Residential scene classification for gridded population sampling in developing countries using deep convolutional...

Area of Science:

• Geographic Information Systems (GIS)
• Machine Learning
• Remote Sensing

Background:

• Surveying in low- and middle-income countries faces challenges due to incomplete sampling frames and outdated census data.
• Geosampling, a probability-based method, uses GIS to create manageable area units for sampling.
• Manual classification of aerial images for residential areas is labor-intensive and prone to errors.

Purpose of the Study:

• To develop a deep learning model for classifying residential areas from aerial images.
• To reduce manual labor and improve the accuracy of sample frame construction in geosampling.
• To eliminate the need for simplifying assumptions in calculating sampling weights.

Main Methods:

• Development of a deep learning classification model.
• Training the model on aerial images to predict residential versus nonresidential grid cells.
• Utilizing geographic information system (GIS) tools to partition areas into grid cells.

Main Results:

• The deep learning model achieved high accuracy comparable to human baselines (94.5% in Nigeria, 96.4% in Guatemala).
• The model outperformed other machine learning approaches using crowdsourced or remote-sensed data.
• The approach is effective even with limited training data in new areas.

Conclusions:

• Deep learning on satellite images offers a novel method for constructing sample frames by identifying residential areas.
• This methodology can reduce the annotation burden in sample frame construction, providing quality comparable to human analysts.
• The approach enhances the timeliness, flexibility, and cost-effectiveness of gridded population sampling.