Abstract
This study introduces a novel Remote Sensing (RS) Urban Prediction (UP) task focused on future urban planning, which aims to forecast urban layouts by utilizing information from existing urban layouts and planned change maps. To address the proposed RS UP task, we propose UP-Diff, which leverages a Latent Diffusion Model (LDM) to capture position-aware embeddings of pre-change urban layouts and planned change maps. In specific, the trainable cross-attention layers within UP-Diff's iterative diffusion modules enable the model to dynamically highlight crucial regions for targeted modifications. By utilizing our UP-Diff, designers can effectively refine and adjust future urban city plans by making modifications to the change maps in a dynamic and adaptive manner. Compared with conventional RS Change Detection (CD) methods, the proposed UP-Diff for the RS UP task avoids the requirement of paired pre-change and post-change images, which enhances the practical usage in city development. Experimental results on LEVIR-CD and SYSU-CD datasets show UP-Diff's ability to accurately predict future urban layouts with high fidelity, demonstrating its potential for urban planning.
Introduction: From Detection to Prediction
In the realm of urban planning and management, the ability to visualize the impact of new developments is paramount. Traditionally, Remote Sensing (RS) research has focused heavily on Change Detection (CD)—a retrospective task that identifies differences between two existing satellite images taken at different times. While useful for monitoring, CD is inherently limited because it requires "paired" historical data and cannot assist in visualizing future, unbuilt projects.
To bridge this gap, this research introduces the task of Urban Prediction (UP). Unlike CD, which takes two images to produce a change mask, UP takes a single "pre-change" image and a "planned change map" (a blueprint mask) to generate a realistic "post-change" image. This shift allows planners to perform "what-if" simulations, visualizing how a specific building or road layout will integrate with the existing environment.
The UP-Diff Framework
The proposed solution, UP-Diff, is a generative framework built upon Latent Diffusion Models (LDMs). The core challenge in UP is ensuring that the generated image strictly adheres to the provided planning map while maintaining the visual consistency of the original satellite imagery.
UP-Diff addresses this through several key architectural innovations:
- Latent Diffusion Foundation: By operating in a compressed latent space using a pre-trained autoencoder, the model achieves high-resolution generation with computational efficiency.
- Condition Embedding: The model is conditioned on three inputs: textual prompts (via CLIP), and spatial layouts encoded by a trainable ConvNeXt model that processes the concatenation of the pre-change image and the change mask into "position-aware embeddings."
- Gated Cross-Attention: Within the iterative diffusion process, the model employs a gated cross-attention mechanism. This allows the UNet to dynamically prioritize the regions specified in the change map, ensuring that new structures are rendered precisely where planned, while the surrounding context (forests, water, existing buildings) remains unchanged.
Experimental Evaluation
The authors evaluated UP-Diff on two major benchmarks adapted for the UP task: LEVIR-CD (high-resolution suburban building changes) and SYSU-CD (diverse urban changes in Hong Kong). The results demonstrate that UP-Diff significantly outperforms traditional generative approaches like GANs and standard VAEs across two critical dimensions:
- Visual Realism: On the LEVIR-CD dataset, UP-Diff achieved a Fréchet Inception Distance (FID) of 117.79, outperforming the next best baseline (DDPM) by a wide margin. This indicates that the generated images are more indistinguishable from real satellite photos.
- Plan Adherence: To verify if the model actually followed the blueprints, the generated images were processed by a separate, pre-trained Change Detection model (DMINet). UP-Diff achieved an Intersection over Union (IoU) of 73.35% on LEVIR-CD, confirming that the generated buildings align accurately with the input masks.
Conclusion
UP-Diff represents a significant step forward in "generative AI for GIS." By moving from retrospective detection to prospective prediction, it provides urban designers with a powerful tool for dynamic and adaptive city planning. By leveraging the strong priors of models pre-trained on natural images and refining them with gated spatial controls, UP-Diff produces high-fidelity visualizations that respect both the geographic context and the human-designed blueprint. Future work will explore pluralistic predictions to offer multiple architectural interpretations of a single plan.
Summary updated