Applying State-of-the-Art Deep-Learning Methods to Classify Urban Cities of the Developing World.

Area of Science:

• Urban studies
• Geospatial analysis
• Artificial intelligence

Background:

• Cities in the Global South face unique challenges in urban development and require tailored analytical tools.
• Existing urban categorization methods may not adequately capture the complexities of formal and informal settlements in developing regions.

Purpose of the Study:

• To develop and validate a novel deep learning-based urban categorization framework for cities in the Global South.
• To assess the efficacy of state-of-the-art deep learning models in segmenting formal and informal urban areas.

Main Methods:

• A novel categorization method was developed, assessing urban space based on urbanization states and architectural forms, defining sixteen sub-categories.
• Expert visual annotation of Google Earth images was performed for seven Global South cities: Dhaka, Nairobi, Jakarta, Guangzhou, Mumbai, Cairo, and Lima.
• Three deep learning models (FCN-8, U-Net, DeepLabv3+) were trained and validated using segmented urban space data, with DeepLabv3+ showing the highest accuracy.

Main Results:

• The DeepLabv3+ model achieved high segmentation accuracy across all tested cities, with results ranging from 82.0% (Guangzhou) to 96.75% (Lima).
• The model demonstrated consistent high performance, indicating its scalability and usability for understanding current urban conditions and forecasting land use.
• The categorization method proved effective in distinguishing between formal and informal urban areas, crucial for socioeconomic analysis.

Conclusions:

• The proposed deep learning framework and categorization method are effective for analyzing urban environments in the Global South.
• DeepLabv3+ is a highly accurate and scalable model for urban space segmentation, supporting real-time socioeconomic comparative analysis.
• This approach provides policymakers with an essential tool for planning sustainable urban futures.