DPG-CD: Depth-Prior-Guided Cross-Modal Joint 2D-3D Change Detection

Researchers introduce DPG-CD, a deep learning framework that detects both 2D semantic and 3D structural changes in urban environments by fusing multi-temporal satellite imagery with Digital Surface Model data. The method addresses the challenge of combining different data modalities to enable high-frequency urban monitoring and disaster assessment without requiring expensive frequent 3D data collection.

This research addresses a critical gap in urban monitoring technology where traditional change detection methods struggle to capture both appearance and height variations simultaneously. Urban environments undergo continuous transformation through building construction, demolition, and structural modifications—changes that satellite imagery alone cannot fully capture. By combining pre-event 3D elevation data with post-event 2D imagery, the DPG-CD framework enables efficient monitoring without the prohibitive costs of frequent LiDAR surveys or drone acquisitions.

The technical innovation centers on solving the spectral-geometric representation mismatch between imagery and Digital Surface Models. The framework employs a depth-prior guidance mechanism that essentially translates 2D imagery into a format compatible with 3D elevation data, allowing the model to understand both modalities in a unified feature space. This cross-modal fusion approach leverages multi-task learning, where auxiliary predictions of DSM values improve overall accuracy by enforcing structural consistency.

For urban planning, disaster management, and infrastructure monitoring sectors, this advancement significantly reduces operational costs while improving detection reliability. Emergency responders can now assess earthquake or flood damage more comprehensively by understanding both what changed visually and how building heights were affected. The creation of the NYC-MMCD dataset provides researchers with real-world urban data for advancing this field further.

Future developments likely involve real-time processing pipelines, integration with existing geospatial platforms, and expansion to additional urban indicators. The methodology could extend beyond height changes to other geometric properties, potentially creating more comprehensive urban digital twins for planning and risk management applications.

• →DPG-CD combines 2D satellite imagery with 3D elevation data to detect both semantic and structural urban changes efficiently
• →Depth-prior guidance bridges the representation gap between different data modalities, improving cross-modal fusion accuracy
• →The framework reduces reliance on expensive frequent 3D data collection while enabling high-frequency urban monitoring
• →Multi-task learning with auxiliary DSM prediction improves structural consistency and height estimation accuracy
• →Results demonstrate superior performance on public datasets with applications for disaster assessment and emergency response