Software Development · Full-Stack & AI · Geospatial ML

Bartek
Tedys

I trained a YOLOv11x segmentation model to detect individual tree canopies from aerial imagery at centimetre resolution. This is that model, running live.

Run the model →See city-scale results

8cm

pixel resolution

YOLOv11x

segmentation model

640px

tile size

PDOK

aerial source

Live inference

Upload aerial imagery

Drop a top-down aerial image and the model will segment individual tree canopies. Use imagery zoomed to roughly street level. The model works best at resolutions of 8–25cm per pixel. No map labels or overlays. Processing takes 20–40s on CPU.

Or try a sample

↑

Drop your own image here

JPG · PNG · GeoTIFF - zoomed to street level, no map labels

→ Or select a region on the map below to pull live PDOK aerial tiles automatically.

Research

About the model

This model is the output of my final year thesis at Munster Technological University: a comparative analysis of three instance segmentation architectures for detecting individual tree canopies in urban aerial imagery.

Urban green spaces are critical to environmental sustainability and quality of life, but effective monitoring at scale remains difficult. This research investigates whether deep learning can automate tree crown delineation accurately enough to be useful for urban forestry and planning applications.

The three models evaluated were YOLOv11, Mask R-CNN, and YOLACT++. YOLOv11x-seg, the model running on this site, outperformed the others across all key metrics.

ArchitectureYOLOv11x-seg
Training dataCustom annotated dataset, Netherlands aerial imagery
Image sourcePDOK Beeldmateriaal, 8cm & 25cm/px RGB orthophotos
Tile size640 × 640px
FrameworkUltralytics / PyTorch
DeploymentONNX → QGIS Deepness plugin / Modal serverless

About

Who I am

Bartek Tedys

Bartlomiej Tedys

Ireland · Open to relocation

I'm Bartek, a software developer who just finished a BSc in Software Development (First-Class Honours) at Munster Technological University in Cork, with a semester abroad at Hogeschool van Amsterdam.

Day to day, I build full-stack web apps and work on applied AI: I write and evaluate prompts for AI model training, and build practical tools that automate real business processes, like a contract and quotation platform for a home renovation company, and a live marketplace platform for a steel trading business, including its Microsoft 365 setup.

My thesis compared three instance segmentation models for detecting individual tree canopies in aerial imagery. YOLOv11x came out on top, and the model on this page is the result of that work. Geospatial ML and computer vision are genuine interests I'd like to build a career in over time.

I'm based in Ireland and have a lot of respect for the Dutch tech scene, both technically and professionally, so I'm open to relocating for the right opportunity. Right now I'm focused on broadening into general software development, AI integration and systems work, with geospatial ML as a specialism I keep coming back to.

Available for work - open to relocation

Technical skills

Frontend

ReactNext.jsTypeScriptTailwind CSS

Backend

PythonFastAPINode.jsPostgreSQLREST APIsSupabase

AI & Automation

Prompt EngineeringLLM EvaluationPyTorchTensorFlowInstance Segmentation

Geospatial

QGISPDOK / WMS/WMTSGeoJSONLeafletRemote SensingDeepness

IT & Systems

Microsoft 365 AdminEmail & MFA SetupGitLinux

Education

Munster Technological University

2021 – 2025

BSc Software Development (Honours)

Cork, Ireland

→ Thesis: Comparative analysis of instance segmentation models for urban tree crown delineation

Hogeschool van Amsterdam

2023 – 2024

Erasmus Exchange - Big Data & Mobile App Development

Amsterdam, Netherlands

Other work

2026

ReactNext.jsMicrosoft 365

EkoDorStal - Steel Trading Platform & MS365 Admin

Designed, built and launched the platform website for a B2B/B2C marketplace trading surplus industrial steel, now expanding into a full e-commerce shop platform. Also set up and administer Microsoft 365 for the business, including email services and multi-factor authentication.

2024 - Present

Prompt EngineeringLLM EvaluationQA

AI Model Training & Prompt Engineering - Outlier

Write and refine prompts to elicit, test and evaluate AI model responses across generalist, multilingual and technical domains. Evaluate multimodal outputs (text, image, speech) and lead quality assurance on training datasets across international projects.

2023 - Present

ReactViteNext.jsSupabase

M&L Home Builds - Business Website & Ongoing IT

Designed and built a conversion-focused website for an Irish home renovation company, later rebuilt in React with Vite for a refreshed design and more interactivity. Separately built a full-stack contract and quotation tool (Next.js, Supabase, automated contract signing) used in the client’s day-to-day operations, and manage ongoing hosting and email for the business.

2024

PythonPyTorchComputer VisionRemote Sensing

Caeli - Computer Vision Internship

Internship at Caeli, a Dutch startup using AI to monitor vegetation from aerial imagery. Trained and evaluated object detection models on satellite and drone data. This experience directly led to my thesis on tree canopy segmentation.

Contact

Get in touch

Whether you have a project in mind, want to discuss software, AI or geospatial ML, or are looking to hire - I'm open to conversations.

barttedys.nl

Model: YOLOv11x-seg · Data: PDOK Beeldmateriaal · Built with Next.js + Modal