Terrestrial

Terrestrial uses stable diffusion to visualize deforestation data, years into the future; answering what our future ecosystems could look like, and offering a new way to visualize terrestrial data.

Comment

Inspiration

What if AI could show us the future of our planet?

Terrestrial is built on the idea of using generative AI to see into the future of terrestrial ecosystems.

We were looking to solve a problem at the intersection between biodiversity and generative AI that has high potential impact, with practical and real world software applications.

We came across the problem that there is currently no way for researchers, policy-makers, or the public to visualize data (at high resolution and in the form of a photo) for what our planet might look like many years into the future if deforestation continues at current rates.

re: United Nations Sustainable Development Goal 15: Protect, restore and promote sustainable use of terrestrial ecosystems, sustainably manage forests, combat desertification, and halt and reverse land degradation and halt biodiversity loss.

What it does

Terrestrial uses stable diffusion to visualize deforestation data, years into the future; answering what our future ecosystems could look like, and offering a new way to visualize terrestrial data.

The application takes raw GIS deforestation data, extrapolates the data, and sends a uniform optimized query to a stable diffusion model; which in turn generates an image.

The application is useful to researchers, policy-makers, activists, and the general public who are looking to proactively: gain insight, and promote the protection and restoration of terrestrial ecosystems.

How we built it

We use GIS data, extrapolate it, create a uniform optimized prompt based on the modeled data, and send it to a stable diffusion model to generate a unique image.

We created a React web app styled with Tailwind and bundled with Vite. The design of the app was prototyped in Figma.

Our backend is written in Python and Typescript and consists of an API with a queueing system. Using crafted prompts and a DALL•E 2 Python library, we are able to use AI to generate unique images.

Through the API, the frontend fetches generated images from the backend.

We also use container services.

Example of prompt: “[country] forest [x] years into the future with [y%] less foliage due to deforestation”

Challenges we ran into

Brainstorming a creative idea within the scope of the topic Time management/pressure cors issues Underestimated how much time it would take to build

Accomplishments that we're proud of

An organized codebase! A very simple and easy to use UI Our first time successfully productionizing a stable diffusion model. Successfully using docker

What we learned

Focusing on a problem is extremely important. Focusing on a problem is an easy step to skip. How to optimize prompts for stable diffusion models. How to use vite for bundling. How to properly turn a csv into a json :) Focus on building core functionality Collaboration - split work among the team efficiently (frontend, backend, product design, copywriting, research) That stable diffusion models could potentially be the future for creative imagery work but also data visualization.

What's next for Terrestrial

In the future, the product can be modified to use other data sets, and deep learning (example features: agricultural expansion, climate, supply chains) for more accurate predictions. It is possible an API could be created for bulk queries. This product has the potential to be venture backed if greater accuracy and public interest is garnered.

Share this project:

Updates