Digitizing Karachi's Decades-Old Cadastral Maps: Leveraging Unsupervised Machine Learning and GEOBIA for Digitization

In urban planning, land-use change is paramount for ensuring sustainable urban ecosystems. Monitoring, analyzing, and quantifying land use change is crucial to making statistical inferences and predicting the economic, environmental, and societal impacts of urban expansion. Recent technologies have enabled robust monitoring, recording, and documenting of spatio-temporal trends. When historical data remain nondigital, integrating modern technologies with traditional paper-based town maps becomes invaluable for digitization. Despite significant efforts in this field, little exploration has been done of the potential of Geographic Object-Based Image Analysis (GOBIA) for digitizing paper-based cadastral maps. This study introduces an innovative approach using unsupervised learning algorithms, K-means and Gaussian Mixture Models (GMM), in conjunction with GEOBIA techniques, to accurately extract land parcels from decades-old cadastral maps of Karachi, Pakistan. Initially, the maps were georeferenced using ArcGIS software, and unsupervised machine-learning algorithms were applied to preprocessed scanned images. Both clustering algorithms were evaluated based on key performance metrics, such as precision, recall, and F1 scores. The experimental results indicated that both algorithms performed well, with GMM slightly outperforming K-means in all aspects. GMM achieved 0.87 precision and recall and 0.86 F1 score of 0.86, while K-means achieved 0.82 precision, 0.78 recall, and 0.78 F1 score. Finally, unwanted features were removed by implementing a geometric criterion based on feature size and shape. This methodology effectively distinguishes between adjoining land parcels and ensures precise extraction of cadastral boundaries and land parcels, providing a reliable foundation for urban research and modeling.