Rapid urbanization and escalating environmental challenges in Latin American cities demand innovative methods for urban analysis and management. In this context, advanced computer vision techniques offer powerful tools for enhancing urban safety, planning, and environmental monitoring. This study explores the use of Fully Convolutional Networks (FCNs) for semantic segmentation of urban scenes in Guayaquil, Ecuador. The objective is to develop an accurate, efficient model for high-resolution pixel-level classification of key urban features, including vehicles, building frontage, vegetation, and pedestrian areas, to inform data-driven urban planning and safety strategies. The proposed model integrates FCNs with multitask learning and attention mechanisms to improve segmentation accuracy. Multitask learning enables the simultaneous identification of multiple urban elements, enhancing generalization, while attention mechanisms refine boundary detection by focusing on relevant image regions. A curated dataset of Guayaquil’s urban scenes was used for training and evaluation, with preprocessing techniques such as data augmentation and normalization employed to improve robustness. Results show segmentation accuracies of 93% on the training set and 92.6% on the test set, outperforming traditional CNN-based methods. The approach excels in detecting street-level features and visual pollution, providing actionable insights for urban safety and environmental assessments. This research highlights the value of region-specific deep learning algorithms tailored to the visual complexity of Latin American cities. By leveraging FCNs for automated, scalable urban analysis, the study can contribute to sustainable development, enhanced public safety, and improved quality of life. These findings can support policymakers, planners, and researchers in building smarter, safer urban environments across the region.