TreeAI Scientific Competition
More information coming soon!
TreeAI Data Science Competition
Advancing Tree Species Identification from Aerial Images with Deep Learning
Details
The TreeAI Global Initiative is a groundbreaking effort to assess tree species diversity using deep learning.
Objective:
The primary goal of the TreeAI Data Science Competition is to:
- Encourage the development of deep learning models for tree species identification from sub-meter resolution aerial imagery and classification.
- Advance biodiversity monitoring through more accurate AI-basedaccurateby improving the accuracy of AI-drivenbased solutions for forest mapping.
- Foster interdisciplinary collaboration among data scientists, ecologists, and remote sensing experts.
The TreeAI Competition consists of three separate challenges, each focusing on different aspects of AI-driven imagetree analysis for tree species identification. Participants can compete in one or multiple challenges, and each will have its own ranking and prize pool.
Challenges:
1. Object Detection
This challenge focuses on detecting individual trees in high-resolution aerial imagery
2. Tree Species Classification
Participants must classify tree species based on provided image annotations.
3. Instance Segmentation
This challenge requires segmenting individual tree crowns, which, includes providing precise boundaries for each detected tree crown species.
Participation
TreeAI is excited to announce the TreeAI Data Science Competition, a global challenge to develop innovative algorithms for identifying tree species using high-resolution aerial imagery. This competition provides an opportunity for data scientists, researchers, and practitioners to test their skills, contribute to ecological research, and help address critical challenges in forest monitoring.
Participants will work with a curated subset of the TreeAI database, the largest open-access archive of annotated aerialforest imagery, to classify and segment tree species. The results of this competition will contribute to the development of scalable, AI-driven solutions for biodiversity conservation and forest health monitoring.
Timeline
Launch Date: 20 February 2025
Registration Deadline: 15 April 2025
Submission Deadline: 15 June 2025
Winners Announcement: 15 July 2025
Pages
Tasks of the TreeAI Competition
1. TreeAI Classification and Object Detection
Participants are tasked with developing models to identify and classify tree species based on RGB aerial imagery provided by the TreeAI database.
Specific objectives include:
- Segmenting individual trees in images.
- Classifying tree species within these segmented areas.
2. Documentation of Applied Methods/Models
- Participants should describe their applied methods and models, including explanations and relevant figures illustrating the method pipeline or model structure.
- If the method is not yet published, a sufficiently detailed description (e.g., 1–2 pages, A4 format) is recommended to ensure the organizers can understand the approach.
- If the method has been published, participants can provide a reference along with a concise method description (e.g., 1-page A4).
- All method descriptions will contribute to the joint publication of the results.
- Organizers may request additional clarifications during the evaluation process.
3. Final Submission
- Participants should submit their final results, including:
ITC delineation/instance segmentation results
Method descriptions (as outlined above)
- Submissions without sufficient documentation may not be considered for prizes.
Dataset
The training data are collected around the world by multiple platforms and cover various forest types and scenarios, e.g., boreal forest, subtropical forest, tropical forest, urban forest, etc.
There are multiple training datasets collected from different areas around the world with different spatial resolutions (see the resolution per dataset).
The competition dataset includes:
High-resolution RGB aerial orthoimages from diverse forest biomes.
Annotations of tree canopies with labeled species.
Training and validation splits for model development and evaluation.
Datasets for object detection and instance segmentation
Dataset Update:
Dataset no. 1: images with spatial resolution of x cm.
Dataset no. 2: images with spatial resolution of x cm.
Dataset no. 3: images with spatial resolution of x cm.
….
Figure. The examples of the training sets in Phase 1
Datasets for tree species classification
Dataset Update:
Dataset no. 1: images with spatial resolution of x cm.
Dataset no. 2: images with spatial resolution of x cm.
Dataset no. 3: images with spatial resolution of x cm.
….
Figure. The examples of the training sets in Phase 1
Evaluation Metrics:
The organizers will evaluate the models bymodelsalgorithm by comparing the submitted predictionspredictions results with the ground truths. Identical evaluation approaches will be applied to the results of all participants.
Models will be evaluated based on:
Accuracy: Correct identification of tree species.
Intersection over Union (IoU): For segmentation accuracy.
F1 Score: Balancing precision and recall for species classification.
The performances of the participants will be evaluated based on the rank of average ap50 in the final evaluation stage.
The top 6 teams are required to submit the dockers of their models to the organizer to confirm the results. The organizers will contact these teams via email in this regard. In case any participant(s) within the top 6 teams cannot be reached after three email notifications, or the provided model in docker cannot resemble the online reported outcomes, or the submitted docker does not work properly, the corresponding team will be removed from the top 6 list, and the next best-rankedbest ranked team(s) will be promoted to the top 6 teams. The final top 6 teams will be announced.
All participants, regardless of their ranking, should send a Word file to describe their models after the evaluation phase by email.
Prize
Each challenge will be evaluated separately, and winners will be awarded based on their performance in the respective task. Participants can submit entries to multiple challenges to increase their chances of winning.
Five teams will be selected as winners for the following prizes:
Object detection:
1st Place: 1500 euros
2nd Place: 500 euros
Tree species classification:
1st Place: 1000 euros
2nd Place: 500 euros
Instance segmentation:
1st Place: 1500 euros
2nd Place: 500 euros
Terms
**Usage**
Participants should not use the training data in this contest for publishing their own results before the joint publication of this project has been published. The training data will be released after the joint publication as open data for non-commercial applications.
**Rules**
Submissions are NOT compulsory in phase
1. You may submit 20 submissions every day and 2400 in total in this phase.
Submissions must be made before the end of the phase
2. You may submit 100 submissions every day and 500 in total in this phase.
Phases
I am not sure how to handle this part. I have uploaded some data, but I still don't know if it works.
TreeAI
Training phase
Evaluation phase
Leaderboard
I am not sure how to handle this part. I have uploaded some data, but I still don't think it works.
Administrators
Stefano and Xinlian
Some screenshots
Details
Example of competitions:
external page https://www.codabench.org/competitions/2018/#/pages-tab
external page https://www.codabench.org/competitions/3667/#/phases-tab