Identifying patterns in urban housing density in developing countries using convolutional networks and satellite...

Area of Science:

• Remote Sensing and Geospatial Analysis
• Machine Learning and Artificial Intelligence
• Urban Planning and Development Studies

Background:

• Growing use of Deep Neural Networks (DNNs) for remote sensing scene analysis.
• Lack of representative public datasets for developing countries hinders algorithm benchmarking.
• Current low-level semantic scene classification lacks broad applicability.

Purpose of the Study:

• To address the unrepresentativeness of existing datasets and limitations of low-level scene classification.
• To apply Convolutional Neural Networks (CNNs) for high-level scene image classification of urban housing density in developing countries.
• To propose a novel method for quantifying spatial extent of urban housing classes and settlement patterns.

Main Methods:

• Developed an end-to-end model training workflow for CNN-based scene image classification.
• Proposed a quantification method based on the ratio of area covered by a housing class to the total area of all classes.
• Implemented the quantification method using grid counts and validated results against OpenStreetMap building density data.

Main Results:

• Achieved scene image classification results comparable to state-of-the-art, even on challenging classes.
• Validated the proposed quantification method against external building density data.
• Contributed a new satellite scene image dataset representative of urban housing in developing countries.

Conclusions:

• The proposed CNN approach and quantification method effectively analyze urban housing density in developing countries.
• The new dataset enhances the representativeness of remote sensing scene analysis for developing regions.
• Findings offer valuable insights into settlement patterns and support urban planning initiatives.