Sentinel Imagery Selection Tool
Selection tool enabling efficient visual inspection of Sentinel satellite imagery across specified points and dates, managing workflows to identify cloud-free images for machine learning training datasets.
Client
Digital Agriculture Services
Key Results
Workflow Reduction
ML Dataset Quality
Team Collaboration
Cloud Detection
The Challenge
Selecting appropriate satellite images for ML training required manually reviewing thousands of images to find cloud-free captures—a process that took weeks per project.
Key challenges included:
- Manual review of thousands of satellite images
- Inconsistent cloud detection across scenes
- Time-consuming quality assessment process
- No standardized workflow for image selection
- Difficulty tracking review progress across teams
ML Data Quality
Machine learning models for remote sensing require high-quality, cloud-free training data. Poor image selection directly impacts model accuracy and downstream agricultural insights.
Training Data Quality
Critical for model accuracy
Cloud Contamination
Major source of bad training data
Our Solution
Built a specialized selection interface with automated cloud detection, allowing rapid visual inspection and classification of imagery. Integrated with existing ML pipelines for seamless data preparation.
Automated Cloud Detection
Pre-filters imagery using cloud probability masks, surfacing only candidate images likely to be usable.
Visual Inspection Interface
Streamlined UI for rapid review and classification of satellite imagery tiles.
Sentinel API Integration
Direct integration with Copernicus Open Access Hub for accessing Sentinel-2 imagery.
Workflow Management
Track review progress across team members with assignment and status tracking.
D3.js Visualizations
Interactive charts showing temporal coverage and quality metrics across regions.
ML Pipeline Integration
Direct export to training data formats with proper labeling and metadata.
Technology Stack
Project Impact
Workflow Efficiency
- Reduced image selection from weeks to days
- Automated pre-filtering eliminates 80% of unusable images
- Standardized workflow across all team members
- Clear progress tracking and assignment management
Data Quality
- Consistent quality standards across projects
- Improved ML model accuracy from better training data
- Reduced rework from poor image selection
- Comprehensive audit trail for data provenance