This paper shows the efficacy of a novel urban categorization framework based on deep learning, and a novel categorization method customized for cities in the global south. The proposed categorization method assesses urban space broadly on two dimensions—the states of urbanization and the architectural form of the units observed. This paper shows how the sixteen sub-categories can be used by state-of-the-art deep learning modules (fully convolutional network FCN-8, U-Net, and DeepLabv3+) to categorize formal and informal urban areas in seven urban cities in the developing world—Dhaka, Nairobi, Jakarta, Guangzhou, Mumbai, Cairo, and Lima. Firstly, an expert visually annotated and categorized 50 × 50 km Google Earth images of the cities. Each urban space was divided into four socioeconomic categories: (1) highly informal area; (2) moderately informal area; (3) moderately formal area, and (4) highly formal area. Then, three models mentioned above were used to categorize urban spaces. Image encompassing $70\%$ of the urban space was used to train the models, and the remaining $30\%$ was used for testing and validation of each city. The DeepLabv3+ model can segment the test part with an average accuracy of $90.0\%$ for Dhaka, $91.5\%$ for Nairobi, $94.75\%$ for Jakarta, $82.0\%$ for Guangzhou city, $94.25\%$ for Mumbai, $91.75\%$ for Cairo, and $96.75\%$ for Lima. These results are the best for the DeepLabv3+ model among all. Thus, DeepLabv3+ shows an overall high accuracy level for most of the measuring parameters for all cities, making it highly scalable, readily usable to understand the cities’ current conditions, forecast land use growth, and other computational modeling tasks. Therefore, the proposed categorization method is also suited for real-time socioeconomic comparative analysis among cities, making it an essential tool for the policymakers to plan future sustainable urban spaces.

The study explores deep learning to perform direct semantic segmentation of bathymetric lidar points to improve bathymetry mapping. Focusing on river bathymetry, the goal is to accurately and simultaneously classify points on the benthic layer, water surface, and ground near riverbanks. These classifications are then used to apply depth correction to all points within the water column. The study aimed to classify the scene into four classes: river surface, riverbed, ground, and other (for points outside of those three classes), focusing on the river surface and riverbed classes. To achieve this, PointCNN, a convolutional neural network model adept at handling unorganized and unstructured data in 3D space was implemented. The model was trained with airborne bathymetric lidar data from the Swan River in Montana and the Eel River in California. The model was tested on the Snake River in Wyoming to evaluate its performance. These diverse bathymetric datasets across the United States provided a solid foundation for the model’s robust testing. The results were strong for river surface classification, achieving an Intersection over Union of (0.89) and a Kappa coefficient of (0.92), indicating high reliability and minimal errors. The riverbed classification also showed an IoU of (0.7) and a slightly higher Kappa score of (0.76). Depth correction was then performed on riverbed points, proportional to the calculated depth from a surface model formed by Delaunay triangulation of ground and river surface points. The automated process performs significantly faster than traditional manual classification and depth correction processes, saving time and expense. Finally, corrected depths were quantitatively validated by comparing with independent Acoustic Doppler Current Profiler measurements from the Snake River, obtaining a mean depth error of 2 cm and an Root mean square error of 16 cm. These validation results show the reliability and accuracy of the proposed automated bathymetric depth correction workflow.