Source-linked AI summary
A Public Image Database for Benchmark of Plant Seedling Classification Algorithms
Thomas Mosgaard Giselsson, Rasmus Nyholm Jørgensen, Peter Kryger Jensen, Mads Dyrmann, Henrik Skov Midtiby
TL;DR
Plant-species recognition lacks benchmark databases that support robust comparison across varying appearances. The paper presents a public seedling-image dataset and proposes a weighted, stratified cross-validation f1 benchmark. The released collection contains 407 images of 12 species, and the benchmark is intended to make classification results easier to compare.
Problem
The authors identify a lack of benchmark databases as a barrier to comparing plant-species recognition methods that must handle substantial variation in plant appearance.
Source: §1 · pp. 1–2Method
The paper presents a public seedling-image database and proposes weighted averages of class-specific f1 scores across stratified cross-validation folds for classification benchmarking.
Source: §1 · p. 1, §1.8 · pp. 5–6Results
407 images of 12 species common in Danish agriculture are made publicly available as a database for species-recognition research.
Source: §3 · p. 9Takeaways & Limitations
The shared dataset and benchmark are intended to support replication and comparison of classification results, including research on plant recognition at early growth stages.
Source: §1 · p. 1, §1.9 · p. 7The image resolution is not practically achievable on field machinery, so downsampling and added noise should be considered to approximate practical recording conditions.
Source: §1.9 · p. 7
Abstract
from arXiv · showhide
A database of images of approximately 960 unique plants belonging to 12 species at several growth stages is made publicly available. It comprises annotated RGB images with a physical resolution of roughly 10 pixels per mm. To standardise the evaluation of classification results obtained with the database, a benchmark based on $f_{1}$ scores is proposed. The dataset is available at https://vision.eng.au.dk/plant-seedlings-dataset
1. Introduction
The paper addresses the lack of usable benchmark databases for plant-species recognition, which makes comparisons difficult despite promising prototypes. It presents a public dataset and benchmark to support replicated evaluation of classification approaches amid substantial variation in plant appearance.
1. Introduction
- The authors identify the lack of benchmark databases as a barrier to robust plant-species recognition and easy comparison of classification methods.They present a freely available database and performance benchmark to support replication and comparison.Source: §1 · p. 1
- Existing botanical and plant-image databases vary in availability and content, and herbarium collections may not capture representative within-species variation.The authors also note that a previously described public database could not be located and that the usability of many databases for machine-vision research was uncertain.Source: §1 · pp. 1–2
- Generic species recognition must handle large variation in plant appearance, and the dataset targets common Danish agricultural weeds to help researchers study that variation.The authors view handling this dataset as a step toward automatic identification for applications such as site-specific weed management.Source: §1 · p. 2
- The database is aimed at ground-based weed or species spotting, an approach researchers have argued is promising for site-specific weed control.Remote sensing must address conditions such as solar angle and cloud cover that users cannot control.Source: §1 · p. 2
1.1. Image recording
The paper motivates a shared seedling-image database because existing data differ in quality, quantity, species, and availability, making method comparisons difficult. It describes the recording plan and species selection, whose choices necessarily bound the database’s generality.
1.1. Image recording
- A public database of plant-seedling images lets researchers test methods on common data and compare results where diverse, sparse, or unavailable datasets make comparisons difficult.The authors describe data recorded with different equipment and varying quality, quantities, and species.Source: §1.1 · p. 2
- Database design choices about species, timing, sample counts, equipment, and illumination constrain generality but make collection achievable.The authors expect thorough documentation of acquisition choices to add value for other researchers.Source: §1.1 · p. 2
- The study planned four boxes per species to obtain about 80 plants at the same growth stage while allowing for 20% germination failure.Each box held about 25 samples, and each contained one species.Source: §1.1 · p. 2
- RGB images were recorded repeatedly over 20 days, every two to three days, beginning a few days after emergence, using a Canon 600D with a fixed 50 mm lens.The early-stage focus supports assessment before weeds compete with crops for nutrition.Source: §1.1 · p. 2
- The recording setup produced images at approximately 10 pixels per millimeter from a native camera resolution of 5184 × 3456 pixels.The camera was positioned approximately 110–115 cm above the soil.Source: §1.1 · p. 3
1.2. Specie selection
- Fourteen important Danish arable-field weed and crop species were selected, but two failed to germinate, leaving 12 recorded species.The selected set included monocotyledon and dicotyledon crop and weed species.Source: §1.2 · p. 3
1.3. Image recording
The recording period targets early growth, when plants are less likely to overlap and are relevant to timely weed control. The acquisition setup also addresses a source of segmentation error from moss growing on bare soil in moist indoor conditions.
1.3. Image recording
- Early recording targets the period before weeds compete with crops and when overlapping leaves are minimal, making individual plants more manageable for species categorization.The paper notes that weed assessment or control is pursued within roughly the first 20 days after crop germination.Source: §1.3 · p. 3
1.4. Acquisition setup
- Small stones cover the soil surface because bare soil in moist indoor conditions can develop green moss that may mislead pixel-based segmentation.Plants were grown in soil in styrofoam boxes at the AU-AgroEcology research facility.Source: §1.4 · p. 3
1.5. Tray and specie tracking
The tracking setup uses single-species boxes with four boxes per species to support sample identification and reduce the influence of box-specific conditions.
1.5. Tray and specie tracking
- Each box contains one labeled species, while four boxes per species reduce the risk that an unforeseen condition in one box misrepresents that species.The repeated boxes also help achieve the intended sample count, and Table 1 lists sample IDs.Source: §1.5 · p. 3
1.6. Recording considerations
A fixed overhead camera rig with diffused flash provides controlled image acquisition. Camera geometry was chosen to provide roughly 10 pixels per mm while keeping plants within the calculated depth of field.
Rig and lighting
- The ITEM rig fixes the camera 1100–1150 mm above the soil, and a diffuser in front of the built-in flash helps avoid hard shadows.The rig's rough white rigid sheets scatter light, while its sides minimize point-like light-source artifacts.Source: §1.6 · p. 3
Rig and capture setup
- An external trigger adds a 2-second delay before image capture to let vibration diminish.The camera is mounted facing downward and uses a fixed 50 mm lens set to autofocus.Source: §1.6 · p. 3, Figure 1 · p. 4
Depth of field
- At camera distances of 1100 mm and 1150 mm, the calculated depth-of-field spans are 125.09 mm and 137.05 mm, respectively.Plants were expected to grow no higher than 100 mm, placing them within the stated focus range.Source: §1.6 · p. 5
Resolution
- 10.6 pixels per mm at 1.1 m and 10.1 pixels per mm at 1.15 m are the calculated image resolutions for the stated object distances.The camera uses a fixed 50 mm lens and a 22.3 by 14.9 mm sensor.Source: §1.6 · p. 5
1.7. Recording procedure
Boxes were photographed individually every two or three days after germination for three weeks or until their plants formed a green cover. A sample image and extracted close-up illustrate the recorded data.
Recording procedure
- Each box was photographed every 2 or 3 days for 3 weeks after germination, or until it consisted of a green cover.Boxes were individually placed in the camera rig; camera settings were ISO 100, 1/60-second shutter speed, and f7.1 aperture.Source: §1.7 · p. 5
Example data
- Figure 2 presents one random sample drawn from the recorded image set and pairs it with a close-up view of the marked rectangle.Source: §1.7 · p. 5
1.8. Proposed benchmark measure
The proposed benchmark uses stratified cross-validation and weighted f1 scores to assess classification algorithms. It estimates confidence intervals by bootstrapping fold-level scores, and is limited to classification because segmentation ground truth is unavailable.
Proposed procedure
- The benchmark measures classifier performance with the fold average of weighted average f1 scores from stratified cross-validation.Data are shuffled into disjoint folds; each run trains on all but one fold and evaluates on the held-out fold, with fold scores weighted by class sample size.Source: §1.8 · p. 5
Confidence intervals
- 1000 bootstrap sets of fold-level scores produce a 95% confidence interval; the authors recommend 10-fold cross-validation.The interval is obtained from the 25th and 975th sorted bootstrap averages.Source: §1.8 · p. 6
Scope
- The proposed benchmark covers species classification, not segmentation, because segmentation ground truth is unavailable.The authors plan to extend the database with annotations of correct segmentations to support a segmentation benchmark.Source: §1.8 · pp. 6–7
1.9. Discussion — The value of the database
The publicly available database is intended for testing plant-species recognition at early growth stages and may support study of early plant development and high within-species variation. Its indoor capture conditions, resolution, and lack of 3D information constrain what it represents; the authors also hope shared implementations will ease further development.
1.9. Discussion — The value of the database
- The publicly available database is intended to let researchers test plant-species recognition approaches at early growth stages.The authors also hope researchers will share algorithmic implementations to ease further development and reuse of pre- or post-processing.Source: §1.9 · p. 7
- The database could help researchers study early plant morphological and structural development, while object-recognition researchers may benefit from its large intra-class variance relative to inter-class variance.Source: §1.9 · p. 7
- Greenhouse growth with supplemented artificial light omits some sources of variation, lacks pest damage, and may alter plant morphology relative to outdoor growth.The authors note that plants may stretch under the lighting conditions.Source: §1.9 · p. 7
- At approximately 10 px/mm, the images may not contain venation features used in taxonomy protocols, and comparable micro-resolution is not practical on field machinery.The authors suggest downsampling and adding noise to approximate practical image quality and resolution.Source: §1.9 · p. 7
- The authors say 3D information would especially benefit representation of monocots, which mainly grow upward during early growth.Source: §1.9 · p. 7
2. Segmentation
The section demonstrates naive Bayes segmentation of vegetation pixels and discusses its visual result, boundary uncertainty, and evaluation limits.
2. Segmentation
- Prior segmentation work examined color mappings and thresholds, while the authors favor training-based approaches as grounded in the apparent dataset.Related work also reports that an environmentally adaptive Bayesian classifier outperformed a static classifier under changing lighting conditions.Source: §2 · pp. 7–8
2.1. Image segmentation example using Na¨ıve Bayes
- The demonstration models RGB pixels as Gaussian class distributions with equal variances and estimates parameters from 40 uniformly sampled image points.Vegetation is encoded as class 1 and background as class 0.Source: §2.1 · p. 8
2.1.1 Segmentation result and discussion
- The method produces functional segmentation emphasizing vegetation pixels over background, with lower certainty at foreground–background edges; its classifier is based on samples and prior distributions collected from a single image.Figure 3(c) shows each pixel’s posterior probability of being foreground, with black and white representing probabilities 0 and 1, respectively.Source: §2.1.1 · p. 8, Figure 3 · p. 9
- Because the classifier uses samples and prior distributions from one image, it is usable only on images with similar foreground and background distributions.Other segmentation approaches using this database are described in Dyrmann and Christiansen (2013).Source: §2.1.1 · p. 9
- Segmentation is difficult to assess at boundaries because camera pixels average over areas and lens aberration can add nearby noise.The boundary may lack a sharp foreground-background edge, making correct ground-truth annotation difficult.Source: §2.1.1 · p. 9
- The database version discussed lacks ground-truth segmentation labels, so defining a segmentation benchmark is not straightforward.Source: §2.1.1 · p. 9
3. Conclusion
The conclusion presents a public image database of Danish agricultural species and proposes a measure for comparing classification results; naive Bayes segmentation is documented as a demonstration.
3. Conclusion
- The database contains 407 publicly available images of 12 species common in Danish agriculture.Images are associated with species IDs and were recorded multiple times over 20 days.Source: §3 · p. 9
- The paper proposes a benchmark measure to help researchers compare classification results obtained with the database.Source: §3 · p. 9
- A naive Bayes classifier is documented as a demonstration of segmentation using the database.Source: §3 · p. 9
A. Sample plants
The supplied figures show edge uncertainty and sample species images used for shape classification.
A. Sample plants
- Figure 4 encodes foreground-background edge uncertainty from blue at 0% to red at 100%.Source: Figure 4 · p. 10
- Figure 5 presents sample species images for shape classification, created by Dyrmann and Christiansen (2013).Source: Figure 5 · p. 11