Deploy AI-powered resolution enhancement and object detection for satellite imagery using OpenShift AI on Red Hat OpenShift.
Welcome to CAIsat, where you can gaze upon Earth from space, enhance satellite imagery, and detect objects with the power of AI. This application lets users navigate live satellite maps, capture regions to enhance from 256×256 to 512×512 resolution using the SwinIR deep learning model, and detect objects like planes, ships, vehicles, and infrastructure using YOLOv8-OBB — all running on Red Hat OpenShift AI.
┌─────────────────────────────────────────────────────────────────┐
│ Frontend (React) │
│ • Three.js 3D rotating Earth │
│ • Leaflet satellite map (Esri World Imagery) │
│ • Drag-and-drop selection box (256×256 region) │
│ • Mouse wheel zoom for detail inspection │
│ • Before/after/detected comparison view │
│ • Collapsible image panels │
│ • Bounding box visualization │
└──────┬──────────────────────────────────────────────────────────┘
│ REST API
├─────────────────────────────────────┬───────────────────►
▼ ▼
┌────────────────────────────────┐ ┌────────────────────────────┐
│ Enhancement Backend │ │ Detection Backend │
│ (FastAPI + Python) │ │ (FastAPI + Python) │
│ • Crop 256×256 region │ │ • Preprocess for YOLO │
│ • Normalize to tensor │ │ • NMS post-processing │
│ • Track statistics │ │ • Bounding box transform │
└──────┬─────────────────────────┘ └──────┬─────────────────────┘
│ KServe v2 │ KServe v2
▼ ▼
┌────────────────────────────────┐ ┌────────────────────────────┐
│ SwinIR Model │ │ YOLOv8-OBB Model │
│ (MLServer / ONNX) │ │ (MLServer / ONNX) │
│ • 2× resolution enhancement │ │ • 15 object classes │
│ • 256×256 → 512×512 │ │ • Bounding boxes │
└────────────────────────────────┘ └────────────────────────────┘
The application consists of five containerized components deployed on OpenShift:
- Frontend (React): 3D Earth globe, interactive satellite map, drag-and-drop selection box, zoom controls, and collapsible detection results
- Enhancement Backend (FastAPI): Image preprocessing, crop extraction, tensor conversion, and KServe communication for SwinIR
- Detection Backend (FastAPI): Image preprocessing, YOLO inference coordination, NMS post-processing, and bounding box visualization
- SwinIR Model Server (OpenShift AI): ONNX model served via MLServer for 2× resolution enhancement
- YOLOv8-OBB Model Server (OpenShift AI): ONNX model served via MLServer for object detection in satellite imagery
Activate satellite view to browse satellite imagery, capture a screenshot, select a 256×256 region of interest, enhance it to 512×512, then detect objects like planes, ships, and vehicles with bounding boxes.
- CPU: 7.5 cores minimum
- Frontend: 50 millicores request, 200 millicores limit
- Enhancement Backend: 100 millicores request, 500 millicores limit
- Detection Backend: 250 millicores request, 1 core limit
- SwinIR Model Server: 4 cores request, 4 cores limit
- YOLOv8 Model Server: 2 cores request, 4 cores limit
- Memory: 13.5 GiB minimum
- Frontend: 128 MiB request, 256 MiB limit
- Backend: 256 MiB request, 512 MiB limit
- Model Server: 6 GiB request, 8 GiB limit
- GPU: Not required (CPU-only inference supported, ~7-15 seconds per image)
- Ability to create projects/namespaces
- Ability to deploy workloads (Deployments, Services, Routes)
- Ability to create InferenceService resources (for OpenShift AI model serving)
- Ability to create ServingRuntime resources
- Red Hat OpenShift 4.2x cluster
- OpenShift AI 3.x
- oc CLI authenticated to your cluster
- Helm 3.10+ installed locally
-
Clone this repository:
git clone https://github.com/rh-ai-quickstart/caisat.git cd caisat -
Create a namespace:
oc new-project caisat
-
Deploy the application using Helm:
helm install caisat ./chart \ --namespace caisat \ --set model.deploy=true
-
Wait for all pods to be ready (this may take 2-3 minutes):
oc get pods -n caisat -w
-
Get the application URL:
oc get route caisat-frontend -n caisat -o jsonpath='{.spec.host}'
-
Check that all pods are running:
oc get pods -n caisat
Expected output showing 5 running pods:
NAME READY STATUS RESTARTS AGE caisat-backend-xxxxxxxxx-xxxxx 1/1 Running 0 2m caisat-detection-backend-xxxxxxxxx-xxxxx 1/1 Running 0 2m caisat-frontend-xxxxxxxxx-xxxxx 1/1 Running 0 2m swinir-predictor-xxxxxxxxx-xxxxx 2/2 Running 0 2m yolov8m-satelite-predictor-xxxxxxxxx-xxxxx 2/2 Running 0 2m -
Verify both model servers are ready:
oc get inferenceservice -n caisat
Both InferenceServices should show
READY: True. -
Access the application:
- Open the route URL from step 5 of Installation in your web browser
- You should see the CAIsat interface with a rotating 3D Earth
- Try the full workflow described in How It Works to verify enhancement and detection are functioning
To completely remove the application:
helm uninstall caisat -n caisat
oc delete project caisat- Activate: Click "Satellite View" to switch to interactive Esri satellite imagery
- Navigate: Pan and zoom to find your area of interest (cities, coastlines, forests, airports, harbors, etc.)
- Capture: Click "Capture & Enhance" to take a screenshot of the current map view
- Select: Drag the red 256×256 selection box over your target region
- Use scroll wheel to zoom in for precision
- Position the box exactly where you want enhancement
- Enhance: Click "Enhance Selected Area"
- Backend extracts the selected 256×256 region
- Image is converted to a normalized tensor
- SwinIR model processes it to 512×512 (2× upscaling)
- Detect: Click "Detect Objects" on the enhanced image
- Detection backend preprocesses the image for YOLOv8
- YOLOv8-OBB model identifies 15 object classes
- Non-maximum suppression filters overlapping detections
- Bounding boxes are drawn with confidence scores
- Compare: View original, enhanced, and detected images (collapsible panels)
- Review: See detected objects with class names and confidence scores
- Download: Save the enhanced or detected image for your records
Processing time: ~10-15 seconds per enhancement, ~5-6 seconds per detection on CPU.
Detectable Objects: plane, ship, storage-tank, baseball-diamond, tennis-court, basketball-court, ground-track-field, harbor, bridge, large-vehicle, small-vehicle, helicopter, roundabout, soccer-ball-field, swimming-pool. Processing time: ~7-15 seconds per image on CPU.
Enhance satellite imagery of agricultural fields to detect crop stress, improve field boundary detection, and analyze historical low-resolution archives. Detect vehicles, storage tanks, and infrastructure to monitor farm operations and equipment distribution.
Improve resolution of building and infrastructure details for development zone analysis. Detect vehicles, bridges, roundabouts, and sports facilities to analyze urban infrastructure distribution and monitor city growth patterns.
Detect ships, harbors, bridges, and storage tanks for maritime traffic analysis, port activity monitoring, and coastal infrastructure assessment. Track vessel movements and monitor harbor capacity utilization.
Detect planes, helicopters, and vehicles at airports and transportation hubs. Monitor aircraft parking positions, analyze airport capacity, and track vehicle movements in large facilities.
Demonstrate AI image processing techniques in Earth observation courses. Show students how deep learning models can be applied to satellite imagery and explore the capabilities and limitations of AI-based enhancement.
- SwinIR Model Paper - Liang et al., 2021
- YOLOv8 Documentation - Ultralytics
- DOTA Dataset - Object Detection in Aerial Images
- Esri World Imagery
Apache 2.0 License - See LICENSE file
- Built by: Red Hat CAI Team
- Powered by: Red Hat OpenShift AI
- Models: SwinIR by Jingyun Liang et al., YOLOv8-OBB by Ultralytics
- Satellite Imagery: Esri World Imagery Service
- Base Images: Red Hat Universal Base Image 9 (UBI9)
Happy exploring! 🌍🛰️