Vehicle Detection in Satellite Imagery – Large vs. Small

nput: High‑resolution satellite imagery (e.g., 0.3–0.5 m/pixel)
Output: Bounding boxes or points with class: large_vehicle or small_vehicle
Goal: Separate vehicle detections into two size‑based categories

Project Overview

Nazru’s AI platform was deployed to detect and classify vehicles from high‑resolution satellite images using a binary size‑based classification: large vehicles (trucks, buses, heavy equipment) and small vehicles (cars, motorcycles, small vans). This simplified taxonomy supports applications such as parking lot analysis (distinguishing car vs. truck spaces), traffic density estimation by vehicle type, and urban infrastructure planning.

The Challenge

Satellite‑based vehicle detection already faces small object sizes, occlusions, and variable orientations. Adding a size‑based binary classification requires the model to differentiate between visually similar roof shapes solely by scale and context (e.g., a small delivery van vs. a large truck). The client needed a clean separation without fine‑grained class labels like “car” or “bus”.

The Nazru Solution

Nazru’s platform used a single‑stage detector (e.g., YOLOv8) trained on satellite images annotated with only two classes: small_vehicle and large_vehicle. The model automatically:

• Detects all vehicles as oriented or axis‑aligned bounding boxes.

• Classifies each detection as small_vehicle (e.g., car, SUV, motorcycle) or large_vehicle (e.g., truck, bus, construction vehicle).

• Outputs bounding boxes with confidence scores.

Threshold for size is defined in pixel area or physical length (e.g., > 5 metres = large vehicle) based on ground sample distance (GSD). The model implicitly learns this distinction from training examples.

Key Results & Benefits

• Accurate binary separation of vehicles by size from satellite images.

• Simplified class schema – easier to annotate and faster to train.

• Useful for:

  • Parking lot management (truck vs. car parking).

  • Traffic monitoring (percentage of large vehicles on roads).

  • Disaster response (identifying large vehicles for evacuation).

  • Military/security applications (distinguishing armored vehicles).

• Scalable to city‑wide or regional satellite coverage.

Key Technologies Used

• Object detection (YOLOv8 / EfficientDet) optimised for small objects in satellite imagery.

• Size‑aware classification – model uses bounding box dimensions and appearance features.

• Oriented bounding boxes (optional) for better fit.

• GIS output with georeferenced detections.