Classifying Informal Settlements in Bogotá Using Graph Neural Networks:

This seminar showcased a master's thesis project focused on the detection and classification of informal settlements in Bogotá through the application of Graph Neural Networks (GNNs). The project originated as a subcomponent of a consulting engagement that explored the use of satellite imagery to identify urban features. As an innovative contribution, this work proposed using GNNs rather than traditional convolutional neural networks.

The core challenge addressed was the difficulty in estimating the degree of informality in urban neighborhoods, particularly regarding land tenure legality. To tackle this, the authors developed a two-stage modeling approach:

Model 1 – Neighborhood-level classification: Based on geographic and satellite data, a GNN was trained to predict whether a neighborhood is formal or informal, even in the absence of prior labels.

Model 2 – Informality prediction at the UPZ level: Using the results of the first model and socio-demographic data (including education level, household composition, emergency call data, and internet access), the team estimated the percentage of informal areas within each Planning Zone (UPZ).

The first model achieved an AUC of 88.9%, indicating strong classification capability. The second model, with an R² of approximately 48%, showed reasonable predictive performance considering the noise carried over from the initial classification. The study highlights the value of GNNs in augmenting traditional census data and supporting urban policy design. Limitations, potential improvements, and practical implications for other cities in Colombia and beyond were also discussed.