Changen2: Multi-Temporal Remote Sensing Generative Change Foundation Model.

Area of Science:

• Earth Science
• Computer Science
• Artificial Intelligence

Background:

• Deep vision models advance Earth's surface temporal dynamics understanding.
• Training these models requires extensive labeled multi-temporal remote sensing images.
• Data collection, preprocessing, and annotation are costly and labor-intensive.

Purpose of the Study:

• To present scalable multi-temporal change data generators using generative models.
• To alleviate data acquisition challenges in remote sensing.
• To develop a method for automatic generation of training data for change detection.

Main Methods:

• Developed a generative probabilistic change model (GPCM) to simulate stochastic change processes.
• Introduced Changen2, a diffusion transformer-based GPCM for generating time series of remote sensing images and labels.
• Employed self-supervision for large-scale training of the generative change foundation model.

Main Results:

• Changen2 demonstrates superior spatiotemporal scalability in data generation, producing high-resolution time series from lower-resolution inputs.
• Pre-trained Changen2 models show strong zero-shot change detection capabilities and excellent transferability across diverse change detection tasks.
• The model narrows the performance gap with fully supervised methods on benchmark datasets.

Conclusions:

• Changen2 offers a cost-effective and automatic solution for generating multi-temporal remote sensing data.
• The generative change foundation model synthesizes crucial training data, enhancing task-specific models.
• This approach significantly improves the efficiency and applicability of deep learning for Earth surface change analysis.