Exploring convolutional neural networks and spatial video for on-the-ground mapping in informal settlements

Area of Science:

• Environmental Health
• Geographic Information Systems
• Computer Science

Background:

• Informal settlements in developing nations often lack spatial data crucial for health interventions.
• Spatial video (SV) offers granular environmental and social data but manual processing is a scalability bottleneck.

Purpose of the Study:

• To explore the use of Convolutional Neural Networks (CNNs) for automated identification of disease-related environmental risks from SV data.
• To assess the potential of machine learning (ML) for health risk mapping in informal settlements and identify training challenges.

Main Methods:

• Utilized CNNs to analyze SV data collected from informal settlements in Haiti.
• Assessed the classification performance of ML models for various environmental features.
• Investigated the impact of frame selection, image resolution, and their combinations on model performance.

Main Results:

• SV is a viable data source for ML-based identification of health risk features.
• Well-defined objects (drains, tires) were classified effectively, while amorphous features (trash, standing water) proved challenging.
• Optimizing frame selection and image resolution improved overall model performance.

Conclusions:

• ML combined with SV can automate environmental risk identification for health issues in informal settlements.
• Training data requirements and success rates for risk identification may vary geographically.
• Best practices for data collection, model training, and performance evaluation were identified, paving the way for automatic health risk mapping tools.