Methodologies for training fully connected neural networks (FCNN/MLP) using both synthetic and real data to address scenario-based spatial analysis of urban environments, exemplified by the city of Novosibirsk, are presented in the paper. A comparative evaluation was conducted between two neural network architectures: a unified model trained on comprehensive scenario-based patterns and a composite model comprising neural networks, each dedicated to an individual elementary scenario. The findings indicate that pre-training on large-scale synthesized datasets guarantees high model accuracy (97–99 %). However, direct application of the unified model to real-world data results in its substantial decrease in accuracy to 60 %, thereby necessitating further fine-tuning. In contrast, composite models demonstrate superior accuracy and robustness across both synthesized and real datasets, as well as greater adaptability to changes in scenario-based patterns. Nonetheless, the unified model exhibits significantly faster prediction speeds compared to the composite one, which is a critical advantage for operational and time-sensitive analyses. Subsequent fine-tuning of the models on real data enhanced their accuracy to 96–97 % while markedly reducing training time. The model outputs showed strong consistency with results produced by specialized software tools. The authors propose a hybrid approach, combining initial pre-training of the models on synthesized data with subsequent fine-tuning on real data, which provides an effective and reliable strategy for improving model performance. The study further outlines promising avenues for future research. Overall, the results underscore the significant practical relevance and potential of neural network methodologies for comprehensive spatial analysis and forecasting urban environment quality