Genome-wide association study of rice (Oryza sativa L.) leaf traits with a high-throughput leaf scorer

Highlight A combination of high-throughput leaf phenotyping and genome-wide association analysis provides valuable insights into the genetic basis of rice leaf traits.


Introduction
Leaves are primarily involved in photosynthesis and transpiration, influencing yield performance in crops (P Wang et al., 2011). The size, shape, and number of leaves determine a plant's photosynthetic potential and play important roles in determining plant yield, disease resistance, and stress responses (Pérez-Pérez et al., 2010). In addition to these traits, the colour (or degree of leaf greenness) is also an important trait related to the leaf nitrogen status (Singh et al., 2002). Because rice (Oryza sativa) is a staple food for billions of people throughout the world, particularly in Asia, Africa, and Latin America (Zhang, 2007), it is necessary to systematically study the genetic basis of leaf traits in rice.
Many previous studies have uncovered how the leaf size and shape is controlled from the perspective of leaf development (Moon and Hakes, 2011;Tsukaya, 2003). Leaf development starts with the recruitment of founder cells from the shoot apical meristem, which is characterized by the down-regulation of knotted1-like homeobox genes (Byrne, 2005). The over-expression of five of these genes (OsH1,OsH6,OsH15,OsH71,and OsH43) in rice profoundly affects leaf formation and results in severely malformed leaves (Sentoku et al., 2000). WUSCHEL-related homeobox genes also play an important role in recruiting founder cells as a meristem organizer (Kessler and Sinha, 2004). In rice, two WUSCHEL-related homeobox genes, namely NAL2 and NAL3 (NAL2/3), affect leaf lateral-axis outgrowth and leaf width (LW) (Cho et al., 2013). After recruitment, several genes act to establish polarity in developing leaves, of which the class III homeodomain leucine zipper (HD-Zip III) genes specify leaf adaxial identity (Kessler and Sinha, 2004). OsHB genes are members of the rice HD-ZIP III gene family, and the ectopic expression of OsH1, OsH3, and OsH5 results in rolled and filamentous leaves (Itoh et al., 2008). In addition, some genes associated with leaf size and shape have been identified via mapbased cloning in rice. For example, the mutant nal1 exhibits reduced LW, a decreased number of longitudinal veins, and a defective vascular system. The mutant nal1 encodes a plant-specific protein that affects polar auxin transport and vascular patterns (Qi et al., 2008). NRL1 and NAL7, which are associated with leaf size and shape, were cloned and characterized using the same approach (Fujino et al., 2008;Hu et al., 2010).
In comparison with leaf size and shape, there is less research on leaf colour in rice. The mutant yellow green leaf 1 (ygl1), a spontaneous mutant isolated from indica rice 'Zhenhui 249', exhibits a yellow-green leaf phenotype. This phenotype is caused by a mutation in the YGL1 gene encoding the chlorophyll synthase as demonstrated via map-based cloning and genetic complementation analyses (Wu et al., 2007). Another mutant yellow green leaf 7 (ygl7) is also characterized by a yellow-green leaf phenotype during its whole lifespan and results from a mutation in YGL7, which encodes a magnesium-chelatase ChlD protein (Deng et al., 2014). As an alternative approach, some quantitative trait loci (QTL) have been mapped for leaf size, shape, and colour (Farooq, et al., 2010;P Wang et al., 2011;Yue et al., 2006). However, the genetic basis of leaf traits is largely unclear because only a limited number of leaf traits such as leaf length (LL) and LW have been studied. To our knowledge, a genome-wide association study (GWAS) of leaf angle, LL, and LW to dissect the genetic architecture has only been performed in maize (Tian et al., 2011); no GWAS of these leaf traits have been conducted on the abundant natural variation in plants with multiple tillers such as rice. It is necessary to develop a reliable, automatic, and highthroughput scorer for leaf size, shape, and colour evaluations of rice and other crops. Current estimation methods for the leaf area or leaf area index (LAI, which is defined as the ratio of the leaf area to a given unit of land area) are either non-destructive or destructive. The non-destructive methods include a canopy analyser (e.g. LAI-2000, LI-COR, Lincoln, NE, USA) (Sone et al., 2009;Wijk and Williams, 2005); digital camera (Shibayama et al., 2011a); near-infrared digital camera (Shibayama et al., 2011b); and hyperspectral bands (F Wang et al., 2011), which have been adopted to estimate or predict the LAI in fields. However, when measuring all the leaves on each individual plant, these methods are not suitable because of their inefficiency and inaccuracy. If using a desktop scanner (DS), the leaf area per plant can be extracted using a subsequent image analysis (Bakr, 2005;Bylesjö et al., 2008;O'Neal et al., 2002); however, this method is inefficient because each clipped leaf must be scanned individually. Even if the leaves are pasted on white paper for scanning, the efficiency is not satisfactory because the leaf pasting is also time-consuming and labourintensive, particularly when dealing with curly leaves resulting from dehydration. Besides measuring leaf size and shape, the traditional evaluation of leaf colour using the leaf colour chart (developed by the International Rice Research Institute, or IRRI) is inefficient, especially when processing a large set of samples.
A high-throughput rice phenotyping facility as recently established that can simultaneously measure multiple morphological traits in rice, including the leaf area per plant, in a non-destructive manner . However, there is also a need for a high-throughput and high-accuracy tool to extract almost all the leaf traits we can define, which is a very important complementary tool to the highthroughput rice phenotyping facility in the rice phenotyping platform. Here, a reliable, automatic high-throughput leaf scorer (HLS) is presented for leaf trait evaluations of a large number of rice accessions by using photonics vision-based techniques. The performance of the HLS for leaf number, area, colour, and shape, was evaluated. Moreover, 29 leaf traits at three different growth stages were extracted with the HLS system, and GWASs of these traits were performed. The results demonstrate that a combination of GWAS and HLS can provide valuable insight into the genetic architecture of leaf traits.

Plant material
Detailed information on the rice accessions and genotyping were described in a previous study . To test the HLS's performance in measuring green leaf area (GLA), total leaf area (TLA), and leaf number (LN), 163 rice accessions were randomly selected. To evaluate the measuring accuracy of LL, LW, and green leaf colour discrimination of the HLS system, three batches of rice leaves (50 leaves for each batch,) were selected using the leaf colour chart developed by the IRRI: green-2 (light green), green-3 (moderate green), and green-4 (dark green). Each batch was then measured with the HLS. Simultaneously, the LL and LW of 150 leaves were measured manually. Moreover, 455 accessions at late tillering stage, 469 at late booting stage, and 389 at milk grain stage were phenotyped with HLS. The experimental arrangement of the rice material is shown in Supplementary Fig. S1.

Constructing the HLS system
The HLS system contains two primary elements: a control unit and an inspection unit (Fig. 1a). The control unit is designed for user interaction, system control, image analysis, and data management. The inspection unit is designed for leaf feeding, image scanning, and image transmission. The control unit includes the following three primary devices: a workstation (HP xw6400, Hewlett-Packard Development Company, USA), a camera-link image acquisition device (IMAQ PCI-1428, National Instruments Corporation, USA) that digitizes the images into 3-colour 8-bit files, and a programmable logic controller (PLC, CP1H-Y20DT-D, Omron Corporation, Japan). The details of the operation procedure, data management, and control unit are introduced in Supplementary Video S1, Supplementary Fig. S2, Supplementary Fig. S3, and Supplementary Notes. The inspection unit The inspection unit contains a camera, a light source, and a leaf feeding device; the details are shown in Fig. 1b. A colour line-scan camera (TAPIX LCD 2048@9000, Tattile International Ltd., Italy), which is equipped with a 28 mm lens (F1.8 EX DG Aspherical Macro, SIGMA, Japan), is used to acquire a colour image with 1280 × 3000 (horizontal × vertical) pixels. A linearray light source (BL-494-40-W, Cogstek Automation Technology Co. Ltd., China; set to 26.4 V and 0.9 A) is fixed to provide stable illumination for the imaging system. A feeding device (shown in Fig. 1b), which includes four stainless steel cylinders and a conveyor, transports the leaves automatically (even if the leaves are curly). The conveyor is driven by a servo motor (MHMD042P1U, Panasonic Corporation, Japan). Finally, all the scanned leaves are transferred and gathered in a collecting box.
The configuration of the inspection unit is shown in Fig. 1c. The pixel size of the camera is 14 × 14 μm, and the spatial resolution of the inspection system is 0.22 × 0.22 mm (0.048 mm 2 /pixel), which is a magnification of ~15.71. The efficient image size of the line-scan camera is 17.94 mm × 0.24 mm; therefore, the field of view (FOV) of the inspection unit is 281.84 mm × 3.77 mm. To ensure that there is a sufficient image FOV with the diameter of the cylinder (30 mm), the centre-line space between the second cylinder and the third cylinder (d v ) must be designed to be more than 33.77 mm. In the inspection unit, the d v is designed to measure 45 mm.
The software interface (designed by LabVIEW, National Instruments Corporation, USA) is shown in Fig. 1d. The original image and green/yellow segmentation image, along with the barcode and 29 leaf traits (defined in Supplementary Table S1), are displayed on the interface in real time. Image analysis The colour line-scan camera consists of three monochrome chips, which are fixed in parallel to trap the red, green, and blue (RGB) spectra, respectively. Thus, the original colour image is distorted in RGB colour space (Fig. 2c, e). After the three greyscale images (RGB components) are extracted using a predefined value, the greyscale images of the green and blue components are shifted up a level with the position of the red greyscale image. Finally, the three greyscale images are merged back together to form a corrected RGB colour image (Fig. 2d, f). The source code for image correction is introduced in Supplementary Notes.
After the RGB colour image is corrected, the image analysis procedures for the two adjoining images (Fig. 2g, h) include the following steps: (i) With the predefined segmentation values in hue, saturation, and illumination (HSL) colour space, the adjoining colour images are transferred to the binary images ( Fig. 2i, j); (ii) the leaves part in the previous image, and they are connected to the lower boundary ( Fig. 2k), cut, and reconnected to the remaining leaf portion in the next image (Fig. 2l); (iii) some dust may appear on the mosaic image ( Fig. 2m), so after removing the isolated pixels, the binary leaf image without impurities can be extracted (Fig. 2o); (iv) the morphological traits of each leaf, including the LL, LW, leaf perimeter, leaf area, and leaf compactness can be extracted; (v) after the images (Fig. 2g, h) are segmented with predefined values in HSL colour space for the GLA ( Fig. 2p and Fig. 2q), the total leaf image without impurities ( Fig. 2o) is also obtained as a mask for the subsequent extraction of the GLA, and the binary images of green leaves without impurities are shown in Fig. 2r, s; and (vi) with a similar procedure and predefined segmentation values, green-2 leaf area, green-3 leaf area, and green-4 leaf area are also extracted by the set intersection of the corresponding segmented regions with impurities and the mask. When the single task is finished (such that the total leaves of an individual plant have all been measured with the HLS), the leaf area is determined by the sum of the pixel numbers for the current task and the spatial resolution of the inspection system (0.048 mm 2 /pixel). The source code for the primary image analysis procedure is introduced in Supplementary Notes.

Desktop scanner method
After the leaves of an individual rice plant were measured with the HLS, a DS (BenQ 8800, BenQ Corporation, China) was used to obtain the actual GLA and TLA (Caldas et al., 1992). The clipped leaves were pasted on white sheets of paper, and then the papers were placed in the scanner. The scanner was set at 300 dpi to obtain 8-bit colour images. Using a fixed HSL colour threshold for green leaves (H: 39-131; S: 34-255; L: 30-233) and yellow leaves (H: 0-38; S: 34-255; L: 30-233), green leaves and yellow leaves were extracted from the background. Based on the foreground pixel counts of all the leaves that belong to the same plant, the GLA and TLA for individual rice plants were obtained.

Genome-wide association study
A total of 4,358,600, 2,863,169, and 1,959,460 single nucleotide polymorphisms (SNPs; minor allele frequency ≥0.05; the number of accessions with minor alleles ≥6) were used in the GWAS for 29 traits at three different growth stages in the whole population and in indica and japonica sub-populations, respectively. The detailed method of GWAS is described in a previous study . Suggestive (1/N) and significant (0.05/N) P-value thresholds were set to control the genome-wide type 1 error rate (Duggal et al., 2008;Li et al., 2013), and N represented the effective number of independent markers calculated using the GEC software tool (Li et al., 2012). The software TASSEL version 3 (Bradbury et al., 2007) was used to calculate phenotypic variation explained by individual SNPs after fitting other model terms in a mixed linear model, effect size of individual SNPs, and heritability [defined as the proportion of variance of the random effect (polygene) over total in mixed linear model]. Epistasis tests of pairwise SNPs were conducted with the function-epistasis in PLINK (http://pngu.mgh.harvard.edu/purcell/plink/) (Purcell et al., 2007), and only tests with P < 10 −4 were shown. Phenotypic variation explained by multiple significant SNPs was calculated by stepwise regression after coding one allele as 1 and the other as 0 (Li et al., 2012). Candidate association analysis was also performed with TASSEL; the genotype data of the Nal1 genomic region including Nal1 itself and the upstream 2 kb was extracted from RiceVarMap (http://ricevarmap.ncpgr.cn) (Zhao et al., 2014).

Performance evaluation of the HLS system Performance evaluation of TLA and LN measurements
To test the HLS performance, 163 rice accessions from germplasm resources were randomly selected to measure the GLA, TLA, and LN. After the leaves of an individual rice plant were measured with the HLS, a DS (BenQ 8800) was used to obtain the actual GLA and TLA (Caldas et al., 1992), and the LNs were counted manually. A comparison between the DS measurements and the HLS measurements for GLA are illustrated in Fig. 3a and TLA in Fig. 3b. The manual measurement versus the HLS measurement for LN is shown in Fig. 3c. The mean absolute per cent errors (MAPE, defined by Eq. 1) were 3.20% for GLA, 4.14% for TLA, and 0.95% for LN. The standard deviation of the absolute per cent error values for GLA was 2.51%, for TLA was 2.81%, and for LN was 1.82%. The coefficients of determination (R 2 ) of the DS (or manual) measurements versus HLS measurements for GLA, TLA, and LN were all 0.99.
where n is the number of samples, x i.DS is the i th measured value using the DS method or manual method, and x i.HLS is the i th measured value using the HLS system.

Performance evaluation of leaf colour discrimination
To evaluate the green leaf colour discrimination of the HLS, three batches of rice leaves (50 leaves for each batch) that belong to green-2 (light green), green-3 (green), and green-4 (dark green) were manually selected using the leaf colour chart developed by IRRI. Afterwards, each batch was measured using the HLS, and the results are shown in Fig. 4a and Supplementary Fig. S4. The discrimination errors for green-2, green-3, and green-4 were 0%, 0%, and 2%, respectively, demonstrating that the HLS is highly accurate in discriminating green leaf colours in rice.

Performance evaluation for LL and LW
To evaluate the measuring accuracy of the HLS for LL and LW, the same three batches of rice leaves used for leaf greenness discrimination were measured using HLS and a Vernier caliper (manual method). The results are shown in Fig. 4b, c.
The MAPE values were 1.36% for LL and 2.13% for LW, and the standard deviation of the absolute per cent error values were 1.24% for LL and 1.56% for LW.

Change of the GLA at different growth stages
The GLA shows visible change during the plant life cycle and in response to environmental cues, but the change pattern has seldom been systematically investigated because of measurement difficulties in traditional methods. By taking advantage of HLS, 367 rice accessions at three growth stages (late tillering stage, late booting stage, and milk grain stage) were chosen to analyse the change in the GLA (Fig. 5a-c). The leaf area ratio of the green-3 leaves changed little among the three growth stages, whereas the green-2 leaf area ratio became  larger and the green-4 leaf area ratio became smaller with the growth and development of the rice plant, particularly from the late tillering stage to the late booting stage. The ratios of the three classes of green leaves are known to be related to the nitrogen content of the leaves. To investigate the relationship between the leaf nitrogen content and different green leaf colours, 32 rice accessions were randomly chosen and tested with HLS and with auto discrete analyzers (Smartchem200; Alliance Corporation, France). It was found that the green-2 and green-3 leaf area ratio were negatively correlated with the nitrogen content in leaves, and the green-4 leaf area ratio was positively correlated with the nitrogen content in leaves ( Fig. 5d-f).

GWAS of leaf traits
Using the HLS system, a total of 29 leaf traits were extracted from 533 rice accessions that have been re-sequenced for GWAS . These leaf traits can be classified into three groups as follows: 6 leaf size-related traits, 7 leaf colour-related traits, and 16 leaf shape-related traits (Supplementary Table S1). GWAS was performed on these traits at three growth stages (late tillering stage, late booting stage, and milk grain stage). Using a Bonferroni correction based on the effective numbers of independent markers (Li et al., 2012), the P-value thresholds were set at 1.21 × 10 −6 (suggestive) and 6.03 × 10 −8 (significant) (Supplementary Table S2). In this study, a total of 542 associations (462 lead SNPs) exceeding the suggestive threshold were identified (suggestive SNPs), of which 104 associations (96 significant SNPs) exceeded the significant threshold. For the convenience of presenting of the GWAS results, a chromosomal region in which the distance of adjacent pairs of associated SNPs was less than 300 kb was defined as a locus . As a result, 291 loci containing 462 suggestive SNPs (suggestive loci), and 78 loci containing 96 significant SNPs (significant loci) were detected ( Fig. 6 and Supplementary  Tables S3 and S4). There were some loci with multiple associated SNPs, and loci were repeatedly detected for different traits or repeatedly detected by the same trait at different growth stages. For example, locus 158 was detected with green-4 leaf area ratio (GLAR 4 ) at late booting stage and milk grain stage, green-2 leaf area (GLA 2 ) at late booting stage, and maximum leaf area at milk grain stage, simultaneously. Furthermore, two independently associated SNPs (LD statistics r 2 = 0.09) with GLAR 4 at late booting stage, sf0620856382 and sf0620860929, were located at this locus at only a 4-kb distance. Among the 78 significant loci, 16 loci were repeatedly detected through different leaf traits; one locus was detected through the same trait at different growth stages; and two loci with multiple associated SNPs were detected (Supplementary Table S4). Manhattan plots for all the traits mentioned above are shown in Supplementary Fig.  S5-S7. In total, 131, 62, and 173 suggestive loci and 37, 7, and 38 significant loci were detected at late tillering stage, late booting stage, and milk grain stage, respectively. Only five suggestive loci but no significant loci were detected at the three growth stages simultaneously. These results suggest that the genetic basis of these leaf traits at different growth stages may not be the same in rice (Fig. 7). After adopting the suggestive P-value threshold, a total of nine loci associated with leaf traits contained leaf-related genes that were previously identified (Supplementary Table  S3, Supplementary Note). For leaf colour-related traits, the lead SNPs at loci 81, 130, 238, and 289 that were associated with GLA 2 at late tillering stage are close to OsH43 (a member of the rice kn1-type class 1 homeobox gene family), YGL1 (a leaf-colour related gene), OsHB2 (a member of the rice Class III homeodomain leucine zipper gene family related to the establishment of polarity in developing leaves), and Oshox1 (a member of the rice homeodomain leucine zipper gene family affecting the leaf size and shape), respectively (Itoh et al., 2008;Meijer et al., 1997;Sentoku et al., 2000;Wu et al., 2007).
For leaf shape-related traits, the lead SNPs at loci 81 and 266 that were associated with leaf width standard deviation (LWSD) per plant at milk grain stage are close to OsH43 and NAL3 (a rice WUSCHEL-related homeobox gene affecting rice LW), respectively (Cho et al., 2013;Sentoku et al., 2000). Lead SNPs at locus 109 were associated with LWSD, average leaf length to leaf width ratio (ALWR), average leaf perimeter to leaf area ratio (APAR) at late booting stage, and APAR, maximum LL, and maximum leaf perimeter to leaf area ratio (MPAR) at milk grain stage are close to NAL1, which affects the rice leaf blade width (Qi et al., 2008). Lead SNPs at locus 45 (associated with average LL at late booting stage) and locus 49 (associated with average leaf perimeter at late booting stage) are close to OsIAA3 in controlling the rice LL (Nakamura et al., 2006).
Moreover, rice nitrate transporter gene OsNRT1 (Lin et al., 2000) was detected at locus 240, which was associated with GLA 2 (a leaf colour related trait) at late tillering stage, which in turn partially supports the correlation between leaf colour and leaf nitrogen content mentioned above. In addition, a large number of new loci were identified. After applying the suggestive P-value threshold, a total of 17, 11, and 47 new loci were detected for six leaf size-related traits at late tillering stage, late booting stage, and milk grain stage; a total of 74, 13, and 48 new loci were identified for seven leaf colour-related traits; and a total of 48, 45, and 105 new loci were associated with sixteen leaf shape-related traits at the three stages. For example, new locus 180 (the P-value of a lead SNP sf0719639267 at this locus was 1.94 × 10 −13 ) was repeatedly detected for GLAR 3 , GLAR 4 , green leaf colour classification, LWSD, and MPAR at late tillering stage, average leaf compactness at the late booting stage, and LWSD at the milk grain stage. All the potential candidate genes within 100-kb upstream and downstream of this lead SNP are shown in Supplementary Table S5. These results suggest that the combination of the HLS and GWAS can provide a powerful tool to identify genetic loci associated with leaf traits in rice.
The lead SNP sf0431040266 was associated with ALWR at the late booting stage (P = 6.59 × 10 −8 ). In this locus, Nal1, a gene encoding an unknown protein with pleiotropic function including LW and photosynthesis rate (Qi et al., 2008;Takai et al., 2013), is only 11 kb away from the lead SNP. The information supported that Nal1 was at least partially responsible for the association signal. Candidate association analysis was conducted to demonstrate the association between Nal1 and ALWR and to detect the causal variant. The genotypes of 64 SNPs were extracted and 23 SNPs remained for the following analysis after excluding SNPs with minor allele frequency <0.05 (Supplementary Table S6). Only sf0431027691 (P = 3.77 × 10 −4 ), located in the third exon with non-synonymous mutation (transition from A to G), surpassed the Bonferroni-adjusted P-value (0.05/total markers = 2.17 × 10 −3 ) in a mixed linear model, while the smallest P-value of other SNPs was 3.86 × 10 −2 . This indicates that the SNP was the possible causal variant for this trait. The results indicate that combining GWAS and HLS would be a promising approach to uncover the genetic basis of leaf traits and could provide sufficient resolution to detect causal variants with candidate association analysis.

Measurement efficiency of the HLS
The HLS was designed for the high-throughput measurement of leaf traits, especially for addressing a large population with hundreds and even thousands of plants in a short time. In this study, with two workers feeding the rice leaves into the machine, the HLS requires approximately 5 minutes to process each rice plant (with approximately 100 leaves). With two similar workers, the DS measurement required a total of approximately 21 minutes per rice plant, including 10 minutes for leaf pasting, 10 minutes for leaf paper scanning, and 1 minute for image analysis. Thus, the efficiency of HLS is quadruple the efficiency of the DS method. Additionally, the pasted leaves from the DS method cannot be used for further measurements, such as measuring the fresh or dry weight. When continuously running, the HLS can measure up to 30 leaves per minute.
The pot-grown rice plants were cut up and then directly fed into the HLS. The time-consumption of leaf harvest was about 1-2 minutes depending on the proficiency of the two workers. Thus, taking the time-cost of leaf-harvesting into account, the total screen time for 100 leaves is approximately 6-7 minutes. The throughput of the HLS will be higher with more workers feeding in leaves. When provided with enough labour, increasing the FOV of the camera and the transverse feeding range can further improve the efficiency of HLS; however, increasing the FOV can also decrease the resolution of the inspection system. Thus, there is a trade-off between the measurement efficiency and measurement accuracy of the HLS. When measuring other species with fewer leaves (such as maize), the efficiency of the HLS will be significantly higher. Moreover, to make the HLS fit the different thicknesses of other plant leaves, the feeding interval between the cylinder and conveyor can be easily altered by adjusting a screw (Fig. 1c). Therefore, the system could be applied to other plants, such as maize, barley, and wheat.

Relationships between leaf traits
By using the HLS, it was possible to extract multiple traits related to the leaf area, shape, and size simultaneously from a large population. The resulting large datasets allowed an analysis of the potential relationships between different types of leaf traits, which have seldom been addressed previously, primarily because of limitations in traditional leaf-phenotyping methods. From the leaf traits of 367 rice accessions, the relationship between the average leaf length × average leaf width × leaf number (ALWN) and the GLA per plant at three growth stages was compared ( Supplementary Fig. S8ac). The coefficients of determination (R 2 ) were 0.93, 0.98, and 0.96, indirectly supporting the reliability of the four trait extractions with the HLS. We also compared the relationship between the maximum leaf length × maximum leaf width × leaf number (MLWN) and the GLA per plant at three growth stages ( Supplementary Fig. S8d-f). The R 2 at the three different growth stages (0.83, 0.82, 0.71) showed that the correlation between MLWNs and the GLA were lower than the correlation between ALWN and the GLA. A comparison of the scatter plots demonstrated that the rice leaves of one plant varied in shape. An advantage, therefore, of the HLS is that the GLA can be predicted with sampling, and that the average LW, LL, and LN can be calculated rather than only calculating these traits in the maximum leaf.
With the resulting leaf traits of 367 rice accessions, the APAR was found to be inversely related to the average leaf width during the three growth stages. Because the rice leaf shape is close to a rectangle, APAR could be calculated as indicated by equation 2.

APAR
ALP ALA ALL ALW ALL ALW ALW ALL = = × + × × = + 2 2 2 2 (2) ALP, ALA, ALL, and ALW are the average leaf perimeter, average leaf area, average leaf length, and average leaf width, respectively (as listed in Supplementary Table S1). Because the value of ALL is much greater than that of ALW, equation 2 can be reduced to equation 3, which is consistent with the relationship in Supplementary Fig. S9a-c, also demonstrating the reliability of the leaf shape trait extraction by the HLS system.

Genetic heterogeneity across different sub-populations
Genetic heterogeneity across different sub-populations is a common phenomenon in rice Zhao et al., 2011), and thus a GWAS of the 29 leaf traits in three growth stages of the indica and japonica sub-populations were also performed. In comparison with the GWAS results in the whole population, new association signals were identified in the indica or japonica sub-population for MPAR at late booting stage and ALWR and MPAR at milk grain stage ( Supplementary Fig. S10). For MPAR at milk grain stage, no clear associations were detected in the whole population, but three strong peak-like association signals appeared in the indica sub-population. These results suggest genetic heterogeneity could reduce the power to detect associations in the whole population, and thus it is necessary to perform a GWAS in different sub-populations separately to identify loci associated with leaf traits.

Genetic architecture of different leaf traits
A difference in genetic architecture was observed for different leaf-related traits. For example, genetic variance explained by a total of 24, 28, and 13 suggestive lead SNPs was 64%, 59%, and 41% respectively for GLAR 4 (leaf-colour related), ALWR (leaf-shape related), and ALA (leaf-size related) at milk grain stage (Supplementary Tables S7 and S8). The remaining genetic variance may be explained by the SNPs with small effect that failed to surpass genome-wide P-value thresholds. A higher genotype-environment interaction could contribute to low heritability of GLAR 4 (h 2 = 0.54). For GLA 2 at late tillering stage, 55 suggestive lead SNPs explained up to 80% of phenotypic variance, suggesting the leading role of additive effect in GLA 2 . In addition to additive effect detected by GWAS, significant epistasis interactions of pairwise lead SNPs (P value < 10 −4 ) were also discovered in this study (Supplementary Table S9). For each trait, different associated variants had different effect size with either positive or negative effect. Also for GLA 2 , 7 out of 22 significant lead SNPs had negative effects and the others had positive effects relative to the reference allele (Supplementary Table S4).
To promote the development of leaf phenomics and the dissection of leaf genetic architecture (Micol, 2009), this paper describes an engineering prototype HLS for high-throughput leaf trait phenotyping that can extract 29 leaf traits from rice. In addition to the traditionally defined leaf traits (such as LL, LW, and leaf area), many newly defined leaf traits were extracted by the HLS. By combining a GWAS with HLS, a large number of genetically associated loci that could not be detected by the previous leaf phenotyping method and/or mapping approaches were identified for both the traditional and newly defined leaf traits in this study. The HLS system was originally designed for use in the greenhouse. If portable power is provided, an HLS system with wheels can be used in the field. In summary, by using multidisciplinary technologies, particularly a photonics-based technique , a combination of the HLS and GWAS has provided valuable insights into the genetic basis of rice leaf traits.