# DDPM-CD: Denoising Diffusion Probabilistic Models as Feature Extractors for Change Detection

**Authors:** Wele Gedara Chaminda Bandara, Nithin Gopalakrishnan Nair, Vishal M. Patel  
**arXiv:** [2206.11892](https://arxiv.org/abs/2206.11892)  
**Date:** 2022-06-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.11892v3/

## Key points

- Understanding dynamics of the planet's surface is crucial for monitoring environmental changes, evaluating human impact, predicting future trends, and supporting decision-making in remote sensing change detection.
- A novel approach leverages off-the-shelf, unlabeled remote sensing images by pre-training a Denoising Diffusion Probabilistic Model (DDPM) for image synthesis.
- DDPMs convert training images into a Gaussian distribution using a Markov chain to generate diverse samples closer to the training distribution during inference.
- The proposed method involves fine-tuning a lightweight change classifier using feature representations produced by the pre-trained DDPM alongside change labels.
- Experiments on datasets like LEVIR-CD, WHU-CD, DSIFN-CD, and CDD show that DDPM-CD outperforms existing state-of-the-art methods in terms of F1 score, IoU, and overall accuracy.
- Pre-trained DDPM plays a crucial role as a feature extractor for downstream applications in change detection tasks.
- The best configuration utilizes multi-timestep features at timesteps 50, 100, and 400 for superior performance.

## Summary

In the field of remote sensing change detection, understanding the dynamics of our planet's surface is essential for monitoring environmental changes, evaluating human impact, predicting future trends, and supporting decision-making. A novel approach introduced in this work leverages off-the-shelf, unlabeled remote sensing images by pre-training a Denoising Diffusion Probabilistic Model (DDPM), a generative model used in image synthesis. DDPMs learn the training data distribution by converting training images into a Gaussian distribution using a Markov chain, enabling them to generate diverse samples closer to the training distribution during inference. The proposed method involves fine-tuning a lightweight change classifier using feature representations produced by the pre-trained DDPM alongside change labels. Experiments conducted on various datasets such as LEVIR-CD, WHU-CD, DSIFN-CD, and CDD demonstrate that the DDPM-CD method outperforms existing state-of-the-art change detection methods in terms of F1 score, IoU, and overall accuracy. This highlights the crucial role of pre-trained DDPM as a feature extractor for downstream applications. Quantitative comparison results show that the proposed method achieves superior performance compared to state-of-the-art methods based on F1 score, IoU, and OA scores. The best configuration utilizes multi-timestep features at timesteps 50, 100, and 400. Further details on performance variations based on timestep usage are discussed in Section VI. Overall,this work showcases the effectiveness of utilizing pre-trained DDPM as a feature extractor for change detection tasks and provides valuable insights into improving remote sensing analysis techniques. The code and pre-trained models are available for further exploration at https://github.com/wgcban/ddpm-cd.