Dynamic forest mapping
Predicting canopy height from aerial imagery, trained and validated against LiDAR, so forest structure can be mapped wherever LiDAR doesn’t reach.
Part of a New Zealand–NASA research partnership with NASA Earth Exchange
Solving real-world problems with AI.
I’m an Applied AI Scientist with a PhD and industry experience. I build practical, reliable AI solutions, from early prototypes to real-world deployment.
NowData Scientist at Dragonfly Data Science
Ask about my work
An assistant that answers from my CV, projects and publications. It’s quicker than reading everything — but answers can be imperfect, so follow the source links.
Profile Q&A
Conversation
Prepared answers from Zoey’s profile and publications.
Selected work
Input
Prediction
Finding every tree
New YOLO and Mask R-CNN variants that detect and outline individual tree crowns in aerial imagery — even where canopies overlap. Drag the image to compare input and prediction.
+10% detection and +8% segmentation accuracy · six papers
GovGPT & PaperBoi
Retrieval-augmented assistants that help New Zealanders find government services and policy in plain language, and explore research from New Zealand institutions.
LLM evaluation & document automation
Structured testing of chatbot answers, plus n8n workflows that draft proposals with AI and route every draft through human review before it’s sent.
Weld seam
Glass flaw
Industrial visual inspection
YOLO detectors that locate weld seams on wheel hubs and find flaws in glass — taken from raw factory images to client deployment.
99.98% detection · under 0.3 cm positioning error
Single-image dehazing
Attention-based and CycleGAN methods that recover colour, contrast and fine texture from a single hazy photograph.
Published in ACM TOMM and Computer Vision and Image Understanding
Where I’ve worked
-
Nov 2025 — PresentWellington
Data Scientist
Dragonfly Data Science
Machine learning for forest mapping from aerial imagery and LiDAR, as part of a New Zealand–NASA research partnership.
-
Aug — Oct 2025Auckland
R&D Intern
ElementX
Chatbot backends, structured LLM response evaluation and n8n document automation.
-
Mar — Aug 2025Wellington
Research Assistant
Victoria University of Wellington
Extended my tree-crown instance segmentation research with Manaaki Whenua – Landcare Research.
-
Oct 2024 — Mar 2025Wellington
Digital & AI Intern
Callaghan Innovation
Improved RAG and semantic retrieval for the GovGPT and PaperBoi assistants.
-
Oct 2021 — Feb 2025Wellington
PhD Researcher
Victoria University of Wellington
Detection and instance segmentation of tree crowns in aerial imagery, with large-scale data pipelines and multi-GPU training.
Publications
Published as Ziyi Sun. Most of my work asks how to make detection and segmentation models hold up on messy, real-world imagery.
- 8
- papers
- 7
- first-author
Tree crowns & remote sensing
-
2026
YOLO-ITC: A New YOLO Method for Instance Segmentation of Individual Tree Crowns
IEEE Transactions on Emerging Topics in Computational Intelligence 10(1), 51–65 · early access 2025
-
2026
YOLOv8-UT: A Unified Training Approach for Cross-Environment Tree Crown Instance Segmentation in Aerial Imagery
PRICAI 2025: Trends in Artificial Intelligence LNCS 16455, 298–314 · Springer
-
2025
Feature Fusion to Improve YOLOv8 for Segmenting and Classifying Aerial Images of Tree Crowns
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, 14752–14765
-
2025
YOLOv8E: An Efficient YOLOv8 Method for Instance Segmentation of Individual Tree Crowns in Wellington City, New Zealand
Journal of the Royal Society of New Zealand 55(6), 1854–1879
-
2025
Efficient Tree Mapping through Deep Distance Transform (DDT) Learning
ISPRS Open Journal of Photogrammetry and Remote Sensing 17, 100095
-
2023
An Improved Mask R-CNN for Instance Segmentation of Tree Crowns in Aerial Imagery
38th International Conference on Image and Vision Computing New Zealand (IVCNZ) 1–6
Image dehazing
-
2022
SADnet: Semi-supervised Single Image Dehazing Method Based on an Attention Mechanism
ACM Transactions on Multimedia Computing, Communications, and Applications 18(2), article 58, 1–23
-
2021
ICycleGAN: Single Image Dehazing Based on Iterative Dehazing Model and CycleGAN
Computer Vision and Image Understanding 203, 103133
What I work with
- Code & data
- Python, SQL, Java, C/C++, LaTeX, Pandas, NumPy, SciPy
- Machine learning
- PyTorch, TensorFlow, scikit-learn, OpenCV, CNNs, Transformers, U-Net, YOLO
- Remote sensing
- Aerial imagery, LiDAR, canopy height prediction, tree-crown segmentation, geospatial datasets
- Generative AI
- LLM applications, RAG, semantic retrieval, LLM evaluation, n8n automation
- MLOps & cloud
- Git, GitHub, CI, reproducible pipelines, multi-GPU training, Azure AI Studio, AWS, Linux
- Analysis & reporting
- Statistical analysis, experimental evaluation, Power BI, Matplotlib, Seaborn
- AI-assisted development
- Codex, ChatGPT, Claude Code, Gemini
- Communication
- Papers, conference talks, technical reports, workshops, client demos, AI/ML tutoring
Where I studied
-
PhD
2021 — 2025
Artificial Intelligence
Victoria University of Wellington
-
M.S.
2018 — 2021
Computer Application Technology
Shandong University of Finance and Economics
-
B.S.
2014 — 2018
Digital Media Technology
Shandong University of Finance and Economics
