunities to utilise these systems throughout the whole pulse-value chain from pre-breeding germplasm-screening through to post-harvest grain assessment and monitoring grain quality in storage (Aviara et al., 2022;Hayes et al., 2017;Nguyen & Norton, 2020). In the following sections of this paper, current standard techniques for pulse-quality assessment are summarised. Then recent developments of MV systems for grain quality evaluation based on digital image analysis (DIA) and near-infrared (NIR) spectroscopy are reviewed, and further opportunities for the development and application of MV systems for pulse grain evaluation are presented.

| ESTABLISHED PULSE ASSESSMENT METHODS
Within traditional pulse markets, quality of the grain product is determined by visual appeal. Characteristics of size, shape, colour, uniformity and the absence of visible defects in the grains impact on market acceptance. In the expanding plant-based protein and health-food markets, the compositional qualities, particularly protein content, of pulses are the key indicators of grain quality.

| Industry assessment at point of receival
Assessment of quality traits for pulses in the major exporting countries is largely determined by manual inspection, such as using sieves to quantitate seed size, broken/split grain, mechanical damage and insect damaged grain (Delwiche & Miskelly, 2017). Visual inspection using colour reference charts is used to determine colour defects because of disease, mould, frost, or heat stress and assign a market class according to grading standards (CGC, 2022;GTA, 2022). The interpretation of these standards by grain inspectors can be highly subjective and further complicating this, and the definitions of acceptable limits of low grade material or contaminant differ between countries (Gupta et al., 2011) often leading to trading disputes.

| Germplasm screening in breeding programmes
Breeding objectives for pulse grain quality traits are determined largely by market specifications and form part of the overall strategy of breeding programmes. Grain quality is ideally measured on the germplasm within the 12-to 16-week period between harvest and the following sowing season. Accurate and timely assessments of the grain quality are crucial to avoid sowing sub-optimal germplasm.

| Grain size
A common method to determine the distribution of grain size is to mechanically pass the grain through a vibrating stack of sieves in order of decreasing aperture size ( Figure 1). This process of mechanical sieving is known to be time consuming and prone to errors (Shahin & Symons, 2005). The irregular shape of some grains, such as chickpea, causes inconsistency in their sieving behaviour, and the sample size can also contribute to misclassifications, for example, a sieve can be blocked by the volume of seed in the majority size fraction preventing smaller seed from falling through. LeMasurier et al.
(2014) reported a misclassification rate of up to 10% in sieving lentils by this method. Through mechanical sieving, the seed size data which can be collected are limited because seed size increments are restricted to those of the sieves (Symons & Shahin, 2000). There is also an undesirable weighting of the resultant size distribution toward the larger seed because the method itself uses mass rather than the seed count in each size fraction.

| Grain colour
Colour of the whole or de-hulled grain is the primary determinant of visual quality within pulse breeding programmes and defines the market classifications in grain trading. Standard assessments of colour within grain laboratories are commonly based on the Commission Internationale de l' Eclairage (CIE) L*a*b* space. This is a globally recognised and device-independent system, and as such it is extremely useful for objective description and comparisons of colour.
However, colorimeters do not have the capability to assess colour variation or distribution within a sample (de Oliveira et al., 2016). This is undesirable in the pulse industry, where uniformity of grain colour is often an important quality for market acceptance and therefore also a breeding priority. Furthermore, the diversity within pulse grains is such that one or two of the CIE L*a*b* coordinates are typically not adequate to describe colour, and because of the difficulty in interpreting the three components concurrently, this colour space is underutilised for pulses.

| Grain processing
Pulse grains undergo minimal grain processing prior to cooking apart from dehulling of the seedcoat and splitting the cotyledons (Singh et al., 2000). The aim of the milling process is to obtain a product with minimal broken cotyledon. Within the laboratory, milling quality is assessed using small-scale equipment which is modelled on industrialscale mills. As there can be significant differences in milling quality between genotypes (Erskine et al., 1991;Wood et al., 2008), the aim is to replicate the milling outcome and identify germplasm with inherently unacceptable milling behaviour. The process involves passing grain between two carborundum discs that results in five fractions (split-dehulled, split with hull, whole-grain dehulled, whole grain with hull and broken grain). Because of its economic impact for milling companies, breeding programmes routinely undertake milling trials on their germplasm although the process is laborious as it involves manually sorting out the five fractions.

| Grain defects
During seed development, pulses are highly susceptible to diseases and pests from which the effects are manifested by undesirable changes to colour of the seed and through seed damage ( Figure 2).
For example, Ascochyta blight and Botrytis grey mould (Botrytis fabae and Botrytis cinerea) are two common diseases that infect the seed pod during seed development resulting in dark shrivelled seed of below marketable quality (Kaiser et al., 2000). In pulse crops, insects which are most active after the plant has flowered have a major effect on the quality of the seed. For example, in field pea crops the larvae of Etiella moth (pea weevil) cause damage by feeding on the developing seed within the pod rendering the seed unsuitable for processing ( Figure 2). Resistance to such pests and diseases has therefore become important germplasm selection criteria within pulse breeding programmes (Erskine et al., 2016;Khan & Croser, 2004). In addition to pests and diseases, storage conditions and handling practices can result in defective grain (Mills et al., 1999). Sub-optimal storage conditions, such as elevated temperature and moisture levels, can cause rapid discoloration of the seed coat (Nasar-Abbas et al., 2009).

| Grain composition: Protein
Analyses of compositional and nutritional grain traits have not typically been included in routine quality-testing procedures for pulses as the traditional markets are driven predominantly by visual properties. For grain types, such as cereals and oilseeds, the composition (e.g. protein, moisture or oil content) determines the market class and processing quality. Therefore, there are established analytical and NIR-predictive methods to quantitate these traits (AACC, 2022).  (Churchill et al., 1992).

| MV systems for agricultural products
Spectral-sensor and computational technologies continue to become more powerful and cost effective, and their application for product evaluation within the food and agriculture industries is continually expanding (Bhargava & Bansal, 2018;Brosnan & Sun, 2004;Mahendran et al., 2016;Nturambirwe & Opara, 2020). Sensors which capture signals or images within the NIR or visible electromagnetic regions are the most common types incorporated into MV systems for evaluation of agricultural products. NIR-based systems are useful for analysing compositional traits (e.g. protein or moisture content), and conventional digital imaging systems are useful for analysing physical and visual traits (Fox & Manley, 2014;Mahajan et al., 2015;Ratish et al., 2017). MV systems also vary in sample presentation (e.g. manual product placement or automated conveyor systems) and data processing software and algorithms, which determine the level of task automation.

| MV developments and opportunities for pulse grains
Although the number of MV applications for assessment of food-grain products is increasing (Vithu & Moses, 2016), there are still relatively few methods developed specifically for evaluation of pulse grains (Meenu et al., 2021;Velesaca et al., 2021). Tables 1 and 2  and either required grain kernels to be non-touching (manually separated) to assist image segmentation (Amaral et al., 2009;Igathinathane et al., 2009;van Dalen, 2004) or required segmentation algorithms ( Figure 3) to identify and separate touching kernels (Mebatsion & Paliwal, 2011;Shahin & Symons, 2005). Therefore, more recent advances in image acquisition incorporate mechanised grain dispersion onto conveyor systems (Figure 4) to automate the separation of grains (Halcro et al., 2020;LeMasurier et al., 2014).
Initial MV developments for seed size distribution (SSD) of lentil can be credited to Shahin and Symons (2001). Similar for rice assessment, van Dalen (2004) proposed a prediction of size distribution using images from a flatbed scanner. These studies demonstrated the potential of rapid seed-size analysis; however, it is difficult to establish their accuracy as the image predictions were not compared to seedsieving. Shahin and Symons (2005) built on previous work developing DIA methods to assess SSD of four pulse types: yellow pea, green pea, soybean and chickpea. An important finding of this study was that the DIA method was robust and repeatable even for irregular shaped grains. The authors later followed with a similar study highlighting the application of DIA to classification of soybean size uniformity and found agreement with visual assessment in excess of 84% (Shahin et al., 2006). Although some grain-size predictions from two-dimensional images have resulted in large errors relative to the size of the seed (Mandal et al., 2012;van Dalen, 2004), it has been shown in several studies that with careful image segmentation and selection of regression models, seed dimensions can be accurately predicted (Igathinathane et al., 2009;Walker & Panozzo, 2012).
LeMasurier et al. (2014) developed a DIA method for prediction of lentil SSD through a statistical model replicating the behaviour of F I G U R E 2 Pulse grains are highly susceptible to disease and insect damage. Some diseases manifest as lesions on the seed coat, such as in the images of lentil (left) and faba bean (centre). 'Pea weevil' damage (right) renders field pea samples unsuitable for processing T A B L E 1 Summary of MV applications based on processing RGB image data for the evaluation of pulse grain quality traits

Grain quality traits evaluated Grain type Reference
Grain size and shape features extracted to predict dehulling efficiency Lentil (Shahin et al., 2012) Grain size, shape and colour features extracted to classify type of lentil Lentil (Venora, Grillo, Shahin, & Symons, 2007) Grain size and colour features extracted to classify seed type and market grade Lentil (Shahin & Symons, 2002) Grain colour features extracted to classify market class Lentil (Shahin & Symons, 2001) Grain size and shape features extracted to replicate mechanical sieving outputs Colour score Field peas (Coles, 1997) Grain size Dry bean, split field pea and split chickpea (Igathinathane et al., 2009) Grain size and shape Dry beans (Kumar et al., 2013) Size and shape features extracted to classify type of bean Dry beans (Koklu & Ozkan, 2020) Grain size and colour features extracted to determine defective grain Dry beans (Kılıç et al., 2007) Grain size, shape and colour features extracted to classify type of bean Dry beans (Venora, Grillo, Ravalli, & Cremonini, 2007) Grain colour and moisture content Soybean (Huang et al., 2014) Grain size, shape, colour and texture features extracted to classify grain defects Soybean (Liu et al., 2015) Grain size Soybean (Shahin et al., 2006) Note: MV, machine vision.

T A B L E 2 Summary of MV applications based on processing NIR spectra and hyperspectral image data for the evaluation of pulse grain quality traits
Grain quality traits evaluated Grain type Sensor-data type Reference

Spectral bands identified for classification of insect damaged grain
Field pea Hyperspectral images (Nansen et al., 2014) Statistical spectral featured extracted to determine grain with fungal infection Chickpea, field pea, lentil and dry beans Hyperspectral images (Karuppiah et al., 2016) Grain protein, moisture, total fat, total fibre, ash, weight and size.
Lentil NIR spectra (Revilla et al., 2019) Grain protein and moisture content Lentil, field pea and two cereal grain types NIR spectra (Assadzadeh et al., 2020) Note: MV, machine vision; NIR, near infrared. seeds through sieves. The DIA predictions were shown to closely match the manual-sieving results with R 2 > 0.99 and Lin's concordance statistic > 0.99. Through image-based 'sieving', seed size traits can be quantified with more detail and consistency than mechanical sieving methods.

| Grain colour
Colour is a diverse and complex trait of pulse grains and intrinsically There are a range of MV applications that have been developed for colour assessment of pulses, including classification of seed type or market grades (McDonald et al., 2016;Shahin & Symons, 2001), detection of discoloured and deteriorated grains (Dell'Aquila, 2006;Jackson et al., 2021;Shahin & Symons, 2002) and quantifying the colour of bleaching-susceptible field pea varieties (Coles, 1997;McDonald et al., 2019). With careful standardisation of pixel intensity values, DIA offers greater flexibility than colorimetric methods to quantify colour uniformity objectively (McDonald et al., 2019). Colour diversity within a grain sample can be indicative of the grain condition (Coste et al., 2005), admixture or other defects, and in the context of breeding programme germplasm, colour distribution may also indicate genetic segregation. This level of objective detail cannot be captured through a colorimeter but can be readily achieved through MV applications.

| Grain processing
Milling quality of pulse grains relates to both the milling potential of whole-grain samples and to traits of the milled product such as cotyledon colour and split yield. There are very few MV applications to date that have been developed to predict aspects of pulse milling quality. The destructive and time-consuming nature of manual millingquality assessments, often referred to as dehulling the seed and splitting the cotyledon, makes MV approaches an appealing option within the breeding selection process. Typically, fewer than 15 Â 100 g samples of dehulled lentil and split cotyledon can be hand-sorted per day, so testing efficiency for split yields would increase substantially through MV assessments (McDonald et al., 2021). Predictions of milling potential from whole-grain images would enable assessments of processing quality on early generation germplasm when the volume of grain is insufficient for destructive testing. Furthermore, extension of the DIA method proposed by Shahin et al. (2012) to analysis of NIRhyperspectral images could further improve model accuracy and robustness because milling potential is impacted by moisture content (Singh et al., 2000) which can be captured through NIR-hyperspectral image analysis (Huang et al., 2014).

| Grain defects
Defective grains are those which do not fit the industry specifications for sound grain within each market class. There are a wide range of defect types which include disease, mechanical damage, dockage (small or foreign grain), discoloration, admixture and contaminants. To identify and assess defects through MV, it is useful to first create the F I G U R E 3 Segmentation of chickpea seed images captured with a flatbed scanner (Shahin & Symons, 2005) F I G U R E 4 BELT image acquisition system for shape and colour phenotyping of seed samples (Halcro et al., 2020) reference profiles for sound grain. Pulse market classes are defined by visual characteristics relating to size, shape and colour of the grain.
Image features based on these characteristics are therefore useful for creating profiles and classifying grain types or market classes and have been utilised for DIA classification of some lentil, field pea and bean types (Adjemout et al., 2007;McDonald et al., 2016;Venora, Grillo, Ravalli, & Cremonini, 2007;Venora, Grillo, Shahin, & Symons, 2007).
Classification models for seed type or market class are directly applicable to the assessment of admixture levels for objects that do not fit the market class profile of a sample (Ramirez-Paredes & Hernandez-Belmonte, 2020). MV analyses of other grain defects that impact grain colour or shape are typically assessed through DIA Dell'Aquila, 2006;Delwiche et al., 2013;Kılıç et al., 2007;Liu et al., 2015), and for the assessment of defects and infestations which impact composition, NIR-based screening (NIR scanning or NIR-hyperspectral imaging) is generally preferred (Barbedo et al., 2018;Nadimi et al., 2021;Peiris et al., 2010;Shahin & Symons, 2011;Singh et al., 2010).
NIR-hyperspectral images contain both spectral and spatial information, and the combination of these has been demonstrated to improve the predictive capacity of MV applications (Amigo et al., 2015;Assadzadeh, Walker, McDonald, et al., 2022;Weng et al., 2020). Analysis of hyperspectral images can be used to detect smaller concentrations or distributions of compounds present in grains. For example, it may be useful for detecting early stages and the location of disease in grains; this detail would be lost in the spectral noise of conventional NIR scanning. Hyperspectral image analysis has been applied to the detection of insect infestation and fungal infection in chickpeas, green field peas, lentils, pinto beans and kidney beans (Karuppiah et al., 2016;Nansen et al., 2014), but it is largely underdeveloped for pulse assessments. Developments for other grain types, particularly wheat, demonstrate the potential of hyperspectral image analysis for detecting admixture and contaminants (Mahesh et al., 2011;Ravikanth et al., 2015Ravikanth et al., , 2016, sprouting (Xing et al., 2010) and seed damage or other grading factors (Shahin & Symons, 2008;Shahin & Symons, 2011;Singh et al., 2009).

| Seed composition: Protein
Traditional pulse markets contractually define quality traits based on visual characteristics of the grain, and grain quality testing protocols have reflected the end-use requirements such as hydration capacity, cooking time, and dehulling and splitting efficiency. With the recent introduction of new markets for plant-based protein products, there is an increasing demand for pulse grain compositional traits, particularly protein, to be routinely included in quality-testing programmes. NIR spectroscopy has been shown to perform well for predictions of major cereal grain constituents and classification of grain and admixture (Barton et al., 2000;Black & Panozzo, 1999). Furthermore, NIR systems have been internationally approved and adopted for quality evaluation of some grain traits, particularly in wheat, such as the determination of ash content of flour, protein content of grain, and flour and hardness of grain (AACC, 2022). Although NIR-based systems are currently underutilised for objective and rapid assessment of pulse compositional traits, their successful application to cereal grains presents a clear opportunity. High prediction accuracies have recently been demonstrated for determination of protein content in lentil and field pea samples (Assadzadeh et al., 2020;Revilla et al., 2019).

| CONCLUDING REMARKS
Most of the global pulse production continues to target traditional markets where the visual grain characteristics are paramount to enduse quality. Meanwhile, new markets which are heavily reliant on the compositional and processing qualities of pulses are rapidly expanding. Therefore, grain traits including size, shape, colour, defects, uniformity, processing quality and protein content are important breeding targets for ensuring the market success of newly released varieties. Current standard methods for quality assessment within breeding programmes are often restricted to advanced germplasm because of the time-consuming or destructive nature of the assessments and the requirement for adequate volumes of grain to conduct the tests. Breeding programmes would therefore benefit from the adoption of rapid and objective MV methods based on visible and NIR sensors for screening grain quality traits.
Application of MV systems within germplasm screening would enable more detailed trait assessments and the capacity to screen for multiple traits simultaneously. The robustness of these systems to evaluate samples of any size would permit comprehensive quality assessments earlier within the breeding cycle, ultimately reducing the number of years between the first inter-cross and resultant variety for commercial release. The range of MV applications developed to date for assessment of pulse grains do not yet encompass the range of grain assessments that are essential within breeding programmes and the wider pulse industry, particularly relating to composition and classification of defects. There is an opportunity and necessity for this further development especially through hyperspectral image processing or the integration of DIA and NIR spectroscopy. Even so, the existing contributions for the analysis of size, shape and specific colour traits as well as the modelling of milling performance would significantly improve the current efficiency and capacity of germplasm assessment. Wider than germplasm screening, MV models have enormous potential throughout the whole pulse value chain from pre-breeding, such as rapid screening of potential breeding candidates or development of genomic prediction models, to post-farmgate product evaluation.

CONFLICT OF INTEREST
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

DATA AVAILABILITY STATEMENT
Data sharing not applicable to this article as no datasets were generated or analysed during the current study.