One of our teammembers, Raghav, is the founder of Eagles For Environment (EFE), a nonprofit organization that tackles Blackland prairie restoration in North Texas. Throughout high school, EFE and members of the community worked hard to restore over 15 acres of native prairie ecosystem after raising over $17,000. The question of planting seeds in the fall vs the spring season was one that sparked heated debate among restoration experts and master naturalists. Given the high costs of seeds and tremendous manpower it takes for restoration initiatives, having some methodology that provides more information on when to seed would be crucial in guiding this decision.
Our project utilizes a ground up Random Forest ML model that trains data based on past restoration site information. Additionally, we make use of a LLM to make the data acquisition process smoother and more conversational rather than relying on mundane surveys. Simply talk to a chatbot which asks you questions about your upcoming restoration site's conditions and provides you with the final suggestion on whether to plant your seeds in the spring or in the fall.
We built our product using Google's Gemini LLM which had the job of asking questions to the user and parsing their answer to gather the most important information: the various data types and quantities of a restoration site. After training our Random Forest ML model with historical restoration data, we were able to accurately run our test data. This test data primarily consisted of the prairie restoration projects conducted in the 3 years that Raghav was in high school working with Eagles For Environment since we know the results of those projects. EFE had two fall seedings and one spring seeding. Like all restoration sites, the three EFE restored had their own unique challenges and features.
Integrating all aspects of the project from the front end UI to the connecting database, to the Random Forest ML model took much much longer than anticipated. What we believed would be a simple 1-2 hour job ended up taking us more than 8 hours of debugging! Realizing that each aspect caused other dependencies to fail when put together was especially infurating at times!! In the end, it would have been better if we started by integrating smaller aspects together sooner instead of attempting to integrate all aspects of the product at once.
Our UI design and animations strongly align to the theme of environmentalism and restoration. Still our working ground up model is easily our favorite aspect in this entire project.
For all of us this was the first time we put our hands on training/testing data sets for an ML model. The same can be said about the LLM API calls that we had to process and train. While these were things we heard often in theory, implementing them ourselves was a huge learning opportunity.
For RestoreAI, our next move relies on providing more insights on the restoration site like geospacial maps to highlight areas of focus for seeding and making use of computer vision technology with satellite imagery and drone imaging to identify and address invasive species early and effectively while also providing deeper insights into a restoration site. The ultimate goal of RestoreAI is to move into the agricultural industry where similar ML models can provide data driven insights to farmers enabling them to harvest greater yields.
ai, appwrite, gemini, javascript, ml, mysql, python, react