Generating Any Changes in the Noise Domain.

Area of Science:

• Earth Observation
• Computer Vision
• Artificial Intelligence

Background:

• Change detection in Earth observation requires large annotated datasets, which are costly to produce.
• Generative models can synthesize data, but struggle with temporal coherence and realistic semantic changes in image pairs.
• Current methods using heuristic rules or text prompts for change simulation lack diversity and spatial consistency.

Purpose of the Study:

• To propose Noise2Change, a framework for simulating realistic changes directly in the noise domain of diffusion models.
• To generate temporally aligned and semantically coherent pre- and post-change image pairs for training change detection models.
• To address the trade-off between realism and consistency in synthetic change detection data.

Main Methods:

• Utilizing the noise space of diffusion models for spatial controllability and generative capacity.
• Manipulating the semantic composition of initial noise to guide the diffusion process.
• Employing a discrete diffusion model to extract semantics and optimize noise for intended changes.
• Generating pre- and post-change label maps with natural transitions and refining them for image generation.

Main Results:

• Noise2Change successfully generates structurally consistent pre- and post-change images with strong temporal alignment.
• The framework produces diverse change types across various scenarios, enhancing realism.
• Experiments show superior performance compared to existing generative approaches on multiple change detection tasks.

Conclusions:

• Noise2Change offers a powerful new approach for synthesizing high-quality training data for change detection.
• The method effectively overcomes limitations of previous generative techniques by leveraging noise manipulation.
• This framework has the potential to significantly advance Earth observation capabilities through improved change detection models.