Specimens were divided into sub-datasets according to species (L. sinensis or L. dorsotubrosa), sex (male or female), and valve (left or right). The panorama function of a Keyence VHX 7000 microscope was used to photograph all specimens from a sub-dataset at once (bulk images) for faster processing. The microscope was centered, and images were captured in a 4×4 grid around the central position at a 150× magnification. The microscope software automatically generates images with extended depth of focus (EDF). This procedure yielded composite overview images (bulk images) for each sub dataset at each locality (Fig. 2).

These bulk images were subsequently processed using the AutoMorph pipeline (Hsiang et al., 2018) (Fig. 2). Processing began with the segment module, created to segment detected objects of a bulk image. While this module was originally designed for images with multiple focal planes, our dataset consisted of already focus-stacked images, and the script was modified according to Kahanamoku-Meyer et al. (2024). Optimal settings for ostracods were determined and refined using the sample mode in the “control file” of the segment module. Crucial settings for the successful segmentation of ostracod valves are threshold and the size range (minimum size and maximum size). For our microscope settings (e.g., light settings) the most effective segmentation results were achieved with a threshold value between 0.3 and 0.4. Furthermore, defining an appropriate size range for object detection proved to be essential to maximize recognition of individual specimens within each bulk image. This range was species-specific and set to 400 to 1500 µm for L. sinensis and 700 to 2000 µm for L. dorsotubrosa. The script was further adapted to test these settings across all images within each sub-dataset folder. To facilitate this, the saving routine was modified so that the original image file name was used instead of the unique ID from the “control file”. This allowed batch processing of all images belonging to a sub-dataset (e.g., different waterbodies or core section) without the need to manually change the control file (i.e., change the unique ID) for each bulk image or perform segmentation testing separately for each image. As a picking tray was used as the background during imaging, the background was insufficiently dark and sometimes contained scratches, irregularities, or particles (e.g., dust, fiber, sediment), which obstructed outline detection. To account for this, an additional step was implemented in which, after segmentation in the final mode, the image background was set to black, prior to saving each object as a single file. This step was necessary to ensure more accurate outline detection in the subsequent run2dmorph execution (Fig. 3). For each processed image, a “settings file” was generated, providing an overview image that displays all identified objects with assigned numbers and individual segmented images of each object labeled with its corresponding number and associated object-specific information from the control file.

The segmented images were used as input for the run2dmorph module. This module was created to detect object outlines through pixel brightness, extract the outline, measure object parameters (like area, major and minor axes, height, and width), and place a predefined number of points along the outline. Several modifications were implemented in this module to enable subsequent geometric morphometric analyses of ostracod valves. For analyses of valve shape using outlines, it is crucial that the coordinate points are consistently positioned and follow the same directional sequence. The number of coordinate points, describing each specimen's outline, was set to 100 in the settings file. However, the outline orientation differs between left and right valves. For left valves, the starting point was defined at the maximum anterior margin on the left side of the image, proceeding counterclockwise along the dorsal margin. In contrast, for right valves, the maximum anterior margin is located on the right side of the image, and the point sequence runs clockwise along the dorsal margin. In order to account for these orientation differences, two separate run2dmorph scripts were created, one for left valves and one for right valves. The run2dmorph output includes folders containing files for aspect ratio, coordinates, intermediates, and outlines, as well as for the settings used for each run. Additionally, a summary file was generated, listing specimen-specific parameters such as SampleID, ObjectID, area, eccentricity, perimeter, major axis length, minor axis length, rugosity, height, width, and aspect ratio.

The automated height and width measurements of the object correspond to the length and height of the ostracod valves, respectively. These measurements were exported in a CSV output file and directly used from the summary file without further modification. For subsequent analyses of the coordinate data, several adjustments were required. The run2dmorph module produced a CSV file for each successfully segmented specimen, containing the specified quantity (100 in this study) of coordinate points describing the valve outline. However, most geometric morphometric software requires a single file that includes all specimens from a dataset (e.g., a TPS file used in tpsRelw; Rohlf, 2015), meaning that all of the specimen files must be merged into a single file. Additionally, the outline coordinates were further processed as sliding semi-landmarks. For this approach, fixed landmarks are required to provide the reference points between which the semi-landmarks can slide. Therefore, an additional Python script was developed to read the coordinate CSV files for each specimen and to identify the maximum rightmost and leftmost points, which serve as fixed landmarks (LM1 and LM2), depending on the valve side. The 100 outlined coordinate points are listed sequentially, resulting in a set of a total of 102 coordinate points per specimen.

The resulting dataset can be used for a generalized procrustes analysis (GPA) of the coordinates, for instance, in tpsRelw (Rohlf, 2015) or within the R packages geomorph (Adams et al., 2025) or gmShiny (Baken et al., 2021).

Object selection: bulk images of the sub datasets were first processed in the segment module using the sample mode. According to the chosen threshold settings, a series of images were generated in which all detected objects were boxed and numbered. These images were visually inspected to verify which and how many objects were successfully identified under the given settings. If too few objects were detected, the parameters were adjusted accordingly to optimize segmentation performance. Each boxed image was visually checked to ensure that the majority, if not all, specimens were correctly identified and segmented before proceeding to the final processing step.

Outline extraction: the outline extraction of the run2dmorph module generates images displaying the extracted outlines superimposed on the original specimens. This output allows visual verification of the accuracy of outline extraction for each specimen. Due to the additional function implemented in our modified segment module, which replaces the image background with black, outline extraction was consistently successful and did not produce false or incomplete contours for any specimen.

Size and shape measurements: the accuracy of size measurements was validated by comparing automated measurements from AutoMorph with manual measurements performed using the Keyence VHX-7000 microscope and its integrated measurement software. Automated and manual measurements were comparable, with an average deviation of less than 2 % (Table 2). For the shape measurements, the coordinate order and orientation were verified by inspecting the coordinate outputs for the first three specimens of each sub-dataset. The correctness of the semi-landmark configuration and point direction was further confirmed using thin-plate-spline (TPS) deformation grids for visualization, which demonstrated consistency across specimens.

Normality of length and height was assessed for each subset (female/male, LV/RV) using the Shapiro–Wilk's test, and multivariate normality for the combined data was evaluated with the Henze–Zirkler test. To test for significant differences, we applied a permutational multivariate analysis of variance (PerMANOVA) as the normality assumption was not met (L. dorsotuberosa). All analyses were performed in R using the MVN package (version 6.2, Korkmaz et al., 2014) and the vegan package (version 2.8, Oksanen et al., 2026).

We used the R package geomorph (version 4.0.10, Adams et al., 2025) to perform a GPA on the 102 generated coordinate points to align all specimens. The TPS deformation grids were generated to visualize shape changes. Principal component analysis (PCA) was conducted based on the procrustes coordinates to explore the morphospace variability within and among species and sexes. Finally, the coordinates for all specimens are merged into a single tps file. The “sliders file”, used to determine the sliding direction of the semi-landmarks, was created in TpsUtil v. 1.82 (Rohlf, 2015). Procrustes analyses of variance (procrustes ANOVA) and linear discriminant analysis (LDA) were used to test for significant differences between the sub-datasets on the first 15 components.

To identify undersampling or oversampling of coordinate points and to assess the robustness of the morphological characterization, we applied the Landmark Sampling Evaluation Curve (LaSEC) function of the LandMark-Based Data Assessment (LaMBDA) R package (Watanabe, 2018) on the left- and right-valve datasets.

Generally, L. dorsotuberosa valves are larger than those of L. sinensis (Table 1, Fig. 5). Valves of L. dorsotuberosa males are generally shorter and narrower than those of females, while, in L. sinensis, male valves are larger in length and height. Size differences and the separation into males and females are more pronounced for L. dorsotuberosa than for L. sinensis. Left and right valves of both males and females of L. dorsotuberosa are similar in their size distribution, while differences in the left and right valves of both sexes of L. sinensis are more pronounced (Table 1).

According to Shapiro–Wilk's test, all datasets of L. sinensis exhibit a normal distribution, while L. dorsotuberosa shows a non-normal distribution, with exceptions regarding the height of female left valves and the length of female right valves (Table 3a). The Henze–Zirkler test for multivariate normality exhibits normal distribution data for L. sinensis and non-normal distribution data for L. dorsotuberosa (Table 3b). The Bonferroni-corrected p values of the PerMANOVA show that all investigated groups (species, sex, and valve side) are significantly different from each other, except for female left and right valves and male left and right valves of L. dorsotuberosa (Table 4).

The PCA reveals clear distinction between adult specimens of species and sexes in both left and right valves despite minor overlap (Figs. 6 and 7). For left valves (LVs), the first three principal components (PC) explain 51.99 %, 14.77 %, and 7.41 % of shape variation, while right valves (RVs) account for 60.81 %, 13.4 %, and 6.36 %.

Regarding left valves, PC1 primarily differentiates between sexes, with positive values being associated with males and negative values being associated with females, whereas PC2 distinguishes between species, with L. dorsotuberosa at positive scores and L. sinensis at negative scores (Fig. 6a). Thin-plate-spline deformation grids reveal that negative PC1 scores correspond to shorter specimens with steep, straight dorsal margins and concave ventral margins (Fig. 6b). Conversely, positive scores indicate elongated, narrow forms with strongly concave dorsal margins and sigmoidal ventral margins. The comparison with the mean shapes (Fig. 6c) of the sub-datasets confirms that these extremes align most closely with female L. sinensis (negative scores) and male L. dorsotuberosa (positive scores). Along PC2, negative values reflect elongated shapes tapering posteriorly, with straight dorsal and slightly sigmoidal ventral margins, while positive values show broader, rounded anterior regions with steep dorsal and strongly concave ventral margins, corresponding to male L. sinensis and female L. dorsotuberosa, respectively. PC3 primarily captures variations in LM1 positioning, shifting upward (negative) or downward (positive).

Right valves, similarly to left ones, differentiate sexes along PC1 and species along PC2, though this pattern is less distinct due to the centered position of L. dorsotuberosa males (Fig. 7a). PC1 distinguishes short specimens with broad anterior regions, tapering posteriorly, and steep dorsal margins (negative) against strongly elongated specimens with straight dorsal and sigmoidal ventral margins (positive), matching female L. dorsotuberosa and male L. sinensis as endmembers (Fig. 7b and c). PC2 opposes slightly convex dorsal with strongly sigmoidal ventral margins (negative) to broader anterior forms tapering rearward with steep dorsal margins (positive), corresponding to male L. dorsotuberosa and female L. sinensis. As in left valves, PC3 reflects upward or downward shifts in LM1 position.

The LDA based on the first 15 principal components of the full dataset clearly distinguished species and sexes for both valves. For left valves, the first axis separated the species, while the second axis discriminated between the sexes (Fig. 8). The right valves also show clear species- and sex-level separation but with a slightly different pattern: the first axis again differentiates the species, whereas the second axis differentiates only male from female L. sinensis, with both sexes of L. dorsotuberosa showing more similar shapes (Fig. 9). Procrustes ANOVA confirms that group means differ significantly for both valves (Table 5).

The distribution of the data (normal vs. non-normal) is inconsistent among the individual datasets (species, males, females, left and right valves) and traits (length, height). Several factors may contribute to this variability: (1) environmental differences among waterbodies, (2) morphological differentiation between left and right valves, and (3) sex-specific morphological patterns.

In the present dataset, individuals originate from different waterbodies where varying conditions may influence specimen size and lead to the occurrence of distinct morphotypes between and/or within waterbodies, as has been documented for C. torosa (Hoehle et al., 2025a; Wrozyna et al., 2022). Furthermore, in L. sinensis and L. dorsotuberosa, the degree of differentiation between species and sexes differs between left and right valves. Similar observations have been reported in previous studies on various taxa, for example, at the population level in Cytheridella and C. torosa (Hoehle et al., 2025b; Wrozyna et al., 2016, 2022). Although the reasons for these pronounced differences between left and right valves remain unclear, these findings indicate that both left and right valves (and both sexes) should be examined as they capture different morphological patterns. Differences between the sexes can be attributed to differences in development due to the different reproductive roles between the sexes, as is known from other groups of organisms. These can also vary in expression between species.

The spatial distribution of ostracods from the Tibetan Plateau has already been investigated through multi-year sampling campaigns (Akita et al., 2016; Mischke et al., 2007; Peng et al., 2013; Wang et al., 2022; Wrozyna et al., 2009; Zhang et al., 2013), but comparative studies between populations of different lake systems are widely missing so far. With nearly 10 000 water bodies (>0.01 km²) spread across a total area of approximately 2.5 ×106 km² (Chen et al., 2024), the Tibetan Plateau provides ideal conditions for studying and understanding morphological adaptations of species to different environmental conditions at a spatial scale. Previous studies have demonstrated that morphological variability in ostracods can be considerable, both within and particularly between habitats, although the underlying causes, whether genetic, environmental, or attributable to phenotypic plasticity, remain poorly understood. Despite the populations originating from distinct waterbodies and the intragroup variability (at the species and sex level) being substantial, the examined groups can nonetheless be clearly distinguished from one another. Whether the observed intragroup variability reflects differences in size and shape across waterbodies and whether these differences may be attributable to environmental factors warrant investigation in future studies. Such research would greatly advance our understanding of the drivers of morphological variability in ostracods more broadly while also considerably enhancing the applicability of the examined species as bioindicators.

Ongoing projects, such as the International Continental Scientific Drilling Program (ICDP) project NamCore, additionally offer the opportunity to investigate much longer temporal scales (Haberzettl et al., 2019; Adolph et al., 2026). Future studies can therefore offer the opportunity to focus on morphological differences among populations in relation to environmental parameters at a spatial scale, and, furthermore, the long-term evolution of morphological variability in ostracods can be examined using sediment core records.

For manual approaches, placement of points along specimen outlines is extremely time-consuming, necessitating careful consideration of the optimal number of points required to capture shape without oversampling, a decision influenced by dataset size. The LaSEC results confirm that at least 80 coordinate points sufficiently represent shape, which can be easily and quickly achieved using the automated approach. Due to the scarcity of homologous points in our dataset and in non-marine ostracod valves in general, only 2 of the 102 coordinate points were defined as fixed landmarks, with 100 points designated as sliding semi-landmarks. The effect of treating all 102 points as landmarks in LaSEC analyses rather than differentiating between landmarks and semi-landmarks remains unclear (Watanabe, 2018). Given that automated data acquisition incurs no additional time costs with increased point numbers, we retained all 100 sliding semi-landmarks describing the valve outlines.

For most non-marine ostracods, the relatively smooth margins require comparatively few points to adequately capture the outline geometry. In contrast, marine taxa may require substantially more points due to their more complex morphologies. Hunt et al. (2017) placed 256 points, sometimes automatically but mostly manually, to describe the outline of several marine species as taxa such as Pterygocythereis nadeauae (Hill, 1955) possess margins with spines and denticles that greatly complicate outline capture. This further highlights the relevance of automated acquisition methods.

During manual landmark digitization, landmark placement can vary between digitizers (inter-observer error) and between specimens for the same digitizer (intra-observer error) (Fox et al., 2020). Combined with other error sources (specimen placement, imaging device), deviations can exceed 30 % of the total variation in biological datasets (Fox et al., 2020). According to Fox et al. (2020), inter-observer error constitutes the largest contributor, followed by specimen placement and intra-observer error.

Automated coordinate generation along the outline drastically reduces these error sources as the automated output remains consistent across specimens. Consequently, the automated approach is substantially less subjective than manual size measurements or hand-placed landmarks and semi-landmarks, thereby enhancing accuracy and reliability. The modified segmentation module, which eliminates background noise and sharpens object–background boundaries, achieves a 100 % success rate in outline extraction. This is a notable improvement over the original implementation.

AutoMorph implementation yields substantial time savings in image processing, reducing task completion times of weeks or months to within hours, depending on dataset size. Technical validations and manual inspections, conducted during several processing steps (object selection, outline extraction), as well as comparison of automated and manual size measurements, confirm the accuracy and robustness of this method. This dramatic reduction in processing time enables morphometric analyses at scales previously impractical for ostracod research.

A primary limitation of this approach is the requirement for specimens with optimal preservation. In manual methods, small fragments or gaps can be reconstructed or accommodated during landmark placement. In contrast, the automated approach operates with high precision, meaning that even minimally damaged valves result in altered outline extraction and consequent shape distortion. While this represents a constraint for some fossil assemblages, such specimens can be identified through outlier detection prior to generalized procrustes analysis and excluded from subsequent analyses. The trade-off between requiring better-preserved specimens and achieving substantially increased sample sizes through automation generally favors the automated approach as the ability to process large numbers of well-preserved specimens outweighs the loss of moderately damaged individuals from analysis.

An additional consideration concerns the initial specimen preparation. Unlike many other organismal groups amenable to automated morphometrics, ostracods require manual picking from sediments and careful arrangement for imaging. This necessity arises from their asymmetric valve morphology, with distinct right and left valves possessing inner and outer surfaces that must be correctly oriented for taxonomic identification and morphometric analysis. This preparatory step remains time-intensive regardless of subsequent automation.