Publication

Abstract : As cities grow, they have more problems with traffic jams and keeping an eye on their infrastructure. This is why smart city development needs intelligent urban management systems. This paper presents a deep learning framework intended to tackle two essential urban analytics challenges: forecasting traffic flow and segmenting building footprints. The first part of the framework uses a Spatio-Temporal Graph Neural Network (ST-GNN) to understand how traffic changes over time, using data from the METR-LA traffic speed dataset. The dataset is made up of many sensors that are spread out over road networks. Graph structures are used to show how the sensors are related to each other in space. The model combines multi-head graph attention mechanisms with gated temporal convolutional networks to effectively capture patterns in traffic data over time and space. The model predicts traffic conditions for the next 12 time steps using the Huber loss function. Its goal is to minimize the normalized Mean Absolute Error (MAE). The second part uses SpaceNet satellite images and an optimized U-Net architecture to segment building footprints. The encoder-decoder structure with skip connections lets classify pixels accurately while keeping spatial details and high-level semantic features. This method makes segmentation work better, even in tough urban settings like shadows, changing light, and dense building structures. All of these parts work together to make an integrated framework that supports AI-driven urban analysis. This helps with smart city planning, monitoring infrastructure, and developing digital twins.