M. truncatula wild type (Jemalong A17), leaf shape mutants (all in the A17 background) and natural variants were grown in a growth chamber or a greenhouse with 16-h/8-h day/night light cycle; 22°C/20°C day/night temperature. For leaf shape mutant analysis, the fourth fully-expanded leaves from 2-month-old plants were collected. For wild type and natural variant analyses, leaves from successive nodes on the primary inflorescence stem of 6 weeks-old or 45 days-old plants were collected.

Leaves were scanned using an Epson Perfection V700 scanner at a resolution of 800 dpi. Images were saved as TIFF files for subsequent analysis. To avoid shadows along leaf edges and thus increase the accuracy in detection and measurement of leaflets, all leaf samples were scanned against black or gray background. To characterize blade and serration features during leaf development, all six compound leaves developed on successive nodes of a 6 weeks-old wild type plant were detached from the stem, arranged in order, and scanned together to generate a single tiff image. Thirty two plants were used for biological replicates. To compare blade and serration features among four different ecotypes, all seven compound leaves developed on successive nodes of a 45-d-old plant were detached, arranged in order, and scanned as a single tiff image. 20–24 plants were used for biological replicates. To characterize the abnormaty of leaf features in induced mutants, fully expanded compound leaves on the 4th node of nine to twenty 2-month-old plants were detached and scanned to generate multiple tiff images.

LeafletAnalyzer was written in MATLAB (2012–2016) with Image Processing and Visualization toolbox. Thousands of Medicago truncatula leaf samples were imaged and analyzed in this study. Original leaf images are deposited into https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/ZPGVPP; https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/QLXGBG; and https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/29PJR1 for public access. Raw data are listed in Supplementary Files 1–3. Screenshots of the operational steps of LeafletAnalyzer are provided in Supplementary File 4.

Except the first, juvenile leaf, which is simple, all other leaves developed in M. truncatula plants are trifoliate (Figure 1A). Because trichomes developed on leaflet margins, if present in a relatively large number, interfere with detection and measurement by LeafletAnalyzer, we implemented an algorithm to automatically detect and remove trichomes from leaf images. Trichomes are distinguished from leaf blades by differences in their color composition. Trichomes are transparent and grayish white in color, whereas leaf blades have mixed green and red colors. The software first converts the original RGB images to the YCbCr color space. By analyzing the blue color distribution, the software detected and removed trichomes from the leaf images (Figure 1B). Alternatively, our elliptic Fourier analysis showed that trichomes were represented as high frequency spectral components and can be removed by a low-pass space filter. The threshold frequency of the filter was determined empirically to remove trichomes but maintain leaflet border features.

Next, leaflets of a trifoliate leaf, one terminal and two lateral, are separated from the rachis and petiole, respectively, to generate three separate leaflet images (Figure 1C). To separate the two lateral leaflets from the petiole, the boundary and centroid (the center-most point) of the leaf are selected from the binary image by the software. Distances between the centroid and each points on the leaflet boundary (margins) are calculated and plotted as a function of their positions (Supplementary Figure 1). Similarly, angles of the centroid to each points on the leaflet margins relative to the horizontal line are also calculated and plotted as a function of positions (Supplementary Figure 1). At each blade and petiole junctions, the calculated distances and angles reached two overlapping points with local bottom or peak values (called turning points), which represent the left and right borders of the blade and petiole junction, respectively (Supplementary Figure 1). Once two adjacent turning points are identified, the software automatically separates the leaflet blade from the petiole and generates two separate leaflet images (Figure 1C; Supplementary Figure 1). After separating the two lateral leaflets from the original leaf image, this process is repeated one more time to separate the terminal leaflet from the rachis (Figure 1C).

Although distances from the centroid to leaflet serration (tooth) tips and sinus (valley) bottoms are also low, their values are much higher than those of the blade and petiole/rachis junctions. By applying a threshold filter, serration tips and sinus bottoms are excluded from the selection. In addition, users can manually select the junction borders if the software did not correctly detect the turning points. This has been seen in some rare cases.

To detect and calculate the leaflet (blade) length, the midpoint of the blade and petiole/rachis junction of a leaflet is selected as the starting point and the distance between the midpoint and the blade margin is calculated as the followings: di= (xi-x0)2+(yi-y0)2 where (x₀, y₀) is the midpoint of the blade petiole/rachis junction and (x_i, y_i) is any point on the blade margin.

The blade length is calculated as the longest distance: L = maxi{di}

The blade width is calculated as the length of the longest line that is orthogonal to the blade length. To calculate this, a series of orthogonal lines with an interval of five pixels are generated to connect points on the blade margin. The longest line is selected as the blade width and calculated as the followings. wi=(xi-xj)2+(yi-yj)2 if (xi-xj)(yi-yj)×xl-x0yl-y0=1             or-1 The blade width is calculated as the longest distance: W = maxi{wi}

where (x_l, y_l) is any point on the blade length (L), which is used to connect two points on the blade margin (x_i, y_i) and (x_j, y_j), in a line orthogonal to the blade length.

In most cases, the blade length selected by the software overlays perfectly with the midvein (Figure 2A). However, in cases in which leaflets are asymmetric, the selected blade length may deviate from the midvein (Figure 2G). In these cases, end users may manually select three or more points on the midvein, and the software will draw a curve to connect the base and tip of the blade through the selected points and measure the length of the curve as the leaflet length (Figure 2G).

Leaflet serrations at the distal margin are important features of M. truncatula leaves. We observed large variations in leaflet serration characteristics during leaf development in wild type M. truncatula plants, developmental mutants and ecotypes (see below). Because this software uses all points (pixels) on the leaflet border, it is designed to detect and calculate accurately leaflet margin parameters (Figures 2B–F).

LeafletAnalyzer detects leaflet serrations (teeth) and sinuses (valleys) as points on the leaflet margin that have local maximal and minimal distances, respectively, to the centroid of the leaflet. The following formula are used to detect the tooth and valley tip (peak) positions. xi>xi+1>xi+2 and xi>xi-1>xi-2 where x is a vector of distances from the centroid to each points on the blade margin. The ith tooth peak position is detected as above. The ith valley peak position is detected as the following: xi<xi+1<xi+2 and xi<xi-1<xi-2

LeafletAnalyzer calculates and exports in excel files more than 50 parameters for each leaflets, including 16 basic leaflet blade and margin parameters, elliptic Fourier frequency component parameters and some derived parameters (Figures 3A,B).

To measure leaflet tooth parameters, a connection line is made between two neighbor valley peak positions. Tooth number is defined as the number of tooth peaks (Figure 2B). Tooth height is the length of the vertical line from the tooth peak position to the connection line (Figure 2B). Tooth area is calculated as the area surrounded by the real tooth border and the connection line (Figure 2D). Tooth perimeter is the total length of the real tooth border and the connection line. To measure tooth angles, including the tooth tip angle, left and right tooth angles, a triangle is made to connect the tooth tip and two neighbor valley peak positions, and the tooth angles are calculated (Figure 2F). Valley parameters are similarly measured as the tooth parameters, except that a connection line is made between two neighbor tooth peak positions (Figures 2C,E).

Additional parameters are calculated in order to comprehensively analyze tooth and leaflet features (Figure 3). The connection lines between two neighboring valley peak positions are combined as the “Tooth Base Line” (Figure 3A; red line), and the length of the tooth base line is defined as the “Tooth Base Length.” A line connecting the starting and ending points of the tooth base line is used to divide the leaflet blade into the top and bottom parts (Figure 3A; cyan line). The total area of the top part is defined as “Outer Toothed Area.” The area surrounded by the tooth base line and the dividing line is defined as “Inner Toothed Area,” which excludes the tooth area (Figure 3A). The total tooth border line is defined as “Length of Outer Edge” (Figure 3A). The inner toothed area and the bottom part of the leaflet blade are combined as “Inner Blade Area,” and the border of the inner blade area is defined as “Blade Inner Perimeter” (Figure 3A). The leaflet length is used to divide the leaflet blade into the left and right parts. The ratio of the left and right blade area is defined as “Leaflet Symmetry” (Figure 3). “Tooth Base Length” and “Inner Toothed Area” are used to describe the degree of tooth distribution. “Inner Toothed Area Ratio” shows the proportion of the toothed area in total leaflet area (Supplementary Figure 3). “Tooth Number Blade Ratio” indicates the density of teeth relative to the leaflet area. Figure 3B lists the parameters and their definitions used in this study. In contrast to previously-published software, LeafletAnalyzer uses the real shape of teeth to describe serration features.

The elliptic Fourier function is integrated into the software for leaflet shape and size analyses. Directional variations in leaflet blades and margins are converted into an one-dimensional array, and applied to elliptic Fourier transformation (Neto et al., 2006). In total, 120 frequency components of both amplitudes and phases are calculated. The frequency components are normalized to be invariant with rotation, translation and the starting point. Therefore, the elliptic Fourier description is only related to the shape and size of the leaflet blade and margin.

In this study, we identified that the amplitude components of the elliptic Fourier translation are sufficient for the description of the blade and margin features. We combined the amplitude components along the x and y directions as the following: A(f)2=A(f)x2+A(f)y2 where f is the frequency, A(f)_x and A(f)_y are the power spectral A(f) projected in the x and y direction, respectively.

Out of the 120 spectral components, the low frequency components (1–19th) are defined as “Delta” components, whereas the middle frequency components (20–55th) and the high frequency components (56–120th) are defined as “Alpha” and “Beta” components, respectively. “50% Frequency” is defined as the frequency at which the sum of the total power spectral A(f) is divided into two equal halves.

The elliptic Fourier description provides a method to analyze blade and serration characteristics in frequency zones. Firstly, elliptic Fourier analysis allows reconstruction of the actual blade and serration using frequency components. By analyzing more than thousands of different groups of M. truncatula leaflets samples, we show that Delta components mainly correspond to the blade outline, and Alpha components to leaflet teeth characteristics. Secondly, elliptic Fourier description reflects integrated features of blade and serration in a leaflet. For an example, when a leaflet contains different sizes of teeth, especially when small serrations are less than 10% of large serrations, the average tooth height and area will not accurately reflect features of all teeth. Thus, the frequency components of the elliptic Fourier description become important parameters to reflect the variation of tooth features. The frequency components of elliptic Fourier description have been successfully used in the classification of different leaflet groups in this study.

Statistical analysis included in LeafletAnalyzer uses the statistical toolbox of MATLAB. Because most of the blade and margin parameters are not in normal distribution, comparisons of multiple leaf groups are performed using both one way ANOVA and Kruskal-Wallis tests. Data are represented by mean ± s.e. Significant differences analysis is made based on 95% CI (confidence interval). Standard deviation/mean was used to evaluate data variations. A majority of leaflet groups has leaflet blade and serration parameters within mean ± 0.2 standard deviation/mean.

Classification of four or more leaflet groups are made by using the k-means or artificial neural network method. Three or four parameters are selected as a group input to the k-means or artificial neural network programs for classification (Figure 4). The accuracy of classification is defined as the percentage of correct classifications. Combinations of three or four parameters that give rise to at least 85% accuracy are selected.

For classification with artificial neural network (Kim, 2010; Saravanana and Sasithra, 2014; Aboukarima et al., 2015), a feedforward neural network with 3–15 hidden layers is used. All leaf samples are mixed together at first. 75% of the leaf samples are used in training, and the remaining 25% are used in identification. This process is repeated almost 30 times, each time with a different set of training samples. We used a genetic algorithm to facilitate the optimal selection of three or four parameters (Figure 4).

Previously, we have shown that the proximodistal axis development varies in leaves developed on different nodes (Wang et al., 2008; Peng et al., 2011; Ge et al., 2014). Visual examination suggests that leaf shape and size are also variable at different developmental stages (Figure 1A). To quantify differences in shape and size during leaf development, we measured and analyzed blade and serration parameters of trifoliate leaves developed on successive nodes of 6 weeks-old wild type M. truncatula plants, using LeafletAnalyzer. Thirty two biological replicates were used for statistical analyses. For each replicate, all six trifoliate leaves, labeled sequentially by the software as L1 (the youngest measurable leaf from the shoot apex) to L6 (the oldest leaf from the shoot apex), as well as the juvenile leaf (the first leaf emerged on the stem) were included in the original images, although the juvenile leaf was excluded from further analysis (Figure 1A).

Analysis results indicate that terminal and lateral leaflets in L1 were similar in size and shape (Figure 5). From L2 to L6, there were significant differences in size and shape between terminal and lateral leaflets with 95% CI (Figure 5). Terminal leaflets were always larger than lateral leaflets (Figure 5A). Interestingly, terminal leaflets were the largest in L3, and lateral leaflets were the largest in L2 (Figure 5A). The leaflet length was the longest for both terminal and lateral leaflets in L2, and reduced gradually from L3 to L6 (Figure 5B). However, terminal leaflets were always longer than lateral leaflets and the differences were statistically significant with 95% CI. The width of leaflets changed very little from L2 to L4, but reduced in L5 and L6, and the differences were significant with 95% CI between terminal leaflets and lateral leaflets (Figure 5C). The analysis results also indicate that lateral leaflets reduced width more than terminal leaflet in L5 and L6 (Figure 5C). The blade length width ratio was significantly different between terminal and lateral leaflets with 95% CI only in L5 and L6. The blade length width ratio was less than 1 only in terminal leaflets in L6, indicating that only terminal leaflets were wider in shape in L6 (Figure 5D).

Analysis results show that newly emerged leaves had the highest number of serrations (teeth) at the distal margin, even though they had the smallest blade area (Figure 6A). From L1 to L6, total serration number was gradually decreased in a nearly linear fashion in both terminal and lateral leaflets (Figure 6A). In L1, terminal and lateral leaflets had a similar number of teeth; but from L2 to L6, terminal leaflets always had more serrations than lateral leaflets (Figure 6A). The average tooth area increased in general from L1 to L6 (Figure 6B). Interestingly, the total tooth area was increased from L1 to L2 but gradually decreased from L2 to L6 (Figure 6C). The average tooth tip angle was gradually increased from L1 to L4 and decreased from L4 to L6 (Figure 6D). The “tooth base length” was increased from L1 to L2 but decreased gradually from L2 to L6 (Figure 6E). On the other hand, the “tooth valley area ratio” remained similarly low in L1 and L2 but increased steadily from L3 to L6 (Figure 6F). Except L1, terminal leaflets in general had significantly (95% CI) higher “tooth number,” larger “total teeth area,” and longer “tooth base length” but smaller “tooth valley area ratio” than lateral leaflets (Figure 6). The average “tooth tip angle” was significantly different between terminal and lateral leaflet only in L5 and L6. Because the “total teeth area” and “tooth base length” shared a similar trend of changes with leaflet length and leaflet area during leaf development, a large leaflet tended to have larger teeth and a longer “tooth base length.” In summary, LeafletAnalyzer facilitated quantitative analyses of leaflet size and shape and enabled identification of directions and degrees of changes in leaflet size and shape during leaf development in M. truncatula.

Using elliptic Fourier analysis, we evaluated the distribution of a total of 120 spectral components of leaflet parameters from L1 to L6 in 6 weeks-old plants. The original leaflet shape can be precisely reconstructed using the 120 frequency components of the elliptic Fourier transformation. To answer how the spectral components of the elliptic Fourier transformation related to leaflet shape, we reconstruct the leaflet shape by an inverse elliptic Fourier transformation with a subset of the spectral components. When the Delta components were used, only the outlines, but not the margin features, of the original leaflets were reconstructed (Figure 7A). When Delta and Alpha components were used, almost all of the blade and margin features were reconstructed (Figure 7B). Thus, the Alpha components are mostly related to the margin features. When all 120 frequency components were used, the blade and margin features were completely reconstructed (Figure 7C). Thus, the “Beta” components are related to small, local variations of the blade margin.

Analysis results show that “50% frequency” and “Alpha components” decreased continually from L1 to L6 (Figures 8A–C), supporting the observation that the tooth number and tooth height decreased but the “tooth valley area ratio” increased from L1 to L6 (Figure 6F).

Because the direction and degree of changes in leaflet blade and margin parameters are in most cases statistically significant among leaflet groups from L1 to L6, we tested whether these parameters can be used to classify leaflet groups. We used an artificial neural network to classify leaflet groups from L1 to L6, using a combination of three or more blade and margin parameters. Supplementary Table 1 lists ten combinations of three parameters that gave rise to the classification accuracy ranging from 90.4 to 97.9% (Supplementary Table 1). Consistent with the statistical analysis results, both the blade and margin parameters are important for the classification. The classification results suggest that the blade and margin features are statistically distinct in leaflet groups from L1 to L6 and these can be used successfully to classify leaflet groups during leaf development in M. truncatula.

Using LeafletAnalyzer, we calculated and compared blade and margin parameters of trifoliate leaves of three M. truncatula natural variants (PI516927, PI516939, and PI577609) and the reference plant, Jemalong A17. For this, seven to 20 biological replicates were used. For each replicate, seven trifoliate leaves were collected from successive nodes on the stem of 45 days-old plants (Figure 9).

Overall, we observed significant differences in blade and margin parameters as well as the degree of changes from L1 to L7 among the natural variants and the reference plants (Figure 10). For examples, from L1 to L7, PI516927 had the largest blade size, and Jemalong A17 and PI516939 had the smallest blade size (Figure 10A). In a quantitative term, compared with PI516927 (normalized as 1), the area of lateral leaflets ranged between 0.656 and 0.787 in Jemalong A17 and 0.681–0.854 in PI516939 (Figure 10A). From L1 to L6, PI577609 had the second largest blade size. Interestingly, in L7, PI516939 had the largest blade size, PI577609 had the smallest blade area and the other two had similar, intermediate blade size (Figure 10A).

In L1 and L2, Jemalong A17, PI577609 and PI516939 had similar blade length (Figure 10B). From L1 to L6, PI516927 had the longest blade length (Figure 10B). From L3 to L5, PI577609 and PI516939 had the second and third longest blade length and Jemalong A17 had the shortest blade length (Figure 10B). However, in L6 and L7, PI577609 had the shortest blade length, and, in L7, PI516939 had the longest blade length (Figure 10B). From L1 to L6, both PI516927 and PI577609 had the longest blade width (Figure 10C). From L1 to L3, Jemalong A17 had the shortest blade width (Figure 10C). From L4 to L6, PI516939 had the shortest blade width (Figure 10C). In L6 and L7, the longest blade width was found in PI516927 (Figure 10C).

By combining the length and width information, the blade length width ratio represents an important shape parameter. From L2 to L6, the blade length width ratio was in general decreasing in all four genotypes, with an exception that PI516939 had increased blade length width ratios in L6 and L7 (Figure 10D). These results indicate that in general leaflets in newly developed leaves grew preferentially along the proximodistal axis; whereas leaflets in older leaves grew preferentially along the mediolateral axis in these genotypes. However, there were some exceptions. For examples, the blade length width ratio was increased from L1 to L2 in PI516939, from L5 to L7 in PI516939, and in L6 and L7 in Jemalong A17, and the increase only occurred in lateral leaflets (Figure 10D). From L2 to L7, the blade length width ratio decreased almost linearly in PI516927 (Figure 10D).

From L1 to L7, PI577609 had the lowest blade length width ratio (Figure 10D). On the other hand, in L1 and L2, Jemalong A17 had the highest ratio, followed by PI516927 and PI516939. From L3 to L7, PI516939 had the highest ratio, followed by Jemalong A17 and PI516927 (Figure 10D). These results indicate that the degree of changes along the leaf proximodistal and mediolateral axes (reflecting compactness) varies among these genotypes during development.

The total tooth number was either similar in L1 and L2 or slightly reduced in L2 in these four genotypes. From L2 to L7, the tooth number was almost linearly reduced in all four genotypes (Figure 11A). Compared with other genotypes, PI516927 had the highest tooth number from L1 to L6 (Figure 11A). From L1 to L4, PI577609 had the second highest number of teeth (Figure 11A). However, in L5 and L6, the tooth number was the lowest in PI577609 (Figure 11A). Jemalong A17 and PI516939 had similar tooth number from L1 to L7 (Figure 11A). In L7, the number of teeth was very similar among Jemalong A17, PI516927 and PI516939 (Figure 11A). Compared with PI516927, the tooth number from L1 to L6 was reduced by 12–27% in the other three genotypes (Figure 11A).

From L1 to L7, the total tooth area was the highest in PI516927 and lowest in PI516939 (37.4% to 65.7% of that of PI516927), except that PI516927 and PI577609 were similar in the tooth area in L1 (Figure 11B). Jemalong A17 and PI577609 had intermediate tooth area in most of the leaflet groups (79.5–95.4% of that of PI516927), except L1 (Figure 11B).

The relative tooth area (total tooth area/blade area) was the highest in Jemalong A17 (116.9–130.7% of that of PI516927) from L1 to L5) and lowest in PI516939 from L1 to L7, except that PI577609 had similar relative tooth areas as Jemalong A17 in L6 and L7 (Figure 11C). From L1 to L7, the relative tooth areas were intermediate in PI516927 and PI577609, except that for PI577609 in L6 and L7 (Figure 11C).

The following parameters were used to further compare the size and shape of serrations: tip, left and right angles, and height and baseline length of teeth (Figures 11D–H). Jemalong A17 had the shortest tooth baseline (Figure 11D), the largest tooth height (106.7–123.1%) (Figure 11E), and the smallest tooth tip angle (84.9–94% compared with that of PI516927) from L1 to L7 (Figure 11F), indicating that leaflet serrations were taller and sharper from L1 to L7 in Jemalong A17 than other genotypes. On the other hand, PI516939 had the smallest tooth height, ranging from 54.7 to 82.7%, and the tooth area, ranging from 52.6 to 80.8% of that of PI516927 (Figures 11B,E).

Because of statistically significant changes in blade and serration parameters seen among the leaf groups within and among genotypes (Figures 10, 11), we tested whether these parameters can be used to classify leaf groups within or among genotypes, using an artificial neural network method as described earlier (Figure 4). The analysis results show that up to 32 sets of three blade and serration parameters can be used to classify leaf groups within or among the four genotypes with at least 90% of accuracy. The classification was based on parameters of left or right lateral leaflets, terminal leaflets, or leaflets as a group. Supplementary Tables 2, 3 listed ten combinations of parameters for each analysis. Figure 12 shows an example of the classification results for seven leaflet groups, L1 to L7 in 45-days-old wild type M. truncatula (A17) plants. The classification accuracy was calculated as 95.7% in this case.

Next, we visualized the distribution of leaf groups (L1–L7) in a three-dimensional morphospace confined by three blade and serration parameters selected from the classification experiments. Figure 13A shows that the seven leaf groups, L1–L7 of M. truncatula cv. Jemalong A17 plants, occupied distinct space in the 3D morphospace confined by tooth number, blade length and length width ratio. Similarly, the seven leaf groups of the three M. truncatula natural variants, PI516927, PI516939, and PI577609, also occupied distinct space in the 3D morphospace (Figures 13B–D). Interestingly, the space occupied by the seven leaf groups of four genotypes appeared to be different and consistent with the differences in blade and serration parameters of the leaf groups observed among the genotypes (Figures 10, 11).

We measured and compared leaflet blade and serration features of five leaf shape mutants isolated from a collection of M. truncatula fast neutron bombardment (FNB)-induced mutants with the corresponding wild type, M. truncatula cv. Jemalong A17 (Figure 14A). Nine to 20 replicates of mature trifoliate leaves of 2 months-old plants were used. The results show that leaflets of FN10068 (m5) and FN1923 (m2) had the longest and shortest blade length, respectively (Figure 14B), leaflets of wild type (WT) and FN1923 (m2) had the widest and narrowest blade width, respectively (Figure 14C), leaflets of FN10068 (m5) and FN1923 (m2) had the largest and smallest blade length width ratio, respectively (Figure 14D), and leaflets of FN3296 (m1) had smooth margins (no serrations) and leaflets of FN10068 (m5) and FN1923 (m2) had the largest and smallest number of teeth (Figure 14E). Figure 14F shows that the tooth number blade ratio was zero for FN3296 (m1), but the largest and smallest for leaflets of FN1923 (m2) and FN21806 (m3), respectively. Analysis results further show that the blade compactness was the highest and lowest for leaflets of FN1923 (m2) and FN10068 (m5), respectively (Figure 14G), supporting that leaflets of FN1923 are close to a circle whereas leaflets of FN10068 are preferentially expanding along the proximodistal direction (Figure 14A).

We also tested whether the quantitative results can be used to classify leaflet groups in the mutants. All of the leaflet samples from mutants and wild type plants were first mixed and then regrouped using the k-means clustering method with a combination of three or more blade and serration parameters optimized by the genetic algorithm (Figure 4). Supplementary Table 4 lists ten combinations of parameters that were used in the classification. The results show that the classification accuracy ranged from 96 to 100% (Supplementary Table 4; Figures 14H,I).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer DM and handling Editor declared their shared affiliation, and the handling Editor states that the process met the standards of a fair and objective review.