Simultaneous Super-resolution and Depth Estimation for Satellite Images Based on Diffusion Model

Area of Science:

• Earth Observation
• Computer Vision
• Remote Sensing

Background:

• Satellite imagery offers large-scale Earth surface observation, crucial for applications like forestry and crop monitoring.
• Current methods for 3D landscape modeling from satellite data are limited, especially using high-resolution RGB images, due to sparse LiDAR data and low resolution of standard satellite images.
• Existing research has not fully explored generating detailed 3D models from enhanced satellite RGB images.

Purpose of the Study:

• To develop a novel methodology for enhancing satellite image resolution and performing depth estimation.
• To enable accurate 3D surface reconstruction and detailed landscape modeling using improved satellite data.
• To leverage the generative power of diffusion models for simultaneous super-resolution and depth estimation.

Main Methods:

• A simultaneous diffusion model learning framework was developed to train models for both super-resolution (SR) and depth estimation (DE).
• The framework enhances low-resolution satellite RGB images to generate super-resolution versions.
• Depth maps were generated corresponding to the super-resolution images, facilitating 3D reconstruction.

Main Results:

• The proposed methodology effectively enhances satellite image resolution.
• Accurate depth estimation was achieved using the developed diffusion model framework.
• Detailed 3D surface reconstruction models were successfully generated from the enhanced images and depth maps.
• Evaluations on multiple satellite datasets confirmed the effectiveness of the approach for both SR and DE tasks.

Conclusions:

• The developed simultaneous diffusion model learning framework significantly improves satellite image resolution and depth estimation capabilities.
• This approach enables detailed 3D landscape reconstruction from readily available satellite RGB images.
• The methodology offers a promising advancement for remote sensing and Earth observation applications requiring high-fidelity 3D models.