StomataBench: Measuring Taxonomic Generalization in Stomatal Detection

Sainath Reddy Gummi, Hankui K. Zhang, Chulwoo Pack, Shyam Solanki, Young Chang

StomataBench is a collection of datasets for stomatal detection and density estimation under controlled distribution shift. StomataBench pairs a single training set, OD-Train, with four held-out test sets designed to investigate distinct deployment failure modes. OD-Train contains 3,154 microscopy images and 134,802 annotated stomata across 21 woody-plant species, all flowering plants: 20 eudicots and one magnoliid. It has no monocots (such as grasses and cereals), gymnosperms (such as conifers), or ferns. T-MAWP tests taxonomic generalization under the same microscope and annotation protocol. T-DCWP tests generalization across multiple woody-plant species with a different microscope [1]. T-MSWP targets imaging-system and annotation shift within woody plants [2]. T-GSH aggregates published multi-source datasets of nine taxa, from eudicots such as soybean and poplar to monocot crops such as wheat, barley, rice, and maize. Its calibrated wheat subset serves as a monocot extrapolation stress test [3].

This repository holds the images, the COCO annotations, the per-image metadata and the prediction files of the 15 detectors evaluated in the paper. This card describes the data only. It reports no model results.

Paper	https://openreview.net/forum?id=kUdZ3znDlI
DOI	https://doi.org/10.57967/hf/8667
Dataset	https://huggingface.co/datasets/Sainath001/StomataBench (this repository)
Code	https://github.com/gummi001/stomatabench_neurips_eval
Model weights	https://huggingface.co/Sainath001/stomatabench_neurips_weights (DOI https://doi.org/10.57967/hf/11473)
Croissant	croissant.json at the repository root, with the core and Responsible AI fields
Result tables	With the evaluation code at https://github.com/gummi001/stomatabench_neurips_eval
Version	v1.1 on main. The submission files are kept under the git tag v1.0 (see Versions).
Licence	CC-BY-NC-4.0 (see Licence and provenance)

Contents

• Key terms
• Dataset at a glance
• The OOD ladder
• The sets
• Repository layout
• Metadata CSV
• Prediction files
• Loading
• Versions
• Datasheet
• Licence and provenance
• Citation
• References

Key terms

• Stoma (plural stomata). A microscopic pore in the leaf surface, opened and closed by a pair of guard cells, through which the leaf takes up CO₂ and loses water vapour. In StomataBench each stomatal complex (the pore with its guard cells) is one annotated object.
• Stomatal density. The number of stomata per unit leaf area, here stomata per mm². For an image with a physical scale, the ground-truth density is the number of annotated stomata divided by the area of the field of view in mm².
• Calibration. The physical size of one pixel (µm per pixel). It was measured with a stage micrometer for the authors' images and taken from scale bars or from the source authors for the other sets. Only calibrated images give a density. The others are used for counts only.
• Abaxial. The lower surface of the leaf. The images collected by the authors show the abaxial surface.
• Accession. A plant or group of plants in a living collection, recorded under one accession number, for example 403-97*C in OD-Train. The authors' images come from accessions at the Arnold Arboretum of Harvard University.
• Scene-primary. Some images are repeated or near-repeated views of the same field of view (a scene). One image per scene is flagged as primary (is_scene_primary in the metadata CSV), and the paper's evaluation counts only scene-primary images.
• Taxonomic ranks. From broad to narrow: clade, order, family, genus, species. Families, orders and clades follow APG IV [4]. The flowering-plant clades in the data are the eudicots, the magnoliids and the monocots.
• OOD ladder. Out-of-distribution (OOD) levels L0–L5 that place each test species by its taxonomic distance from the training species (see The OOD ladder).

Dataset at a glance

StomataBench comprises one training set (OD-Train), one validation set (OD-Valid), and four held-out test sets (T-MAWP, T-DCWP, T-MSWP, T-GSH). Table 1 and Figure 1 below summarize the splits and taxonomic distance.

Table 1. StomataBench dataset composition and evaluation subsets (scene-primary: one image per scene). Shift from OD-Train: T taxonomic, I imaging system, A annotator.

Dataset	Shift	Taxa	OOD levels	Released images	Released stomata	Scene-primary images	Scene-primary stomata	Calibrated images	Calibrated stomata
Training and validation
OD-Train	–	21	–	3,154	134,802	3,139	134,614	3,139	134,614
OD-Valid	–	5	L0	138	8,792	138	8,792	138	8,792
Held-out test sets
T-MAWP	T	17	L1, L2, L3, L4	456	17,164	456	17,164	456	17,164
T-DCWP	T+I	39	L0, L1, L2, L3, L4	119	2,858	119	2,858	119	2,858
T-MSWP	T+I+A	18	L1, L3, L4	10,715	514,568	10,284	503,898	9,479	421,290
T-GSH	T+I+A	9	L4, L5	8,229	319,161	6,943	288,543	1,930	40,993

Released counts every image in this repository. Scene-primary counts one image per scene, the images used for detection and count evaluation. Calibrated counts the scene-primary images with a physical scale that passed the calibration check, the images used for density evaluation.

Figure 1. StomataBench at a glance. Representative micrographs from the training set (OD-Train), the OD-Valid validation set, and the four test sets, arranged along the OOD ladder (L0–L5) (Table 2).

The OOD ladder

Species in the test sets are arranged on a six-level OOD ladder (L0–L5) ordered by clade divergence from the training distribution [1, 5]. Taxonomic distance provides a meaningful shift axis for stomatal phenotyping because it orders test species by how closely they are related to the training species, and stomatal complex type and epidermal morphology carry taxonomic information. Measuring generalization along this OOD ladder shows, species by species, where detector performance changes, but it does not replace validation in the intended deployment domain.

Species are grouped into OOD levels by the first taxonomic rank at which they differ from the species in OD-Train (Table 2), with families, orders, and clades following APG IV [4]. The levels are ordinal, not a continuous phylogenetic distance, and detection difficulty is not assumed to rise monotonically with every level. A Lauraceae species introduces a new family within a represented order and is L3. Magnolia virginiana introduces order Magnoliales within the represented magnoliid clade and is L4.

Table 2. The six-level OOD ladder (illustrated in Figure 1): the biological holdout and generalization regime of each level, with the held-out test species, scene-primary images and calibrated images at that level.

Level	Biological holdout	Generalization regime	Species	Images	Calibrated
L0†	Familiar species, new accession	No taxonomic shift (control)	1	3	3
L1	New species, familiar genus	Within-genus shift	24	6,455	5,650
L2	New genus, familiar family	Within-family shift	6	60	60
L3	New family, familiar order	Within-order shift	21	783	783
L4	New order, same clade	Within-clade shift	24	4,868	3,671
L5	New clade: monocots	Cross-clade shift	6	5,633	1,817

† L0 also holds the OD-Valid validation set (5 species, 138 images, 138 calibrated). It is not a held-out test and is not counted above.

The level of every image is in the ood_level column of the metadata CSV, next to its family, order and clade, so the assignment can be checked or redone if the taxonomy changes.

The sets

Section S1 of the paper is the datasheet for StomataBench. It describes what is in each set, species by species, so that a reader can see the taxonomic distance behind every OOD level of the paper (its Section 3, and Table 1 and Figure 1 above). Tables S1–S5 list the species of OD-Train and of each held-out test set with their family, order, images and stomata, and Figures S1–S5 show representative micrographs of each set. Table S6 lists the images that enter the count metrics but not the density metrics, either because no calibration exists or because a nominal calibration failed our consistency check. This is also where the T-MSWP density decision is documented. All images are released under CC-BY-NC-4.0.

The tables and figures below are those of Section S1 of the paper, set by set, with the paper's numbering.

OD-Train (training set)

The training set consists of 3,154 microscopy images and 134,802 annotated stomatal complexes across 21 species in 10 botanical families (Table S1, Figure S1). Coverage is concentrated in Rosales, with secondary representation from Ericales, Cornales, Ranunculales, Vitales, Fagales, Sapindales, Dipsacales, and Laurales. Each stomatal complex is annotated as a single-class (stomata) axis-aligned bounding box.

Table S1. OD-Train species composition (21 species across 10 families). Mean density per species ranges from 39.81 to 277.27 stomata mm^-2.

Accession ID(s)	Scientific name	Family	Order	Images (N)	Stomata (N)	GT density (mm-2)
403-97*C & 905-85*A	Actinidia arguta	Actinidiaceae	Ericales	161	4,662	61.61
257-2006*B	Actinidia macrosperma	Actinidiaceae	Ericales	145	5,110	74.99
107-2004*B	Calycanthus 'Venus'	Calycanthaceae	Laurales	55	3,165	122.44
1690-77*B	Clematis terniflora	Ranunculaceae	Ranunculales	308	6,900	47.67
276*A	Cotinus coggygria	Anacardiaceae	Sapindales	56	5,285	200.81
14714-1*A	Hydrangea paniculata 'Praecox'	Hydrangeaceae	Cornales	434	6,306	30.92
427-2016*A	Malus 'Blanche Ames'	Rosaceae	Rosales	134	10,609	168.46
1951-80*A, 1951-80*B, 210-2001*A, 608-2016*A & 12198*B	Malus hupehensis	Rosaceae	Rosales	240	12,063	106.95
28-2013*A	Malus ioensis	Rosaceae	Rosales	49	4,966	215.64
79-2009*B, 51-2009*A, 78-2009*B, 48-2009*A, 49-2009*A, 79-2009*A, 75-2009*A & 74-2009*B	Malus sieversii	Rosaceae	Rosales	204	12,591	131.33
49-98*A, 46-98*A, 228-2013*B & 228-2013*C	Malus spontanea	Rosaceae	Rosales	128	6,724	111.77
80-2017C, 47-98*B & 50-2011B	Malus toringoides	Rosaceae	Rosales	62	4,685	160.78
510-94*A	Malus × zumi	Rosaceae	Rosales	172	7,846	97.06
596-86*A	Prinsepia sinensis	Rosaceae	Rosales	226	6,552	61.69
119-96*A	Prunus 'Washi-no-o'	Rosaceae	Rosales	75	3,712	105.31
13-2016*A	Quercus rubra	Fagaceae	Fagales	116	15,116	277.27
80-2012*A	Rosa 'Queen Elizabeth'	Rosaceae	Rosales	63	1,533	51.78
713-2010*MASS	Rosa 'Seafoam'	Rosaceae	Rosales	162	7,513	98.68
531-34*C	Stewartia gemmata	Theaceae	Ericales	54	1,308	51.54
158-2004*A	Viburnum sargentii 'Onondaga'	Adoxaceae	Dipsacales	55	1,029	39.81
332-84*A	Vitis vinifera × amurensis 'Beihong'	Vitaceae	Vitales	255	7,127	59.47
Total (21 species, 10 families)				3,154	134,802

Figure S1. Representative stomata micrographs of diverse woody plants from OD-Train (300 µm scale bar).

OD-Valid (validation set) and T-MAWP (test set)

OD-Valid. The validation set contains 138 images and 8,792 stomata from five species (Table S2a, Figure S2a). All five species are in OD-Train, but the images come from different accessions (L0).

T-MAWP. The first test set contains 456 images and 17,164 stomata across 17 species (Table S2b, Figure S2b). It uses the same microscope, working distance, magnification, calibration, and annotation protocol as OD-Train and OD-Valid, but includes species-level taxonomic holdouts spanning the OOD ladder used for density evaluation (Table 2).

Table S2. Species-level composition of (a) OD-Valid and (b) T-MAWP, grouped by OOD level relative to OD-Train (Table 2). (a) OD-Valid, the validation set: 138 images and 8,792 stomata from 5 OD-Train species on different accessions (L0). (b) T-MAWP: 456 images and 17,164 stomata across 17 held-out species at L1–L4.

OOD	Accession(s)	Scientific name	Family	Order	Images	Stomata	Mean density (mm-2)
(a) OD-Valid (validation set)
L0	593-87*A	Actinidia arguta	Actinidiaceae	Ericales	37	1,434	82.5
L0	7241*A	Malus hupehensis	Rosaceae	Rosales	41	3,276	170.0
L0	103-2013*A, 216-2013*A	Malus spontanea	Rosaceae	Rosales	40	3,069	163.3
L0	202-2001*A	Malus toringoides	Rosaceae	Rosales	9	671	158.6
L0	331-84*A	Vitis vinifera × amurensis	Vitaceae	Vitales	11	342	66.2
(b) T-MAWP (held-out test set)
L1	248-86*D	Actinidia coriacea	Actinidiaceae	Ericales	37	1,034	59.5
L1	130-2018*A, 130-2018*B	Malus domestica	Rosaceae	Rosales	45	1,532	72.4
L1	457-2010*A	Prunus Tai-Haku	Rosaceae	Rosales	47	1,675	75.8
L1	439-2008*A	Prunus maritima	Rosaceae	Rosales	12	1,177	208.7
L1	931-51*A	Prunus sargentii	Rosaceae	Rosales	18	907	107.2
L1	319-90*F	Prunus subhirtella	Rosaceae	Rosales	20	1,749	186.1
L1	198-2010*A	Rosa Radtkopink	Rosaceae	Rosales	12	222	39.4
L1	327-84*C	Rosa centifolia	Rosaceae	Rosales	10	285	60.6
L1	920-83*A	Rosa chuckles	Rosaceae	Rosales	29	913	67.0
L1	130-2019*B	Vitis empire	Vitaceae	Vitales	9	231	54.6
L1	1739-81*A, 302-89*C	Vitis riparia	Vitaceae	Vitales	33	1,242	80.1
L2	7058*B	Amelanchier sanguinea	Rosaceae	Rosales	32	1,307	86.9
L2	24-98*B	Fagus crenata	Fagaceae	Fagales	17	1,330	166.5
L3	866-50*B	Rhododendron Roseum Superbum	Ericaceae	Ericales	32	1,238	82.3
L4	547-59*MASS	Populus tremula	Salicaceae	Malpighiales	40	706	37.6
L4	170-2017*A	Populus tremuloides	Salicaceae	Malpighiales	23	569	52.6
L4	168-2015*E	Populus trichocarpa	Salicaceae	Malpighiales	40	1,047	55.7

Figure S2. Representative stomata micrographs, one per species, of (a) OD-Valid, the validation set, and (b) T-MAWP, the held-out test set (300 µm scale bar).

Imaging and annotation protocol for OD-Train, OD-Valid and T-MAWP

Images from OD-Train, OD-Valid, and T-MAWP were sampled from living accessions at the Arnold Arboretum of Harvard University, Boston. All images were captured by the authors using a ProScope HR handheld digital microscope (USB 2.0, 400× magnification) connected to a tripod-mounted laptop running ProScope HR v1.2.1. Image resolution is 1280×1024 px, with calibration scale s = 0.59 µm px^-1 measured using a stage micrometer. Imaging was performed on the abaxial leaf surface. The clear contact dome was gently pressed against the leaf to stabilize the field of view, and the integrated LED ring provided coaxial fill illumination. Field sessions were conducted under natural daylight, and images with strong specular highlights from direct solar reflections were excluded during annotation. For hardwood species, the lowest accessible branch was sampled. Visible stomata were annotated as bounding boxes in RectLabel Pro under the supervision of a botanist. No image augmentation was applied before annotation.

T-DCWP (test set)

The second test set contains 119 images and 2,858 stomata across 39 species (Table S3, Figure S3). Source images are publicly available and were captured in Washington, DC, using a high-resolution Olympus BX-63 microscope under a non-destructive imaging protocol [1]. A curated subset was annotated by the authors using the same single-class bounding-box protocol as OD-Train and T-MAWP. Images with calibrated scale bars are used for species-level density evaluation (Table 2).

Table S3. Species-level composition of T-DCWP by OOD level relative to OD-Train (Table 2): images, stomata and per-image mean GT density per species. Most of T-DCWP sits at L3/L4 (74% of images, 82% of stomata, in novel families and orders). Totals: 119 images and 2,858 stomata across 39 species.

Scientific name	Family	Order	Images (N)	Stomata (N)	GT density (mm-2)
L0: Familiar species, new accession
Quercus rubra	Fagaceae	Fagales	3	94	337.84
L0 subtotal (1 species)			3	94	–
L1: New species, familiar genus
Viburnum acerifolium	Adoxaceae	Dipsacales	3	43	154.54
Viburnum dentatum	Adoxaceae	Dipsacales	3	54	194.08
Viburnum nudum	Adoxaceae	Dipsacales	3	26	93.45
Viburnum prunifolium	Adoxaceae	Dipsacales	3	48	172.51
Stewartia malacodendron	Theaceae	Ericales	2	12	64.69
Prunus lasiocarpa	Rosaceae	Rosales	3	43	154.54
L1 subtotal (6 species)			17	226	–
L2: New genus, familiar family
Camellia oleifera	Theaceae	Ericales	3	30	107.82
Amelanchier stolonifera	Rosaceae	Rosales	3	31	111.42
Aronia arbutifolia	Rosaceae	Rosales	3	93	334.25
Crataegus viridis	Rosaceae	Rosales	2	39	210.25
L2 subtotal (4 species)			11	193	–
L3: New family, familiar order
Cornus alba	Cornaceae	Cornales	3	50	179.70
Kalmia latifolia	Ericaceae	Ericales	3	43	154.54
Sideroxylon lycioides	Sapotaceae	Ericales	3	88	316.28
Betula nigra	Betulaceae	Fagales	3	36	129.39
Betula papyrifera	Betulaceae	Fagales	3	30	107.82
Carpinus caroliniana	Betulaceae	Fagales	3	118	424.10
Carya glabra	Juglandaceae	Fagales	3	86	309.09
Morella cerifera	Myricaceae	Fagales	3	108	388.16
Lindera benzoin	Lauraceae	Laurales	3	69	247.99
Persea borbonia	Lauraceae	Laurales	3	153	549.89
Sassafras albidum	Lauraceae	Laurales	3	118	424.10
Zanthoxylum americanum	Rutaceae	Sapindales	3	49	176.11
Zanthoxylum simulans	Rutaceae	Sapindales	3	22	79.07
Acer rubrum	Sapindaceae	Sapindales	3	117	420.50
Aesculus parviflora	Sapindaceae	Sapindales	3	151	542.70
L3 subtotal (15 species)			45	1,238	–
L4: New order, same clade
Fatsia japonica	Araliaceae	Apiales	3	48	172.51
Ilex cornuta	Aquifoliaceae	Aquifoliales	3	92	330.65
Moringa oleifera	Moringaceae	Brassicales	3	52	186.89
Euonymus americanus	Celastraceae	Celastrales	3	63	226.42
Stephanotis floribunda	Apocynaceae	Gentianales	5	195	200.47
Fraxinus pennsylvanica	Oleaceae	Lamiales	3	42	150.95
Olea europaea	Oleaceae	Lamiales	3	83	298.31
Magnolia virginiana	Magnoliaceae	Magnoliales	3	148	531.92
Passiflora coriacea	Passifloraceae	Malpighiales	5	200	198.30
Hibiscus clayi	Malvaceae	Malvales	3	35	125.79
Edgeworthia chrysantha	Thymelaeaceae	Malvales	3	37	132.98
Fothergilla major	Hamamelidaceae	Saxifragales	3	37	132.98
Itea virginica	Iteaceae	Saxifragales	3	75	269.55
L4 subtotal (13 species)			43	1,107	–

Figure S3. Representative stomata micrographs of diverse woody plants from T-DCWP.

T-MSWP (test set)

The third test set, T-MSWP, contains 10,715 images and 514,568 annotated stomata across 18 taxa (Table S4, Figure S4). It primarily evaluates imaging-system and annotation shift because images were acquired using a different protocol from OD-Train, and the annotations were made by the original authors and obtained from their public release, in YOLO TXT format with two classes, stomata and stomata pore. We retained only the stomata class and converted the annotations to COCO format. Source images are publicly available and were captured at Mississippi State [2].

T-MSWP contains both order-overlapping and order-unseen eudicot taxa relative to OD-Train. The Level-4 OOD subset consists of taxa from orders absent from OD-Train: Ilex opaca in Aquifoliales, Fraxinus caroliniana in Lamiales, Populus spp. and Salix spp. in Malpighiales, and Dirca palustris in Malvales. Density uses the per-image calibration supplied by the source authors. X100 images were used for detection and counting only (Table S6).

Table S4. Species-level composition of T-MSWP. Calibration is resolved per image from acquisition magnification and sensor dimensions (four distinct scales among calibrated images, 239–930 px per 0.1 mm). Quercus texana, Q. shumardii, and Populus spp. hold about 80% of annotated stomata. Totals: 10,715 images and 514,568 stomata across 18 taxa. Calibrated (N): released images whose physical scale passed calibration QC (others in Table S6). Density evaluation uses the 9,479 scene-primary images among them.

OOD	Scientific name	scientific_name in the metadata CSV	Family	Order	Images (N)	Stomata (N)	Calibrated (N)
L1	Quercus michauxii	Quercus michauxii Nutt.	Fagaceae	Fagales	129	4,735	129
L1	Quercus nigra	Quercus nigra L.	Fagaceae	Fagales	653	16,783	653
L1	Quercus pagoda	Quercus pagoda Raf.	Fagaceae	Fagales	286	14,704	286
L1	Quercus phellos	Quercus phellos L.	Fagaceae	Fagales	568	34,027	568
L1	Quercus shumardii	Quercus shumardii Buckley	Fagaceae	Fagales	1,612	115,138	1,348
L1	Quercus stellata	Quercus stellata Wangenh.	Fagaceae	Fagales	196	7,868	196
L1	Quercus texana	Quercus texana Buckley	Fagaceae	Fagales	2,975	198,757	2,434
L3	Vaccinium stamineum	Vaccinium stamineum Linneaus	Ericaceae	Ericales	85	1,074	85
L3	Carya ovata	Carya ovata (Mill.) K. Koch	Juglandaceae	Fagales	91	2,097	91
L3	Nyssa sylvatica	Nyssa sylvatica Marshall	Nyssaceae	Cornales	50	500	50
L3	Acer rubrum	Acer rubrum L.	Sapindaceae	Sapindales	100	1,759	100
L3	Ulmus alata	Ulmus alata Michx.	Ulmaceae	Rosales	205	7,374	205
L3	Ulmus americana	Ulmus americana Planch	Ulmaceae	Rosales	185	4,054	185
L4	Ilex opaca	Ilex opaca Aiton	Aquifoliaceae	Aquifoliales	100	3,752	100
L4	Fraxinus caroliniana	Fraxinus caroliniana Miller	Oleaceae	Lamiales	200	3,282	200
L4	Populus spp.	Populus L.	Salicaceae	Malpighiales	3,068	94,666	3,068
L4	Salix spp.	Salix L.	Salicaceae	Malpighiales	82	1,206	82
L4	Dirca palustris	Dirca palustris L.	Thymelaeaceae	Malvales	130	2,792	130
	Total (18 species)				10,715	514,568	9,910

Rows are grouped by OOD level, taken from the ood_level column of the metadata CSV. The third column gives each name as it appears in the metadata CSV, with the source authority.

Figure S4. Representative stomata micrographs of diverse woody plants from T-MSWP.

T-GSH (test set)

The fourth test set contains 8,229 images and 319,161 stomata aggregated from Stomata Hub source datasets (Table S5, Figure S5) [3, 6, 7, 8, 9, 10, 11, 12]. Original annotations were provided in XML and converted to COCO format for evaluation. They were made by the authors of the source datasets, then checked by botanists at Stomata Hub, who applied Contrast Limited Adaptive Histogram Equalization (CLAHE) to reduce inter-source color bias and host the images on their website [3]. T-GSH is the deepest taxonomic-shift slice in the detection benchmark. For density evaluation, only the calibrated images are used (Table 1), and the calibrated Gibbs wheat subset (Triticum aestivum, Table S5) serves as the L5 monocot stress test (Table 2).

Table S5. Source-level composition of T-GSH: eight identified species plus one unresolved Triticum group, from published datasets. T-GSH is the deepest taxonomic-shift slice: about 84% of images and about 82% of stomata are monocots (Poaceae, Arecaceae), with the three LabelStoma species, including Hordeum vulgare, counted individually. Totals: 8,229 images and 319,161 stomata. Calibrated (N): released images whose physical scale passed calibration QC (others in Table S6). Density evaluation uses the 1,930 scene-primary images among them.

OOD	Scientific name	Family	Order	Source	Images (N)	Stomata (N)	Calibrated (N)
L4	Glycine max	Fabaceae	Fabales	Casado-García et al. (2020) [8]	600	33,054	0
L4	Phaseolus vulgaris	Fabaceae	Fabales	Casado-García et al. (2020) [8]	600	19,132	0
L4	Populus balsamifera	Salicaceae	Malpighiales	Fetter et al. (2019), Gibbs et al. (2021) [1, 6]	113	4,225	113
L5	Elaeis guineensis	Arecaceae	Arecales	Kwong et al. (2021) [12]	321	11,787	0
L5	Hordeum vulgare	Poaceae	Poales	Casado-García et al. (2020) [8]	600	15,915	600
L5	Oryza glaberrima	Poaceae	Poales	Cowling et al. (2021) [9]	253	11,931	0
L5	Triticum aestivum	Poaceae	Poales	Gibbs et al. (2021) [6]	349	1,889	349
L5	Triticum aestivum	Poaceae	Poales	Sun et al. (2023) [10]	2,161	42,984	0
L5	Triticum spp.	Poaceae	Poales	Toda et al. (2021) [11]	732	56,556	0
L5	Zea mays	Poaceae	Poales	Liang et al. (2022) [7]	2,500	121,688	2,000
	Total				8,229	319,161	3,062

Rows are grouped by OOD level, taken from the ood_level column of the metadata CSV.

Gibbs wheat. Of the 349 Triticum aestivum images from Gibbs et al. (2021) [6] in Table S5, 348 are scene-primary and calibrated. These 348 images are the Gibbs wheat subset of the paper, the L5 monocot stress test for density.

Figure S5. Representative stomata micrographs from T-GSH.

Images without density values

Table S6. Images excluded from species-level density evaluation. Detection and count evaluation use all 22,811 images. Density evaluation uses 16,839 (73.8%). Exclusions are of two kinds: no calibration exists, or a nominal calibration failed an internal consistency check in which the same species imaged at a different magnification gave a discrepant density. Ground-truth annotations are unaffected in both cases.

Species	Source subset	Set	Images	Stomata	Reason
No calibration available (4,667 images, 175,444 stomata)
Triticum aestivum	StomataTracker [10]	T-GSH	2,161	42,984	No scale bar in image, no published scale
Triticum spp.	Toda et al. [11]	T-GSH	732	56,556	No scale bar in image, no published scale
Phaseolus vulgaris	LabelStoma	T-GSH	600	19,132	No scale bar in image, no published scale
Glycine max	LabelStoma	T-GSH	600	33,054	No scale bar in image, no published scale
Elaeis guineensis	Kwong et al. [12]	T-GSH	321	11,787	No scale bar in image, no published scale
Oryza glaberrima	Cowling et al. [9]	T-GSH	253	11,931	No scale bar in image, no published scale
Calibration failed consistency check (1,305 images, 161,696 stomata)
Quercus texana	X100 group	T-MSWP	541	54,069	Density below the same species at X200/X400
Quercus shumardii	X100 group	T-MSWP	264	28,539	Density below the same species at X200/X400
Zea mays	TipScope [7]	T-GSH	500	79,088	No published scale for this instrument
Total excluded			5,972	337,140

Repository layout

The layout of main (v1.1):

Path	What it holds
datasets/OD-Train/coco.json	COCO annotations: 3,154 images, 134,802 boxes, one category stomata
datasets/OD-Train/images/	3,154 JPEG images
datasets/OD-Valid/coco.json	COCO annotations: 138 images, 8,792 boxes, one category stomata
datasets/OD-Valid/images/	138 JPEG images
datasets/T-MAWP/coco.json	COCO annotations: 456 images, 17,164 boxes, one category stomata
datasets/T-MAWP/images/	456 JPEG images
datasets/T-DCWP/coco.json	COCO annotations: 119 images, 2,858 boxes, one category stomata
datasets/T-DCWP/images/	119 JPEG images
datasets/T-MSWP/coco.json	COCO annotations: 10,715 images, 514,568 boxes, one category stomata
datasets/T-MSWP/images/	10,000 JPEG images
datasets/T-MSWP/images_2/	715 JPEG images
datasets/T-GSH/coco.json	COCO annotations: 8,229 images, 319,161 boxes, one category stomata
datasets/T-GSH/images/	8,229 JPEG images
predictions/<model>/<set>.json	Boxes predicted by each of the 15 detectors on OD-Valid, T-MAWP, T-DCWP, T-MSWP, T-GSH, in the COCO results format (image_id, category_id, bbox, score)
predictions/<model>/<set>_meta.json	Run record of each prediction file (category_id, coco_json, cuda_available, elapsed_seconds, framework, gpu, image_dir, model_id, num_images, out_json, score_threshold, seed, torch_version, and the weights or checkpoint used)
metadata/StomataBench_master_v10.csv	One row per image (22,811 rows), see Metadata CSV
croissant.json	Croissant metadata (core and Responsible AI fields) for this version

• The <model> folders are FRCNN-R101-1x, FRCNN-R50-1x, FRCNN-R50-3x, FRCNN-X101, GDINO-SB, RT-DETRv2, RT-DETRv4, Retina-R101-3x, Retina-R50-1x, Retina-R50-3x, SAM3-OD, YOLOv11, YOLOv12, YOLOv26, YOLOv8. They are the model IDs of the paper with x in place of ×, the same folder names as in the weights repository.
• T-MSWP images are split across images/ (10,000 files) and images_2/ (715 files) because the Hub allows at most 10,000 files per folder.
• Prediction files keep every box down to a score of 0.001 (or a model's higher score floor), as described in the evaluation section of the paper. There are no predictions on OD-Train.
• Annotations are COCO JSON only. The per-image XML files of the submission are under the git tag v1.0.

Image file names. The file_name of each image in a coco.json starts with a set prefix that is not the folder path on the Hub. Base names are unique within a set, so an image is found by its base name under datasets/<set>/images*/ (see the resolver in Loading).

Set	file_name prefix in coco.json	Folders on the Hub
OD-Train	OD-Train/	images/
OD-Valid	OD-Valid/	images/
T-MAWP	T-MAWP/	images/
T-DCWP	T-DCWP/	images/
T-MSWP	T-MSWP/	images/, images_2/
T-GSH	images/	images/

Metadata CSV

metadata/StomataBench_master_v10.csv has one row per image (22,811 rows, 45 columns). It agrees image by image with the coco.json files: coco_image_id is the image id in the set's coco.json, and gt_stomata_count is the number of boxes on that image. The key columns:

Column	Meaning
dataset	Set of the image: OD-Train, OD-Valid, T-MAWP, T-DCWP, T-MSWP, T-GSH
global_image_id	Image id unique over the whole release
coco_image_id	Image id in the set's coco.json (unique within a set)
file_name	file_name of the image in coco.json, with the set prefix
basename	File name of the image under datasets/<set>/images*/
width_px	Image width in pixels
height_px	Image height in pixels
gt_stomata_count	Number of annotated stomata (boxes) on the image
scientific_name	Species name as given by the source
genus	Genus
family	Family (APG IV)
order	Order (APG IV)
clade	Clade: Eudicots, Magnoliids, Monocots
ood_level	OOD level relative to OD-Train: L0, L1, L2, L3, L4, L5. OD-Train rows carry L0
accession_id	Accession number at the Arnold Arboretum, for the authors' images (empty elsewhere)
source_subset	Source collection or published subset, for example wheat_gibbs or labelstomata_bean
scene_id	Scene group: images that show the same field of view share a scene_id
is_scene_primary	True for the one image per scene used in the evaluation
is_repeat	True for a repeated or near-repeated view of a scene
calibration_status	retained (physical scale passed the calibration check, used for density), excluded (nominal calibration failed the consistency check, Table S6), uncalibrated (no calibration exists, Table S6)
calibration_group	Calibration group of the image, for example tmswp_X200_1024 (empty without calibration)
scale_um_per_px	Physical scale in µm per pixel (empty without calibration)
image_area_mm2	Area of the field of view in mm2 (empty without calibration)
gt_density_per_mm2	Ground-truth stomatal density, gt_stomata_count / image_area_mm2 (empty unless calibration_status is retained)
microscope	Microscope, where recorded
leaf_surface	Leaf surface imaged, where recorded
augmentation	Image processing recorded for the image (CLAHE, no), where recorded
annotator	Who made the annotations, where recorded (pending where not yet filled in)
imaging_source	Who captured the images, where recorded (pending where not yet filled in)

Other columns: dataset_id, epithet, species_group, scene_size, scene_primary_coco_id, scene_member_coco_ids, repeat_coco_ids, cultivar, has_density, stomata_density, calibration_note, pixels_per_0_1mm, sampling, imaging_environment, location, country.

The names in scientific_name follow each source and can differ in form from the names in the tables above (cultivar names without quotes, source authorities for T-MSWP).

Prediction files

predictions/<model>/<set>.json holds the boxes that one of the 15 fine-tuned detectors of the paper predicted on one set: OD-Valid and the four test sets T-MAWP, T-DCWP, T-MSWP and T-GSH. Each file comes from one run of the checkpoint with the same folder name in the weights repository (https://huggingface.co/Sainath001/stomatabench_neurips_weights) over every image listed in that set's coco.json. There are no predictions on OD-Train, the training set.

Format. Each file is a JSON list in the COCO results format, one entry per predicted box:

{"image_id": 1, "category_id": 1, "bbox": [729.2819, 259.4064, 48.2784, 55.2472], "score": 0.71337}

• image_id is the image id in the set's coco.json, the same value as coco_image_id in the metadata CSV.
• category_id is the single class stomata.
• bbox is [x, y, width, height] in pixels from the top-left corner of the image, as in COCO.
• score is the detector's confidence.

Score floor. Boxes are kept down to a confidence of 0.001 so that COCO average precision can use the full ranked list, unless a detector releases no lower scores. GDINO-SB and SAM3-OD return no boxes below their own floor, which the table shows as the lowest score in their OD-Valid files.

Model	Framework	Text prompt	Export score floor (meta)	Lowest score in OD-Valid file	Checkpoint hash matches weights repo
FRCNN-R50-1x	Detectron2	none	0.001	0.001	yes
FRCNN-R50-3x	Detectron2	none	0.001	0.001	yes
FRCNN-R101-1x	Detectron2	none	0.001	0.001	yes
FRCNN-X101	Detectron2	none	0.001	0.001	yes
Retina-R50-1x	Detectron2	none	0.001	0.001	yes
Retina-R50-3x	Detectron2	none	0.001	0.001	yes
Retina-R101-3x	Detectron2	none	0.001	0.001	yes
YOLOv8	Ultralytics	none	0.001	0.001	yes
YOLOv11	Ultralytics	none	0.001	0.001	yes
YOLOv12	Ultralytics	none	0.001	0.001	yes
YOLOv26	Ultralytics	none	0.001	0.001	yes
RT-DETRv2	Hugging Face Transformers	none	0.001	0.001	yes
RT-DETRv4	official RT-DETRv4 code	none	0.001	0.001	yes
GDINO-SB	Hugging Face Transformers	"a stomata ."	0.001	0.250	yes
SAM3-OD	official SAM 3 code	"stomata"	0.001	0.500	yes

The checkpoint hash that each run recorded is compared with the file in the weights repository in the last column.

How the paper uses these files. The paper reports no result in this card, but it computes every result from these files in one of two ways.

• COCO average precision and recall use every box in a file (with pycocotools), on the scene-primary images (is_scene_primary in the metadata CSV).
• Counts, true and false positives and density use only the boxes whose score is at or above a confidence threshold. The main threshold is a fixed 0.25. The paper also reports the shared threshold and each detector's own threshold, both selected on OD-Valid (Table S8 of the paper), and the fixed thresholds 0.50 and 0.75.
• The predicted count of an image is its number of boxes at or above the threshold. The predicted density divides that count by the image's field-of-view area (image_area_mm2), on calibrated images only (calibration_status is retained).

Run records. predictions/<model>/<set>_meta.json records how each file was made. The fields depend on the framework.

• Detectron2 (detectron2): category_id, coco_json, config, cuda_available, device, elapsed_seconds, framework, gpu, image_dir, model_id, num_images, num_predictions, out_json, score_threshold, seed, torch_version, weights, weights_sha256
• official RT-DETRv4 code (rtdetrv4_custom): category_id, checkpoint, checkpoint_sha256, coco_json, config, cuda_available, device, elapsed_seconds, framework, gpu, image_dir, image_size, max_det, model_id, num_images, num_raw_predictions, num_saved_predictions, out_json, repo_dir, score_threshold, seed, torch_version
• official SAM 3 code (sam3_custom): base_weights, base_weights_sha256, bpe_path, category_id, coco_json, cuda_available, device, elapsed_seconds, finetuned_checkpoint, finetuned_checkpoint_sha256, framework, gpu, image_dir, max_det, model_id, num_images, num_raw_predictions, num_saved_predictions, out_json, score_threshold, seed, text_prompt, torch_version
• Hugging Face Transformers (transformers): category_id, checkpoint_sha256, coco_json, cuda_available, device, elapsed_seconds, framework, gpu, image_dir, max_det, model_dir, model_id, model_kind, num_images, num_raw_predictions, num_saved_predictions, out_json, processor_id, score_threshold, seed, text_prompt, text_threshold, torch_version
• Ultralytics (ultralytics): category_id, coco_json, cuda_available, elapsed_seconds, framework, gpu, image_dir, imgsz, max_det, model_id, nms_iou, num_images, num_predictions, out_json, score_threshold, seed, torch_version, weights, weights_sha256

The paths in these records (coco_json, image_dir, weights, checkpoint, out_json) are where the run took place on the authors' computing cluster. They are kept as provenance and do not point into this repository. Use the *_sha256 fields to match a run to its checkpoint.

Example. Counts and density at the main threshold for one detector and set, then COCO average precision:

import json
from collections import Counter

import pandas as pd
from huggingface_hub import hf_hub_download
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval

repo = "Sainath001/StomataBench"
get = lambda path: hf_hub_download(repo, path, repo_type="dataset")
model, subset, tau = "YOLOv8", "T-MAWP", 0.25

preds = json.load(open(get(f"predictions/{model}/{subset}.json")))
meta = pd.read_csv(get("metadata/StomataBench_master_v10.csv"), low_memory=False)
rows = meta[meta["dataset"].eq(subset)]

counts = Counter(p["image_id"] for p in preds if p["score"] >= tau)
calibrated = rows[rows["is_scene_primary"] & rows["calibration_status"].eq("retained")].copy()
calibrated["pred_count"] = calibrated["coco_image_id"].map(counts).fillna(0)
calibrated["pred_density"] = calibrated["pred_count"] / calibrated["image_area_mm2"]

gt = COCO(get(f"datasets/{subset}/coco.json"))
gt.dataset.setdefault("info", {})  # recent pycocotools versions expect this key
ev = COCOeval(gt, gt.loadRes(get(f"predictions/{model}/{subset}.json")), "bbox")
ev.params.imgIds = sorted(rows.loc[rows["is_scene_primary"], "coco_image_id"])
ev.evaluate()
ev.accumulate()
ev.summarize()

Loading

Download one set and the metadata with allow_patterns, then open the COCO file with any COCO reader (for example json or pycocotools).

import json
from pathlib import Path

import pandas as pd
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Sainath001/StomataBench",
    repo_type="dataset",
    allow_patterns=["datasets/T-MAWP/*", "metadata/StomataBench_master_v10.csv"],
)


def image_path(root, set_name, file_name):
    """Find an image by its base name under datasets/<set>/images*/. Base names are unique within a set."""
    name = Path(file_name).name
    for folder in sorted((Path(root) / "datasets" / set_name).glob("images*")):
        if (folder / name).is_file():
            return folder / name
    raise FileNotFoundError(f"{set_name}: {file_name}")


set_name = "T-MAWP"  # the resolver works for every set, also T-MSWP with images/ and images_2/
coco = json.loads((Path(root) / "datasets" / set_name / "coco.json").read_text(encoding="utf-8"))
paths = {im["id"]: image_path(root, set_name, im["file_name"]) for im in coco["images"]}

meta = pd.read_csv(Path(root) / "metadata" / "StomataBench_master_v10.csv", low_memory=False)
rows = meta[meta["dataset"] == set_name]
scene_primary = rows[rows["is_scene_primary"]]                              # detection and count scope
calibrated = scene_primary[scene_primary["calibration_status"] == "retained"]  # density scope

To load the submission files instead, pass revision="v1.0".

Versions

• v1.1 is on main and is described by this card.
• v1.0 is kept under the git tag v1.0. It holds the submission files: the per-image XML annotations, the submission-era results and scripts, and the model weights before they moved to https://huggingface.co/Sainath001/stomatabench_neurips_weights.

Datasheet

Motivation

• Purpose. Stomatal density is widely measured in plant biology because it links cellular anatomy to photosynthesis, transpiration, water-use efficiency, and global change responses across Plantae. Most existing stomatal datasets evaluate detectors within a single acquisition pipeline, leaving open whether their predictions remain reliable under out-of-distribution (OOD) deployment. StomataBench evaluates stomatal detection as a scientific measurement problem under structured OOD shift. StomataBench provides a benchmark for testing whether stomatal detectors remain useful as measurement tools under deployment-relevant biological and imaging shifts.
• Creators. The authors listed at the top of this card.
• Funding and acknowledgements. This work was supported by the Hatch-Multistate project (SD00R730-23) through the SDAES and in part by SDSU's HPC resources. We acknowledge the Arnold Arboretum of Harvard University for providing access to living collections and financial support through a Sinnott Award.

Composition

• Instances. Each instance is a microscopy image of a leaf surface. Each stomatal complex is annotated as an axis-aligned bounding box of a single class, stomata. There are no other classes.
• Size. 22,811 images and 997,345 annotated stomata over all sets (Table 1).
• Taxonomic scope. OD-Train contains 21 woody species, 20 eudicots and one magnoliid, and no monocot, gymnosperm, or fern.
• Sensitive data. No crowdsourcing or human-subject research was conducted. The observations and annotations concern plant microscopy images.

Collection and annotation

Set	Images	Annotations	Source
OD-Train, OD-Valid, T-MAWP	Captured by the authors at the Arnold Arboretum of Harvard University	Annotated by the authors in RectLabel Pro under the supervision of a botanist	This paper
T-DCWP	Public images captured in Washington, DC, with an Olympus BX-63 microscope	A curated subset annotated by the authors with the OD-Train protocol	Fetter et al. [1]
T-MSWP	Public images captured at Mississippi State	Made by the original authors (YOLO TXT), stomata class kept, converted to COCO	Wang et al. [2]
T-GSH	Stomata Hub source datasets	Made by the source authors (XML), checked by botanists at Stomata Hub, converted to COCO	See Table S5 [3, 7, 8, 9, 10, 11, 12]

Preprocessing and labelling

• Annotations are COCO JSON only, with one class. No image augmentation was applied before annotation of the authors' images.
• T-MSWP annotations came with two classes, stomata and stomata pore. Only stomata is kept.
• For T-GSH, botanists at Stomata Hub applied CLAHE (Contrast Limited Adaptive Histogram Equalization) to reduce colour bias between sources [3].
• No image was removed for being a repeated view. Repeated views stay in the release and are marked by scene_id and is_scene_primary.
• Density is computed only where the physical image area is known and passed the calibration check (Table S6). A missing scale is left empty, not estimated.

Uses

• Intended use. StomataBench is a benchmark for testing whether stomatal detectors remain reliable under taxonomic, imaging-system, annotator, and multi-source shift. In the paper, all pretrained checkpoints were fine-tuned end-to-end on OD-Train. The paper's validation thresholds are selected on OD-Valid and frozen before any test set is used.
• Out of scope. Training or tuning on a test set and then reporting it as held out. Reading the density results as measurements of stomatal aperture, conductance or physiology, which the data do not contain. Reading the L5 results as a full estimate of monocot generalization.
• Risks. Biased density estimates can mis-rank genotypes or distort drought-screening decisions. Benchmark performance is not a deployment guarantee. Target-domain calibration, confidence selection on independent validation data, and human review remain necessary.

Limitations

• Field of view. Density estimation is constrained by the physical field of view. Larger calibrated leaf areas could provide more stable density estimates, but a wider field of view lowers the optical resolution needed to detect individual stomata [7]. At the same time, OD-Train, OD-Valid, and T-MAWP were collected from intact leaves under field conditions, making the benchmark relevant to scalable tree breeding and field phenotyping workflows.
• Taxonomic scope. OD-Train is taxonomically broad within woody eudicots (plus one magnoliid) but does not include monocots, gymnosperms, ferns, or basal angiosperms (the ANA grade). These are priorities for future expansion. Species names follow each source dataset, so a few are older synonyms. The L5 results, from the T-GSH monocots (wheat, barley, rice, maize, and oil palm), should therefore be interpreted as a stress test for eudicot-to-monocot extrapolation, not as a complete estimate of monocot generalization.
• Low-replication species. Several T-MAWP and T-DCWP species have small image-level sample sizes or low within-species density variance. Low-replication species are retained because they reflect a realistic biological-data regime rather than a perfectly balanced machine learning benchmark.
• Single-class boxes. StomataBench currently uses single-class bounding-box annotations for stomatal complexes. It does not distinguish guard cells from pavement cells, annotate pores, estimate aperture state, or measure pore-level geometry. Extending the benchmark to keypoint detection, instance segmentation, and pore-level labels would support a broader range of stomatal phenotyping tasks [7].
• No physiological metadata. Images were also collected without controlled CO₂, humidity, water status, light history, developmental stage, or time-of-day metadata [10]. StomataBench is therefore a multidisciplinary benchmark for detection generalization and density measurement under distribution shift, not a controlled physiological experiment.

Maintenance

• Corrections are released as new versions of this repository. v1.0 stays available under its git tag.
• Maintainer: Sainath Reddy Gummi (gummisainath@gmail.com).

Licence and provenance

All images and annotations are released under CC-BY-NC-4.0. The T-DCWP, T-MSWP and T-GSH images come from the published sources below. Please cite them together with StomataBench.

Set	Images	Reference	Link
T-DCWP	Images captured in Washington, DC	Fetter et al. (2019) [1]	https://doi.org/10.1111/nph.15892
T-MSWP	Images captured at Mississippi State	Wang et al. (2024) [2]	https://doi.org/10.1038/s41597-023-02657-3
T-GSH	Elaeis guineensis	Kwong et al. (2021) [12]	https://doi.org/10.1038/s41598-021-94705-4
T-GSH	Glycine max, Phaseolus vulgaris, Hordeum vulgare	Casado-García et al. (2020) [8]	https://doi.org/10.1016/j.compag.2020.105751
T-GSH	Oryza glaberrima	Cowling et al. (2021) [9]	https://doi.org/10.1101/2021.12.01.469618
T-GSH	Populus balsamifera	Fetter et al. (2019) [1]	https://doi.org/10.1111/nph.15892
T-GSH	Populus balsamifera, Triticum aestivum	Gibbs et al. (2021) [6]	https://doi.org/10.3389/fpls.2021.780180
T-GSH	Triticum aestivum	Sun et al. (2023) [10]	https://doi.org/10.1016/j.compag.2023.108120
T-GSH	Triticum spp.	Toda et al. (2021) [11]	https://doi.org/10.3389/fpls.2021.715309
T-GSH	Zea mays	Liang et al. (2022) [7]	https://doi.org/10.1111/pbi.13741

Citation

@inproceedings{gummi2026stomatabench,
  title     = {StomataBench: Measuring Taxonomic Generalization in Stomatal Detection},
  author    = {Gummi, Sainath Reddy and Zhang, Hankui K. and Pack, Chulwoo and Solanki, Shyam and Chang, Young},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2026}
}

References

1. Fetter et al. (2019). StomataCounter: a neural network for automatic stomata identification and counting. New Phytologist. https://doi.org/10.1111/nph.15892
2. Wang, Renninger, Ma (2024). Labeled temperate hardwood tree stomatal image datasets from seven taxa of Populus and 17 hardwood species. Scientific Data. https://doi.org/10.1038/s41597-023-02657-3
3. Gibbs, Burgess (2024). Application of deep learning for the analysis of stomata: a review of current methods and future directions. Journal of Experimental Botany. https://doi.org/10.1093/jxb/erae207
4. The Angiosperm Phylogeny Group et al. (2016). An update of the Angiosperm Phylogeny Group classification for the orders and families of flowering plants: APG IV. Botanical Journal of the Linnean Society. https://doi.org/10.1111/boj.12385
5. Magallón et al. (2015). A metacalibrated time-tree documents the early rise of flowering plant phylogenetic diversity. New Phytologist. https://doi.org/10.1111/nph.13264
6. Gibbs et al. (2021). A Deep Learning Method for Fully Automatic Stomatal Morphometry and Maximal Conductance Estimation. Frontiers in Plant Science. https://doi.org/10.3389/fpls.2021.780180
7. Liang et al. (2022). StomataScorer: a portable and high-throughput leaf stomata trait scorer combined with deep learning and an improved CV model. Plant Biotechnology Journal. https://doi.org/10.1111/pbi.13741
8. Casado-García et al. (2020). LabelStoma: A tool for stomata detection based on the YOLO algorithm. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2020.105751
9. Cowling et al. (2021). Stomata Detector: High-throughput automation of stomata counting in a population of African rice (Oryza glaberrima) using transfer learning. bioRxiv. https://doi.org/10.1101/2021.12.01.469618
10. Sun et al. (2023). StomataTracker: Revealing circadian rhythms of wheat stomata with in-situ video and deep learning. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2023.108120
11. Toda et al. (2021). An affordable image-analysis platform to accelerate stomatal phenotyping during microscopic observation. Frontiers in Plant Science. https://doi.org/10.3389/fpls.2021.715309
12. Kwong et al. (2021). Automated stomata detection in oil palm with convolutional neural network. Scientific Reports. https://doi.org/10.1038/s41598-021-94705-4