Machine Learning

Object Detection

A computer vision technique that identifies and locates specific objects within satellite imagery by drawing bounding boxes around them. Used for counting buildings, detecting ships, identifying vehicles, and monitoring infrastructure from space.

Overview

Object detection locates and classifies individual objects in satellite imagery with bounding boxes and class labels. Unlike segmentation, it identifies discrete countable objects — vehicles, buildings, ships, solar panels — and reports their positions, sizes, and categories.

How It Works

Two-stage detectors like Faster R-CNN generate region proposals then classify each one — more accurate for small objects but slower. Single-stage detectors like YOLO predict boxes and classes in one pass — faster, suitable for large archives. DETR uses transformers for set prediction, eliminating hand-designed components. Feature Pyramid Networks handle the wide range of object scales in satellite imagery.

Key Facts

  • YOLO models process satellite tiles in milliseconds; Faster R-CNN offers higher accuracy on small objects.
  • Specialized datasets like DOTA and xView contain hundreds of thousands of labeled objects.
  • Small object detection is the primary challenge — cars may be only 3-5 pixels wide.
  • Mask R-CNN extends detection with instance segmentation masks.

Applications

Ship Detection

Detecting vessels from optical and SAR imagery for maritime surveillance and illegal fishing monitoring.

Building Footprint Extraction

Creating building inventories for urban planning, population estimation, and damage assessment.

Vehicle Counting

Estimating economic activity and traffic patterns from parking lot and road imagery.

Infrastructure Mapping

Locating solar panels, wind turbines, and energy infrastructure across large areas.

Limitations & Considerations

Objects in satellite imagery are often extremely small, viewed from overhead with no perspective variation, and densely packed. Cloud cover and shadows degrade accuracy. Models trained on one sensor often fail on different satellites without retraining. Processing large scenes requires tiling strategies that can miss objects at tile boundaries.

History & Background

Transformed by R-CNN (2014), Fast/Faster R-CNN (2015), and YOLO (2016). DETR introduced transformers in 2020. Large benchmarks like xView (2018) accelerated progress.

Analyze Object Detection data with LYRASENSE

Use our agentic notebook environment to work with satellite data and compute indices like Object Detection — no setup required.