This study proposes a multi-task learning framework with an attention mechanism for semantic and instance segmentation in coastal urban spatial perception. The framework integrates an attention mechanism to enhance the model's ability to extract and match urban spatial attributes. The model is tested on urban street scenes, and the results show that it achieves high prediction accuracy for urban spatial attributes such as beauty, boredom, depression, liveliness, safety, and richness, with accuracy rates of 79.54%, 78.62%, 79.68%, 77.42%, 78.45%, and 76.98%, respectively. Additionally, the model achieves high accuracy in segmenting urban spatial scene elements such as roads, walls, sky, vehicles, and buildings, with accuracy rates of 95.4, 94.8, 96.2, 92.1, and 96.7, respectively. The model's performance is compared with other models, and it is found to have higher recognition accuracy for urban spatial buildings. The study also evaluates the model's performance in terms of user satisfaction, and the results show that the proposed model has higher user satisfaction than other methods. The study contributes to the field of urban planning and design by providing scientific, data-driven support for creating more accurate and comprehensive guidance for urban planning and design. It also helps to create a more livable, sustainable, and human-centered urban environment. However, the model requires a large amount of data for training and has a higher computing power requirement, which may limit its practical application.This study proposes a multi-task learning framework with an attention mechanism for semantic and instance segmentation in coastal urban spatial perception. The framework integrates an attention mechanism to enhance the model's ability to extract and match urban spatial attributes. The model is tested on urban street scenes, and the results show that it achieves high prediction accuracy for urban spatial attributes such as beauty, boredom, depression, liveliness, safety, and richness, with accuracy rates of 79.54%, 78.62%, 79.68%, 77.42%, 78.45%, and 76.98%, respectively. Additionally, the model achieves high accuracy in segmenting urban spatial scene elements such as roads, walls, sky, vehicles, and buildings, with accuracy rates of 95.4, 94.8, 96.2, 92.1, and 96.7, respectively. The model's performance is compared with other models, and it is found to have higher recognition accuracy for urban spatial buildings. The study also evaluates the model's performance in terms of user satisfaction, and the results show that the proposed model has higher user satisfaction than other methods. The study contributes to the field of urban planning and design by providing scientific, data-driven support for creating more accurate and comprehensive guidance for urban planning and design. It also helps to create a more livable, sustainable, and human-centered urban environment. However, the model requires a large amount of data for training and has a higher computing power requirement, which may limit its practical application.