Onyxia for Geospatial — A Satellite Imagery Use Case

Raya Berova

25 June 2026

Outline

1️⃣ Introduction

2️⃣ Onyxia

3️⃣ Pipeline Walk-through

4️⃣ GeoServer

5️⃣ Web Application

Hello everyone. Today, we will demonstrate the use of Onyxia in a geospatial context, drawing on a concrete use case: the INSEE Satellite Imagery project.

1️⃣ Introduction

The Project: Satellite Imagery at Insee

• Goal: detect buildings on satellite imagery across French overseas territories (OTs)
• Identify areas with the strongest changes in built-up land — constructions & demolitions
• Priority targets: informal settlements in French OTs
• Used to assist and optimise cartographic surveys conducted before each census

The goal of this project is to detect buildings in satellite imagery in order to identify areas experiencing the most significant changes in the built environment—whether through new construction or demolition. The objective is to monitor areas with high levels of “activity,” such as shantytowns in the French Overseas Departments and Regions (DROM).

Final product and pipeline

Demo: web application

Full technical stack of the project

Entire pipeline built and run on Onyxia

Use Case: Cyclone Chido

• Cyclone Chido hit Mayotte in late 2024 — one of the most devastating in decades
• Using our pipeline, we produced a full post-cyclone land cover map and identified destructions in a single night of computation

This project was developed entirely on Onyxia, and its pipeline consists exclusively of Onyxia components. During Cyclone Chido in Mayotte in late 2024, it was possible to generate a post-cyclone land-cover map and identify areas of destruction in a single night of processing. We will walk through this pipeline step by step to explain it in detail.

Note: The project’s backend is developed entirely in Python.

Outline

1️⃣ Introduction

2️⃣ Onyxia

3️⃣ Pipeline Walk-through

4️⃣ GeoServer

5️⃣ Web Application

2️⃣ Onyxia

What is Onyxia SSPCloud?

• Browser-based data science workspace — no local installation
• Launch Python, R environments with CPU / GPU allocated on demand
• Built on Kubernetes → scalable, reproducible, cloud-native
• Integrated S3 object storage — or connect to an external S3 bucket
• Project spaces: share services & data across a team

Onyxia SSPCloud is a data science workspace accessible directly via your web browser. It allows you to launch development sessions (Python, R, etc.) using on-demand CPU and GPU resources, without requiring any local installation. The platform is built on Kubernetes, ensuring flexibility and scalability.

It features integrated S3 object storage for storing and accessing data. You can also connect to an external S3 bucket to retrieve third-party data.

You can request the creation of an Onyxia “project” to share services and data among multiple users. In our case, we have a project dedicated to the Satellite Imagery initiative, as well as an isolated sandbox—this is the environment where I will conduct my demonstrations to avoid interfering with production services and data.

Services Used in Our Project

• VSCode / VSCode GPU
  • Browser-based IDE, central tool for development & experimentation
  • GPU variant for deep learning training (SegFormer fine-tuning)
• MLFlow
  • Experiment tracking: metrics, hyperparameters, artefacts
  • Model Registry: promote a run to Production with one click

Here are the Onyxia services we use in our use case:

[Show Onyxia service catalog]

• VSCode / VSCode GPU: An integrated development environment accessible via the browser. The GPU version enables deep learning model training by leveraging graphics accelerators. It is our core tool for development, debugging, and experimentation.

• MLFlow: A machine learning model lifecycle management platform. It allows us to track experiments (metrics, hyperparameters, artifacts), compare runs, and deploy models to production via its Model Registry.

Services Used in Our Project - 2

• Argo Workflows
  • Kubernetes-native pipeline orchestration
  • DAG-based: each step is a container, dependencies are explicit
  • Enables parallel processing across image tiles
• ArgoCD
  • GitOps continuous deployment
  • Watches our Git repo, auto-syncs the cluster on push
  • Used to deploy & scale our inference API

• ArgoCD: A GitOps-based continuous deployment tool. It monitors a Git repository and automatically synchronizes the Kubernetes cluster state with the declared configuration. We use it to deploy and update our inference API.

• Argo Workflows: A Kubernetes workflow orchestration engine. It allows us to run complex computational pipelines (training, inference) in a parallelized and reproducible manner, directly on the cluster.

Outline

1️⃣ Introduction

2️⃣ Onyxia

3️⃣ Pipeline Walk-through

4️⃣ GeoServer

5️⃣ Web Application

3️⃣ Pipeline Walk-through

Full Pipeline — Live Walk-through

• Step 1 — Explore Data on S3
• Step 2 — Launch Training via Argo Workflows
• Step 3 — Monitor Experiment on MLFlow
• Step 4 — Promote Model to Production on MLFlow
• Step 5 — Deploy Inference API with ArgoCD
• Step 6 — Run Inference & Check Results

All steps demonstrated live in the sandbox project space. 1. Data exploration on S3 — visualization of raw satellite images stored in our bucket. 2. Launching a training job via Argo Workflows — execution of a training pipeline on the cluster. 3. Experiment tracking on MLflow — visualization of metrics (loss, IoU, etc.) and generated artifacts. 4. Promoting the model to production on MLflow — promoting a run to the Model Registry and moving it to the Production stage. 5. Deploying the inference API with ArgoCD — deployment of a FastAPI API connected to the MLflow model, supporting multiple replicas for parallel processing. 6. Launching inference via Argo Workflows — execution of the prediction pipeline on new images. 7. Verifying results on S3 — inspection of generated land cover maps and change detection files.

Outline

1️⃣ Introduction

2️⃣ Onyxia

3️⃣ Pipeline Walk-through

4️⃣ GeoServer

5️⃣ Web Application

4️⃣ GeoServer

What is GeoServer?

• Open-source geospatial data server
• Publishes raster & vector data via OGC standards:
  • WMS (Web Map Service) — rendered map tiles
  • WFS (Web Feature Service) — raw vector features
• Layers are directly consumable by web apps

GeoServer is an open-source geospatial data server. It enables the publishing, storage, and serving of geographic data (rasters, vectors) via standardized protocols such as WMS (Web Map Service) or WFS (Web Feature Service). This makes it possible to distribute map layers that can be consumed directly by web applications.

Deploying GeoServer on Onyxia

• Not in the Onyxia service catalogue — but deployable on the underlying Kubernetes cluster
• Requirements:
  • VSCode service with Kubernetes admin role
  • Define Deployment, Service, Ingress resources manually
• Once deployed: accessible via a cluster Ingress URL like any other service

Demo: GeoServer interface — example WMS layer

GeoServer is not part of the Onyxia service catalog, but it is entirely possible to deploy it manually on the underlying Kubernetes cluster. This requires a VSCode service with the Kubernetes admin role, as well as familiarity with Kubernetes resources: Deployment, Service, and Ingress. I will not go into that level of detail here, but rest assured that the operation is feasible.

Pipeline → GeoServer Integration

• Inference pipeline automatically pushes results to GeoServer on completion:
  • Land cover and Building change polygons → new WFS layers
• GeoServer creates and versions the layers automatically

Our inference pipeline is directly connected to our GeoServer: with each execution, the predicted land cover and detected changes in built structures are automatically pushed to the GeoServer, which creates and updates the associated geographic layers.

Outline

1️⃣ Introduction

2️⃣ Onyxia

3️⃣ Pipeline Walk-through

4️⃣ GeoServer

5️⃣ Web Application

5️⃣ Web Application

Observable Framework

• Frontend built with Observable Framework — JS-based dataviz platform
• Combines markdown, JavaScript, and live data in reactive notebooks
• Consumes GeoServer layers via WMS / WFS for live map rendering

Full stack: Python backend · JavaScript, HTML, CSS frontend

Finally, our web visualization application was developed using Observable Framework (formerly Observable.js). This is a JavaScript framework geared towards data visualization and analysis, which allows for the creation of interactive dashboards combining code, text, and map visualizations. We use it to display geospatial layers from our GeoServer and allow our users to explore changes in built environment across the territory.

I mentioned that our backend is entirely in Python—our frontend, on the other hand, uses JavaScript, HTML, and CSS.

[Show code: WFS calls to GeoServer]

Development & Deployment Workflow

• Develop & preview locally in a VSCode (Onyxia) service with npm
• Git push → GitHub Actions build → auto-deploy to GitHub Pages
• Zero server to maintain, free hosting

Demo: code source of the web app

The development workflow is as follows: changes are made within a VSCode Onyxia service and previewed locally using npm. Once the changes are approved, a Git push automatically triggers the application build and its deployment to GitHub Pages.

Further Resources

• Onyxia tutorials & training courses
• Funathon on Satellite Imagery

• The SSPCloud catalogue hosts many tutorials on Onyxia services (VSCode, MLFlow, Argo, etc.)
• The Funathon is a self-contained hands-on project to discover Onyxia, VSCode and MLFlow using the satellite imagery use case — a great starting point for the full technical stack

Thank you!

Any questions?

Be ready to discuss: - How to get access to SSPCloud / Onyxia - Kubernetes admin role requirements for GeoServer - MLFlow model registry workflow in detail - ArgoCD replica scaling for large inference jobs - Observable Framework vs other frontend options