Data report: batangas liberica coffee bean morphology image dataset with derived morphometrics (BaLiCoM)

1 Introduction

Open, well-documented datasets are increasingly treated as reusable scholarly outputs because they enable verification, secondary analysis, and downstream method development (Pasquetto, Randles, and Borgman, 2017). The FAIR principles provide a widely used framework for making research data findable, accessible, interoperable, and reusable (Wilkinson et al., 2016). In crop and food informatics, curated image datasets are especially useful because they allow researchers to evaluate segmentation pipelines, feature-extraction workflows, and quality-control algorithms using shared data rather than one-off image collections (Hong, Zhou, and Han, 2025; Niu et al., 2024).

Image-based plant and food phenotyping has expanded rapidly as computer vision and deep learning have become increasingly available for agricultural applications (Hong, Zhou, and Han, 2025). However, the usefulness of image-analysis workflows depends strongly on the availability of well-described image collections with stable identifiers, provenance metadata, transparent preprocessing steps, and validation records. Dataset reports are therefore important not only for making files accessible but also for making their context, reuse boundaries, and limitations clear to future users.

Liberica coffee (Coffea liberica), commonly associated with Kapeng Barako in the Philippines, remains locally important but comparatively underrepresented in open image repositories. Philippine coffee planning documents identify Liberica/Barako as one of the country’s commercially important coffee types. Coffee diversity is also relevant to long-term crop sustainability because wild and cultivated coffee resources are increasingly discussed in relation to conservation, climate stress, and production resilience. The present dataset addresses part of the open-data gap by providing a fixed, versioned archive of controlled single-bean images and derived morphometric descriptors from Batangas, Philippines.

The dataset should be interpreted as a controlled-image morphology resource rather than a comprehensive physical, sensory, or agronomic characterization of Liberica coffee production in Batangas. Its principal contribution is to support reusable image-analysis workflows for beans photographed under consistent laboratory-style conditions. Broader biological, production-level, food-quality, or physical-dimension inferences require additional sampling, calibration, and quality data beyond the image-derived variables reported here.

2 Methods used to collect the data

2.1 Dataset identity, repository deposit, and version control

For this report, the dataset is designated as the Batangas Liberica Coffee Bean Morphology Image Dataset with Derived Morphometrics (BaLiCoM). The fixed archive is deposited in Zenodo as version v1 under DOI 10.5281/zenodo.18917349 (Balazon, 2026). The public landing page is https://zenodo.org/records/18917349, and the archived package is KapengBarako_Datasets-main.zip. Future corrections, additional metadata, or validation files should be released through versioned repository updates so that reuse can be linked to a stable dataset version.

2.2 Sampling design and bean provenance

Beans were obtained from eight participating farms or collection locations in Batangas Province, Philippines: Rosario, San Juan, Padre Garcia, Taysan, Malvar, Sto. Tomas City, Lipa City, and Nasugbu. The dataset was balanced at the farm/location level, with 500 archived bean images per source group for a total of 4,000 records. Samples came from different lots harvested between 7 June 2025, and 20 December 2025. Beans were manually harvested, and visually good, mature, intact beans were selected for imaging. Damaged, defective, immature, broken, or visibly atypical beans were excluded during pre-imaging sorting and were not included in the image archive.

The metadata allows users to subset the dataset by provenance (Table 1). Each record is linked to a stable image identifier and should include farm/location code, municipality or city, lot or batch code, harvest or collection date, imaging date, inclusion status, and quality-control flag. These provenance fields enable users to determine whether downstream analyses use the full balanced archive or a defined subset of farms, lots, dates, or curation groups. The dataset is therefore suitable for controlled reuse in image analysis while allowing users to account for source-location structure during model training, validation, or the generation of descriptive morphometric summaries.

TABLE 1. Recommended metadata fields that allow provenance-based subsetting and reuse.

2.3 Imaging setup and acquisition protocol

Each image contains a single Liberica coffee bean photographed against a matte, non-reflective white background. Images were acquired using the main camera of a Samsung S24 smartphone mounted on a tripod at a fixed distance. Illumination was provided by a diffuse LED ring light to reduce shadows and improve edge contrast. During acquisition, beans were positioned individually within the field of view, and the camera-background setup was kept constant to support reproducible foreground extraction.

The controlled acquisition setup is important because segmentation quality and derived shape descriptors depend on edge contrast, object separation, illumination uniformity, and background consistency. The archive is therefore best suited to evaluating methods under controlled single-object conditions. It should not be treated as a field-imaging dataset with variable backgrounds, occlusions, or uncontrolled illumination.

2.4 Image preprocessing, segmentation, and morphometric extraction

For model input and standardized processing, images were resized to 256 × 256 pixels and normalized to the range [0,1]. Binary bean masks were generated with a U-Net-style segmentation model (Ronneberger, Fischer, and Brox, 2015). The source workflow reports training for 10 epochs using the Adam optimizer with a learning rate of 0.001 and a batch size of 32 on a Google Cloud TPU v3-8 (Kingma and Jimmy, 2014).

Morphological features were extracted from the masks using deterministic image-analysis routines implemented in Python and scikit-image (Walt et al., 2014). Reported descriptors include area, perimeter, eccentricity, solidity, centroid_r, and centroid_c. Supporting tabular processing and validation workflows used standard scientific Python tools, including NumPy, SciPy, pandas, and Matplotlib (Virtanen et al., 2020; Harris et al., 2020; McKinney, 2010; Hunter, 2007). Area and perimeter are pixel-based variables because the archive does not currently describe the use of a physical calibration target. Consequently, these values support within-archive comparison and computational benchmarking but should not be converted to real-world units unless appropriate calibration information is added to the repository. The image-processing and validation workflow used Python, OpenCV, scikit-image, NumPy, SciPy, pandas, Matplotlib, and scikit-learn (Bradski, 2000; Pedregosa et al., 2011; Walt et al., 2014; Harris et al., 2020; Virtanen et al., 2020; Hunter, 2007).

2.5 Quality control, exclusion criteria, and dataset counts

Quality control was applied before imaging, during mask generation, and during dataset inventory checking. Pre-imaging sorting retained visually good beans and excluded damaged, defective, immature, broken, or atypical beans. After acquisition and processing, no captured image was retained as an excluded or failed record in the final archive. The final v1 inventory, therefore, contains exactly 4,000 raw images, 4,000 standardized/resized images, 4,000 masks, 4,000 morphometric records, and 4,000 metadata records. The inventory is summarized in Table 2, and the balanced source-location structure is summarized in Table 3.

TABLE 2. Exact inventory of the archived BaLiCoM dataset.

TABLE 3. Balanced source-location structure and sampling summary.

3 Dataset characteristics and reuse potential

The release combines raw single-bean images, standardized image files, segmentation masks, per-image morphometric descriptors, and supporting metadata or workflow files. Each record is linked via a stable image identifier, allowing images, masks, provenance fields, and tabular outputs to be joined without ambiguity. This structure supports reproducible segmentation benchmarking, precomputation of descriptors from masks, and descriptive summaries of two-dimensional bean shape under the specific imaging conditions represented in the archive.

Three forms of reuse are especially appropriate. First, the images and masks can serve as a controlled benchmark for foreground segmentation under a uniform white-background imaging setup. Second, the morphometrics table enables descriptive comparison of pixel-based two-dimensional shape descriptors across curated subsets, provided users account for provenance and quality-control metadata. Third, the package can support automated quality-control workflows in which masks and derived descriptors are audited together.

The dataset should not be used to infer representative physical dimensions of all Liberica coffee beans from Batangas unless additional calibration and sampling information are supplied. Similarly, potential applications in food-quality assessment, grading, or classification should be framed as future methodological uses rather than as conclusions established by the present Data Report.

Figure 1 summarizes the sampling, curation, processing, validation, and interpretation workflow used to define the dataset as a controlled-image morphology resource.

4 Technical validation

4.1 Validation design and reference workflow

Technical validation compared the archived derived morphometrics with measurements from a separate reference image-processing workflow that was manually checked during dataset generation. A stratified random validation subset of 400 images, representing 10% of the final archive, was manually checked. The subset included 50 images from each of the eight farm/location groups and was selected across available lot and harvest-date metadata, ensuring validation covered all source locations rather than a single subset of the dataset.

Reference masks were generated independently from the archived U-Net outputs using a classical image-processing workflow. Images were converted to grayscale, foreground-background separation was performed using an image-specific Otsu threshold (Otsu, 1979), and the resulting binary masks were refined using connected-component filtering, hole filling, and morphological opening and closing. The largest bean component was retained, and manual boundary correction was applied in ImageJ/Fiji when edge artifacts, small shadows, or thresholding irregularities were observed (Schneider, Rasband, and Eliceiri, 2012). Measurements were recomputed from the final reference masks using Python, NumPy, SciPy, and scikit-image region measurements (Harris et al., 2020; Walt et al., 2014; Virtanen et al., 2020).

Initial manual checking was performed by FB and DA. JM and RS independently reviewed the validation records, and unresolved discrepancies were checked by ME and resolved by consensus. Independent review was therefore used as a consensus quality-control step. Formal inter-rater coefficients were not computed because the final validation records were consensus masks and measurements rather than separate competing annotations. The validation protocol is summarized in Table 4.

TABLE 4. A technical validation protocol is used to document the reference workflow and manual checking procedure.

4.2 Descriptor-level agreement

Across the reported descriptors, agreement between the archived workflow and the manually checked reference workflow was strong overall, with R2 values ranging from 0.932 for perimeter to 0.976 for area. The lower agreement for perimeter is expected because boundary-based features are more sensitive to small segmentation differences than area or centroid position. These results indicate that the archived descriptors are suitable for methodological reuse, descriptive characterization, and quality-control development under the deposit’s-controlled imaging conditions. Table 5 reports the descriptor-level agreement statistics.

TABLE 5. Agreement between archived derived morphometrics and reference measurements.

4.3 Mask-level segmentation validation metrics

Because BaLiCoM includes segmentation masks, descriptor-level agreement alone is not sufficient to characterize mask quality. In addition to MAE, MSE, and R2 for derived morphometrics, segmentation datasets are commonly assessed using pixel-wise overlap and boundary metrics such as Intersection over Union (IoU), Dice coefficient, precision, recall, and a boundary-distance metric like mean symmetric boundary distance or Hausdorff distance (Vlăsceanu et al., 2024). For the 400-image consensus validation subset, archived masks should then be compared with the independently generated reference masks using these metrics.

For each validation image, let denote the archived mask and denote the consensus reference mask. The following metrics should be computed from paired binary masks:

I o U = M a ∩ M r M a ∪ M r

D i c e = 2 M a ∩ M r M a + M r

P r e c i s i o n = T P T P + F P , R e c a l l = T P T P + F N

Boundary error should be computed from the distance between the archived and reference mask contours. Reporting the mean, standard deviation, median, and range for each metric provides users with a direct measure of segmentation reliability and allows them to judge whether the masks are suitable for segmentation benchmarking or downstream descriptor extraction. Table 6 provides the reporting structure for these standard mask-level metrics. Table 6 summarizes the standard mask-level validation metrics recommended for the 400-image consensus reference subset. These values should be computed directly from paired archived and consensus reference masks before the dataset is used as a segmentation benchmark. Because these metrics evaluate pixel-wise mask agreement, they complement the descriptor-level MAE, MSE, and R² statistics reported in Table 5.

TABLE 6. Standard mask-level validation metrics recommended for the 400-image consensus reference subset.

4.4 Interpretation of validation evidence

The validation evidence should be interpreted within the scope of the dataset. The descriptor-level agreement indicates that the archived morphometric variables are internally consistent under the controlled imaging setup, while the proposed mask-level metrics provide the appropriate additional evidence for evaluating segmentation-mask quality. Together, these validation layers support the use of BaLiCoM for controlled-background segmentation benchmarking, mask audit workflows, and reproducible extraction of pixel-based shape descriptors.

The validation does not establish that the dataset represents all Liberica beans from Batangas, nor does it support direct conclusions about physical bean dimensions, sensory properties, market quality, or farm-level production traits. Those applications would require additional sampling, calibration targets, physical measurements, and quality or sensory reference data.

5 Dataset validity, limitations, and reuse considerations

BaLiCoM represents two-dimensional projected morphology under a controlled white-background imaging setup. It is most appropriate for within-archive comparison, segmentation benchmarking, descriptor recomputation, and method-development workflows. It is not designed to characterize three-dimensional bean structure, cup quality, sensory properties, or farm-level production traits.

The most important reuse limitations are sampling scope, calibration, and validation scope. First, although the dataset is balanced across eight Batangas farm/location groups, it should not be generalized to all Liberica coffee production in the province without additional sampling. Second, area and perimeter are pixel-based measurements and should not be treated as physical units without a calibration target. Third, descriptor-level agreement should be supplemented with mask-level validation metrics whenever the dataset is used as a segmentation benchmark.

To avoid overstatement, claims about automated quality assessment, grading, food-quality classification, or comparative physical characterization should be presented as potential future uses rather than as conclusions supported by this Data Report. The present dataset supports image-analysis reuse under controlled conditions; it does not by itself validate broader biological, sensory, commercial, or physical interpretations.

To improve long-term reuse, the repository documentation should retain exact dataset counts, provenance fields, exclusion criteria, split-policy documentation, reference-mask files if released, and validation records. These additions will allow users to evaluate whether the archive is suitable for segmentation benchmarking, descriptor extraction, descriptive morphology, or other downstream applications.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Tatiana Colombo Pimentel, Federal Institute of Education, Science and Technology of Paraná, Brazil

Reviewed by: Diego Fernando Roa-Acosta, University of Cauca, Colombia