WeedDetect-AI 2026

Big challenge 03 · AI in Agriculture

Beyond Weed Recognition: Field Species Identification and Efficient Computer Vision for Agriculture

Team registration openSubmissions open

WEED

Category: Artificial Intelligence in Agriculture
Format: Online, November – December 2026 · final presentations on 17 December
Tracks: 2 ranked + demo · up to 3 per team
Teams: Solo or teams of up to 4
Submission deadline: 10 December 2026

Overview

Weeds compete with cultivated plants and pasture vegetation for light, water and nutrients. Their identification is an important step in crop monitoring and the management of agricultural land.

Field photographs are difficult to classify consistently because plants are observed under changing light, varied backgrounds and different stages of growth. Leaves may be partly hidden, and several species can have similar visual characteristics.

WeedDetect-AI 2026 invites participants to develop image classification methods that recognise weed species from outdoor photographs and remain reliable when image quality changes.

The competition uses DeepWeeds, a publicly released dataset containing 17,509 field images from eight weed species and a negative category.

Objectives

01Develop accurate AI-based methods for weed species recognition.
02Distinguish eight target weed species from images without those species.
03Handle class imbalance and complex field backgrounds.
04Study the effect of blur, illumination changes and image compression.
05Develop efficient models that can run on CPU-based systems.
06Compare computer vision architectures and training strategies.
07Encourage reproducible research in agricultural image analysis.

The data

DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning — Scientific Reports, 2019.

Dataset source

Imaging modality: RGB field photographs
Images: 17,509
Target weed species: 8
Classification categories: 9 (including Negative)
Image size: 256 × 256 pixels in the published dataset
Collection environment: Outdoor sites in Queensland, Australia
Annotations: Image-level species labels
Data format: RGB image files with accompanying labels and splits
Licence: Creative Commons Attribution 4.0 (CC BY 4.0)

The eight species are Chinee apple, Lantana, Parkinsonia, Parthenium, Prickly acacia, Rubber vine, Siam weed and Snake weed. The Negative category contains images without the eight target species.
The dataset includes 9,106 Negative images, slightly more than half of all photographs. For this reason, classification accuracy alone is not sufficient for comparing models.
Images and annotations are released under CC BY 4.0; the original repository also contains code under Apache 2.0. Dataset attribution and the corresponding licences must be retained.
The original labels and split files are publicly accessible. Results are therefore checked through code reproduction and declared data use. A separate collection of unseen field images would be required for a blind test leaderboard.

Tracks

A team may enter up to 3 of these. Each ranked track has its own leaderboard.

01Track A

Leaderboard · primary metric: macro F1

Weed Species Classification

Develop an AI model that takes an RGB field photograph as input and predicts the corresponding weed category among nine classes.

Classes

Chinee apple
Lantana
Parkinsonia
Parthenium
Prickly acacia
Rubber vine
Siam weed
Snake weed
Negativenone of the eight target species

Each image must receive a probability for every class, allowing the prediction to be assessed even when the model confuses visually similar plants.

Main technical challenges

Distinguishing species with similar leaves or branching patterns.
Handling variable lighting and outdoor backgrounds.
Recognising partly occluded plants.
Managing the large Negative class.
Avoiding false weed predictions on ordinary vegetation.

Evaluation

Metric	Purpose
Macro F1-scorePrimary	Measures classification performance across all nine classes with equal class weighting; primary ranking metric.
Balanced accuracy	Measures average recall across the classes.
Top-1 accuracy	Measures the proportion of images assigned to the correct category.
Confusion matrix	Shows the principal errors between species and the Negative class.

Every category, including Negative, contributes to the score. Ties are resolved by balanced accuracy.

Expected output

Nine class probabilities for every evaluation image.

You submit

Track A predictions
One row per requested image; probabilities between 0 and 1 that sum to one.
image_id,p_Chinee_apple,p_Lantana,p_Parkinsonia,p_Parthenium,p_Prickly_acacia,p_Rubber_vine,p_Siam_weed,p_Snake_weed,p_Negative
CSV · up to 5 MB
Trained model weights
Over 10 MB? Share a download link (Drive, Hugging Face, Zenodo…).
PT, PTH, ONNX, H5, KERAS, TFLITE, ZIP · up to 10 MB, or a link

02Track B

Leaderboard · primary metric: robust macro F1

Robust and Efficient Weed Recognition

Develop a compact classification model that maintains its performance when input image quality changes. It is tested on the original images and on standardised image variations prepared by the organisers.

Tested under

Changes in brightness and contrast
Moderate image blur
JPEG compression
Limited sensor-like image noise
Small changes in image scale or crop

All transformations preserve the original class label. The final set of transformations and their intensity levels will be published before the competition begins.

Main technical challenges

Maintaining classification accuracy under image degradation.
Reducing model size and memory use.
Achieving acceptable inference speed without a GPU.
Balancing efficiency and performance across the nine classes.
Ensuring consistent results on the reference evaluation machine.

Evaluation

Metric	Purpose
Robust macro F1-scorePrimary	Mean of macro F1 across the original images and the published image-quality conditions; primary ranking metric.
Worst-condition macro F1	Measures performance in the most difficult evaluated condition.
Inference time	Measures median CPU inference time per image at batch size one.
Model size and memory	Measures the stored model size and peak inference RAM.

Models run on a standard CPU environment without internet access. The proposed size limit is 50 MB for model weights; a CPU memory limit and maximum inference time will be confirmed after the organisers test the baseline model on the reference machine. Ties are resolved by worst-condition macro F1, then lower inference time.

Expected output

A compact model, prediction probabilities and a report of inference time, memory use and model size.

You submit

Track B predictions
One row per requested image; probabilities between 0 and 1 that sum to one.
image_id,p_Chinee_apple,p_Lantana,p_Parkinsonia,p_Parthenium,p_Prickly_acacia,p_Rubber_vine,p_Siam_weed,p_Snake_weed,p_Negative
CSV · up to 5 MB
Compact model weights (≤ 50 MB)
Upload if it is 10 MB or less; otherwise share a download link.
PT, PTH, ONNX, H5, KERAS, TFLITE, ZIP · up to 10 MB, or a link
Efficiency report
Model weight size, peak memory usage and median CPU latency (batch size one).
PDF · up to 10 MB

+Optional

Advanced demonstration · not a leaderboard

Field Weed Recognition Assistant

Teams may combine species recognition and lightweight deployment in an offline demonstration application that displays the predicted class, alternative predictions and model confidence. This is an advanced demonstration category rather than a separate mandatory leaderboard.

You submit

Demonstration (video or live link)optional
A short screen recording or a link to the application.
Link

How to submit

Register your team, then upload each deliverable from the participant portal. Prediction files are checked the moment you upload them, and your latest upload before the deadline is the one evaluated.

Suggested approaches

Convolutional neural networks (CNNs)
MobileNet and other lightweight networks
EfficientNet and compact residual networks
Vision transformers and hybrid architectures
Transfer learning and data augmentation
Model compression, pruning and quantisation

Track A · Weed Species Classification

Track A predictions
One row per requested image; probabilities between 0 and 1 that sum to one.
image_id,p_Chinee_apple,p_Lantana,p_Parkinsonia,p_Parthenium,p_Prickly_acacia,p_Rubber_vine,p_Siam_weed,p_Snake_weed,p_Negative
CSV · up to 5 MB
Trained model weights
Over 10 MB? Share a download link (Drive, Hugging Face, Zenodo…).
PT, PTH, ONNX, H5, KERAS, TFLITE, ZIP · up to 10 MB, or a link

Track B · Robust and Efficient Weed Recognition

Track B predictions
One row per requested image; probabilities between 0 and 1 that sum to one.
image_id,p_Chinee_apple,p_Lantana,p_Parkinsonia,p_Parthenium,p_Prickly_acacia,p_Rubber_vine,p_Siam_weed,p_Snake_weed,p_Negative
CSV · up to 5 MB
Compact model weights (≤ 50 MB)
Upload if it is 10 MB or less; otherwise share a download link.
PT, PTH, ONNX, H5, KERAS, TFLITE, ZIP · up to 10 MB, or a link
Efficiency report
Model weight size, peak memory usage and median CPU latency (batch size one).
PDF · up to 10 MB

Optional · Field Weed Recognition Assistant

Demonstration (video or live link)optional
A short screen recording or a link to the application.
Link

Every team

Source code & inference instructions
Repository link (GitHub, GitLab, Hugging Face…) with installation and execution instructions.
Link
Technical description
Architecture and training methodology; preprocessing and augmentation; evaluation results and computational requirements.
PDF · up to 10 MB

Technical requirements

The organisers provide the official class order, partition files, baseline model, evaluation code and image-transformation protocol.
The prediction file contains one row per requested image. All probabilities are between 0 and 1 and sum to one within the tolerance announced by the organisers.
Track B submissions also include model weight size, peak memory usage and CPU latency measurements. The final ranking is based on the same reference machine for all participants.

Competition rules

1Participants must use the dataset and official data partitions specified by the organisers.
2Training and validation must use only the permitted partitions; test labels must not be used for model selection.
3The Negative class must be included in training and evaluation.
4All external datasets and pretrained models must be declared.
5Model predictions must be generated from the image pixels; image-ID or hash-based label lookup is prohibited.
6Track B models must satisfy the published CPU inference and model-size requirements.
7Source code and trained weights must be supplied for reproducibility.
8The organisers reserve the right to verify submissions and reject results obtained through test-data leakage.

The dataset contains Australian weed species. Performance on DeepWeeds must not be presented as evidence that the model is validated for Algerian crops or local weed populations.

The final submission format and evaluation instructions will be announced before the competition begins.

Timeline

Key dates.

Call for participation
November 2026
Registration
November – December 2026
Dataset and technical documentation release
13 November 2026
Model development
November – December 2026
Submission deadline
12 December 2026
Code verification and final evaluation
13 – 16 December 2026
Final presentations and demonstrations
17 December 2026
Results and awards
17 December 2026

The final dates will be confirmed by the organising committee.

Recognition

What’s at stake.

Best Overall Performance
Best Weed Classification Model
Best Robust and Efficient Model
Most Innovative Approach
Best Technical Presentation

Awards and certificates are subject to confirmation by the symposium committee.

Expected outcomes

Encourage research in agricultural computer vision.
Introduce participants to species classification under real field conditions.
Promote the use of class-balanced evaluation metrics.
Develop efficient models for CPU-based recognition.
Compare methods for handling image-quality changes.
Identify approaches for future precision agriculture studies.

Outstanding teams may be invited to present their approaches and discuss future work on locally collected agricultural data.

Why it matters

Weed recognition is a useful component of field monitoring and precision agriculture. Classification models can help identify target species in photographs and support the development of more specialised agricultural tools.

By evaluating both recognition quality and CPU-based efficiency, the competition addresses practical constraints that can affect the use of AI systems outside controlled environments.

Important

DeepWeeds contains image-level labels rather than object locations or segmentation masks. This competition does not evaluate weed localisation, infestation area or treatment recommendations; additional annotated data would be needed for those tasks.

References

[1]Olsen, A., Konovalov, D. A., Philippa, B., Ridd, P., Wood, J. C., Johns, J., Banks, W., Girgenti, B., Kenny, O., Whinney, J., et al. (2019). DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning. Scientific Reports, 9, 2058.doi.org/10.1038/s41598-018-38343-3
[2]Olsen, A., et al. DeepWeeds: dataset, annotations and original implementation.github.com/AlexOlsen/DeepWeeds
[3]Creative Commons Attribution 4.0 International Public License (CC BY 4.0).creativecommons.org/licenses/by/4.0/

WeedDetect-AI 2026 · ARC'26

Bring your best model.

Weeds compete with cultivated plants and pasture vegetation for light, water and nutrients. Their identification is an important step in crop monitoring and the management of agricultural land.

WeedDetect-AI 2026 invites participants to develop image classification methods that recognise weed species from outdoor photographs and remain reliable when image quality changes.

The competition uses DeepWeeds, a publicly released dataset containing 17,509 field images from eight weed species and a negative category.

Objectives

01Develop accurate AI-based methods for weed species recognition.
02Distinguish eight target weed species from images without those species.
03Handle class imbalance and complex field backgrounds.
04Study the effect of blur, illumination changes and image compression.
05Develop efficient models that can run on CPU-based systems.
06Compare computer vision architectures and training strategies.
07Encourage reproducible research in agricultural image analysis.

Metric

Purpose

Macro F1-scorePrimary

Measures classification performance across all nine classes with equal class weighting; primary ranking metric.

Balanced accuracy

Measures average recall across the classes.

Top-1 accuracy

Measures the proportion of images assigned to the correct category.

Confusion matrix

Shows the principal errors between species and the Negative class.

Metric

Purpose

Robust macro F1-scorePrimary

Mean of macro F1 across the original images and the published image-quality conditions; primary ranking metric.

Worst-condition macro F1

Measures performance in the most difficult evaluated condition.

Inference time

Measures median CPU inference time per image at batch size one.

Model size and memory

Measures the stored model size and peak inference RAM.

Register your team, then upload each deliverable from the participant portal. Prediction files are checked the moment you upload them, and your latest upload before the deadline is the one evaluated.

Suggested approaches

Convolutional neural networks (CNNs)
MobileNet and other lightweight networks
EfficientNet and compact residual networks
Vision transformers and hybrid architectures
Transfer learning and data augmentation
Model compression, pruning and quantisation

1Participants must use the dataset and official data partitions specified by the organisers.
2Training and validation must use only the permitted partitions; test labels must not be used for model selection.
3The Negative class must be included in training and evaluation.
4All external datasets and pretrained models must be declared.
5Model predictions must be generated from the image pixels; image-ID or hash-based label lookup is prohibited.
6Track B models must satisfy the published CPU inference and model-size requirements.
7Source code and trained weights must be supplied for reproducibility.
8The organisers reserve the right to verify submissions and reject results obtained through test-data leakage.

The dataset contains Australian weed species. Performance on DeepWeeds must not be presented as evidence that the model is validated for Algerian crops or local weed populations.

The final submission format and evaluation instructions will be announced before the competition begins.

Key dates.

Call for participation
November 2026
Registration
November – December 2026
Dataset and technical documentation release
13 November 2026
Model development
November – December 2026
Submission deadline
12 December 2026
Code verification and final evaluation
13 – 16 December 2026
Final presentations and demonstrations
17 December 2026
Results and awards
17 December 2026

The final dates will be confirmed by the organising committee.

What’s at stake.

Best Overall Performance
Best Weed Classification Model
Best Robust and Efficient Model
Most Innovative Approach
Best Technical Presentation

Awards and certificates are subject to confirmation by the symposium committee.

Expected outcomes

Encourage research in agricultural computer vision.
Introduce participants to species classification under real field conditions.
Promote the use of class-balanced evaluation metrics.
Develop efficient models for CPU-based recognition.
Compare methods for handling image-quality changes.
Identify approaches for future precision agriculture studies.

Outstanding teams may be invited to present their approaches and discuss future work on locally collected agricultural data.

Real fields. Robust models.

DeepWeeds

Choose your track.

Ship it from your team workspace.

Play fair, show your work.

Key dates.

What’s at stake.

Bring your best model.

WeedDetect-AI 2026

Real fields. Robust models.

DeepWeeds

Choose your track.

Ship it from your team workspace.

Play fair, show your work.

Key dates.

What’s at stake.

Bring your best model.