Inspiration
All 3 of us are allergic to pollen, and it's hard during springtime for us, because we never know which spots to be near and which to avoid. None of us like to stay inside, so we decided to make an app to help us know where we should be and where we should avoid. One of us also has asthma, so we thought it would be good to see how certain plants affect health risk factors.
What it does
It's simple. The first step is to add risk factors to the app-- you add different health conditions you have that could affect your health, and the app saves that. You either take a picture or upload one that you already took, and the app will tell you what species it is, what kind of microscopic pollen it has, and how that affects your risk factors.
How we built it
To detect the species of the plant, we trained an EfficientNetV2 Convolutional Neural Network using PyTorch on the Pl@ntNet-300K image dataset, which contains over 300,000 images of over 1,000 plant species. After that, we used deep learning to map the plant names to their microscopic pollen. Then, we hosted that model on AWS SageMaker using an S3 database. The frontend then takes an image and runs it through the backend server. Then, the app uses the gemini-2.5-flash API to generate text regarding the risk factors given the provided risk factors in the app and the plant(s) identified from the picture.
Challenges we ran into
Even though we used EfficientNet, which is a fast-training model, training still took roughly a day, since each epoch had to go through over 300,000 images. We also had to run this locally on our own computers, which added to the high training time. AWS is also a relatively new tool, and implementing it to host our backend was also difficult, as we had not worked with it much in the past. Also, connecting everything together definitely took longer than we had imagined.
Accomplishments that we're proud of
We are very proud that we managed to pull off a full stack app, with a quality frontend that looks clean that integrates with an efficient backend hosted on AWS and not locally. We are also proud that we managed to train a Convolutional Neural Network to gain almost 90% accuracy identifying plants-- trained on local hardware on a laptop.
What we learned
We learned how to use AWS to host a backend on a server and connect it with a frontend. We learned about the EfficientNetV2 architecture and how it is more efficient and more accurate than most CNNs and Vision Transformers.
What's next for PlantHealth AI
We want to make the UI cleaner, and add a YOLOv11 model from ultralytics to add real-time scanning into the app with bounding boxes so the user can identify plants more easily and live scan their surroundings for risk factors.
Log in or sign up for Devpost to join the conversation.