The inspiration behind FloodSafe was the need to address the significant impact of floods caused by atmospheric rivers in California, including financial, environmental, and human losses.
What it does
FloodSafe utilizes machine learning, satellite imagery, and weather data to predict atmospheric river floods in California with 98% accuracy. It is accessible to the general public through a website and aims to provide a long-term solution for flood prediction.
How we built it
FloodSafe was built around a ResNet convolutional neural network and trained on over 2000 satellite water vapor images that we manually scraped from NASA’s Worldview, along with weather data from the Open Meteo World Weather API. The model was trained over 30 epochs and tested on unseen data to achieve 98% accuracy. The system is wired to a website for accessibility.
Challenges we ran into
One of the biggest challenges was collecting and processing the vast amount of data needed for training the machine learning model. We also had a lot of trouble working with the EC2 instance to set up the FastAPI server that the model is hosted on.
Accomplishments that we're proud of
We are proud to have developed an accurate and accessible flood prediction system that can help mitigate the impact of atmospheric river floods in California. Achieving 98% accuracy in predicting AR-based floods is also a significant accomplishment for us.
What we learned
We learned about the complex nature of atmospheric river floods and the importance of utilizing machine learning and satellite imagery to accurately predict them. We also learned about the challenges of working with large datasets and ensuring data balance in machine learning models.
Additionally, we were inspired by the workshops that Lynbrook ML hosted, namely the Web Dev workshop on deployment and the CAA workshop about environmental ML. We incorporated these ideas into our project as well.
What's next for FloodSafe - Detecting Atmospheric River Flooding
In the future, we plan to expand FloodSafe to other areas beyond California and improve its accuracy by incorporating additional data sources. We also aim to collaborate with emergency response agencies to make FloodSafe a more integral part of flood response efforts.
NOTE: We built this project previously as part of the Synopsys Science Fair. For this hackathon, we decided to modify the model architecture after attending the Intro to ML workshop hosted by ML club, in which they talked about overfitting. To eliminate the possibility of overfitting, we decided to remove one of our intermediate layers of neurons because we felt it would fit the numeric data better and be more accurate overall.