Inspiration
I got the inspiration for this project from the ever-increasing global average temperature, global warming, climate change, and most importantly, the Wildfire Catastrophe of 2019-2020 in Australia. I thought, why not use AI to help with that?
What it does
My project is designed to be accessible to anyone, which is why all three models—satellite, camera, and meteorological wildfire detection models—are accessible through a Flask app. These models use satellite imagery, images from cameras and drones, and weather data from a particular location to detect wildfires or at least indicate the probability of a wildfire occurring. It can also send hourly alerts if there's any sign of a wildfire in the pre-selected area.
How we built it
I trained the satellite imagery wildfire detection and camera image wildfire detection models using ResNet50v2 as the backbone for feature extraction and dense layers for predictions. For the weather data-based wildfire prediction model, I built a simple feed-forward neural network. After training all three models, I developed a Flask app from which each of these models is easily accessible. For the interactive map on the website, I've used Mapbox's free API. For sending alerts, I've used the free API of Mailersend. All of this was built using Python, with the frontend using HTML, JavaScript, and Tailwind CSS.
Challenges we ran into
The most difficult part was gathering the weather data as it wasn't diverse and large enough, affecting the model's performance in different regions. Another challenge was the large but not diverse satellite imagery data, which again impacted the model's generalization for different regions. To tackle this issue, I've made the app provide separate as well as combined predictions for both models. This will reduce the errors if any model makes a mistake.
Accomplishments that we're proud of
Under the test dataset, we achieved the following accuracies:
- Satellite Classification CNN: Using ResNet50v2, this model detects wildfire probabilities from satellite images with an accuracy of 97%.
- Image Classification CNN: Also based on ResNet50v2, this model identifies fires in uploaded images with an accuracy of 98%.
- Weather Data Model: This model predicts wildfire risks from meteorological data with an accuracy of 100%, albeit on a limited dataset.
Additionally, it's an innovative and never-before-made project that uses a multi-modal approach and deploys it on an easy-to-access Flask app.
Another potential accomplishment will be winning this hackathon.
What we learned
I learned a lot about the causes of wildfires, mitigation practices, ways of detection, using deep learning in this field, and how AI could be used to reduce environmental damage. I also learned how a multi-modal approach can be used to tackle inaccuracies in deep learning.
What's next for Wildfire Detection System - AI-Powered Solution
Next, I'm planning to implement wildfire burnt area segmentation to highlight the areas that have burned on the map. This will help reduce further damage caused by wildfires. Additionally, as mentioned, the limited quantity and diversity in data have affected the model's generalization, so I look forward to collecting data that is more diverse and in greater quantity to improve the model's generalization.
Log in or sign up for Devpost to join the conversation.