Generative AI For Synthetic Height Bands on Ryzen AI PCs

Project Title: Generative AI For Synthetic Height Bands

You can try it out in streamlit

Project Overview:

My project, "Generative AI For Synthetic Height Bands," focuses on creating synthetic height data for geospatial analysis. This innovative solution addresses the high costs and limitations associated with traditional methods of obtaining Digital Elevation Models (DEMs) and Digital Surface Models (DSMs).

Motivation:

I embarked on this project to revolutionize the field of geospatial analysis by making height data more accessible and cost-effective. Traditional methods of acquiring DEMs and DSMs, such as field surveys and satellite data acquisition, are expensive and time-consuming. By leveraging generative AI, I aim to provide a more efficient and precise alternative.

How It Works:

Data Collection: I started by collecting available DEM and DSM data to train generative AI models. This includes satellite imagery, and other geospatial data sources.

Model Training: Our AI models are trained on this extensive dataset, learning the patterns and features that characterize height variations in different terrains.

Synthetic Generation: Once trained, the AI models can generate synthetic height bands for any given area. The generated data is both accurate and reliable, closely matching the quality of traditional DEMs and DSMs.

Validation and Refinement: The synthetic height bands are validated against real-world data to ensure accuracy. Continuous refinement of the models ensures improvements in data quality over time.

Future Work: Enhancing Satellite Image Resolution Using Generative AI

The original scope of this project was to leverage generative AI techniques to improve the resolution of open-source satellite images. While significant progress has been made, several challenges were encountered that have shaped the direction of future work. This section outlines the envisioned steps to address these challenges and achieve the project's goals.

Current Progress:

The project successfully integrated various machine learning models to process and enhance satellite images. Key accomplishments include:

Model Training: Training regression models to predict higher resolution from lower resolution inputs.

Data Handling: Developing efficient methods to read, preprocess, and scale large satellite image datasets.

Post-Processing: Implementing basic post-processing techniques such as sharpening and deblurring to enhance image clarity.

Challenges:

Despite these advancements, several challenges were faced during the project:

Model Limitations: Traditional regression models struggled to capture the complex patterns needed for significant resolution enhancement.

Computational Resources: Processing high-resolution satellite images requires substantial computational power, which limited the ability to train and test more sophisticated models.

Blurriness and Artifacts: Post-processing techniques, while helpful, were not sufficient to completely eliminate blurriness and artifacts introduced during image resampling and prediction steps.

Integration Complexity: Combining various techniques (e.g., machine learning, image processing) into a cohesive pipeline proved to be more complex than anticipated.

Future Work

To overcome these challenges and achieve the project's goals, the following steps are proposed for future work:

Advanced Generative Models:
Super-Resolution GANs (SRGANs): Explore the use of Super-Resolution Generative Adversarial Networks (SRGANs) to generate high-resolution satellite images from low-resolution inputs. SRGANs have shown promising results in enhancing image resolution by learning intricate details through adversarial training.