UniChange: A New Paradigm for Unified Change Detection with Multimodal Large Models

The UniChange framework proposed by the HLT Lab of Nankai University for the first time introduces multimodal large language models (MLLMs) into the field of change detection. It achieves unified change detection capabilities across datasets and sensors, solving the generalization challenges of traditional methods and providing a breakthrough unified solution for this field.

What is Change Detection

Change detection automatically identifies surface changes by comparing remote sensing images of the same area at different times, and is applied in urban planning, environmental protection, agriculture, disaster response, and other fields.

Dilemmas of Traditional Methods

1. Data Heterogeneity: Traditional models only process data from specific sensors (e.g., optical, SAR) and are difficult to generalize across sensors;
2. Diverse Change Types: Models need to be designed separately for each type of change (e.g., new building construction, vegetation growth);
3. Scarce Annotation Data: The cost of paired temporal images and pixel-level annotations is high, limiting model scale and generalization.

Core Innovation: Introducing Multimodal Large Language Models

Model change detection as a visual-language understanding task: The visual encoder extracts features from bi-temporal images, uses the semantic understanding ability of MLLMs to analyze changes, and leverages pre-trained knowledge to improve generalization.

Unified Framework Design

• Data Level: Supports multimodal data such as optical, SAR, and multispectral, and learns cross-modal shared representations;
• Task Level: Outputs pixel-level change masks + natural language descriptions, enabling precise localization and semantic understanding;
• Knowledge Level: Uses MLLM pre-trained knowledge and has zero-shot/few-shot learning capabilities.

Visual Encoding and Alignment

A flexible encoding strategy adapts to images from different sensors. Through contrastive learning, it aligns visual features with the semantic space of the language model, laying the foundation for MLLMs to understand visual information.

Temporal Feature Fusion

A temporal fusion module using attention mechanisms adaptively focuses on changed regions, suppresses interference from unchanged regions, and improves detection accuracy and robustness.

Language Decoding and Output

Fused features are sent to MLLM for decoding, generating change masks and natural language descriptions, and supporting multi-granularity outputs (option to output only masks or both masks and text descriptions).

Cross-Dataset Generalization Ability

It performs excellently on optical datasets such as LEVIR-CD and WHU-CD, as well as SAR datasets. When applied across datasets, it maintains high accuracy and reduces dependence on specific annotated data.

Cross-Sensor Adaptability

After training on optical images, it can be directly applied to SAR image detection without additional SAR data training, solving the problem of incomplete sensor data in real scenarios.

Accuracy of Change Description

It can generate accurate and coherent natural language descriptions, explaining the type, location, and degree of changes, which is suitable for manual review or report generation scenarios.

• Urban Dynamic Monitoring: Automatically identifies new buildings, road construction, etc., providing decision support for urban planning;
• Precision Agricultural Management: Monitors crop growth and pest/disease areas, optimizing resource input;
• Environmental Protection: Monitors deforestation and wetland degradation, evaluating the effects of ecological policies;
• Disaster Response: Compares pre- and post-disaster images to quickly identify affected areas; cross-sensor capability can handle cloud cover (using SAR data).

Technical Insights

It verifies the feasibility of introducing large language models into remote sensing analysis, which can be extended to other remote sensing tasks such as object detection and land cover classification.

Future Outlook

1. Multimodal Fusion: Fuse more data sources such as LiDAR and geographic vectors;
2. Open World Detection: Leverage the open vocabulary capability of MLLMs to identify new change types not seen during training.

Conclusion

UniChange achieves a leap from pixel classification to semantic-level change cognition, and will play an important role in fields such as Earth observation and resource management.