SynthEcoWatch

Inspiration

Wildfires are a serious environmental issue in Canada. In Canada, wildfires burn an average 2.5 million ha/year, nearly half the size of Nova Scotia. To facilitate controlling wildfires, we proposed this project to detect the smoke generated by forest fires to enable a more rapid, and targeted approach to tackling them.

What it does

Our project is to segment the smoke of forest fires. We used an AI synthetic dataset, which is beneficial due to the limited availability of forest fire datasets. We used FCN to segment the smoke.

How we built it

We generated the background images (forest) using OpenAI DALL-E-2. The benefit of using AI-generated images is that it is easier to set different environments, locations, wind, and sun. For smoke and fire, we used Blender to simulate instances of a fire, and its accompanying smoke and atmospheric conditions for our training set. This included its behavior in the presence of wind and in highly combustible scenarios, which might lead to dense, cloud-like smoke. The smoke was generated by first centering a mesh UV Sphere and simulating an inflow of fire and smoke in Blender. To simulate realistic motion, a wind factor of 1.0, a fuel setting of 1.0, an initial temperature of 1.0, and a surface emission of 1.5 are fixed. Wind strength and smoke density are adjusted to account for variations in the intensity and weather conditions present during wildfires.

Challenges we ran into

A challenge that arose in creating the training dataset was that upon rendering our fire and smoke simulation, we encountered a significant deficiency in the lighting of the frames. To fix this, we attempted to increase the lighting by 20x, but it turns out our camera angle was not projecting from an optimal angle and needed to be adjusted. Additionally, our model started out performing poorly with negative samples, which we fixed by training it further with additional negative instances.

Due to the limited time of the hackathon, our dataset quality was limited by the short amount of time, the seasons, and the weather for the forests. This can be improved easily by rendering and generating more images by AI if given more time.

Accomplishments that we're proud of

We are proud that our model can recognize the general smoke pattern in a majority of instances and even differentiate between clouds and smoke. We didn’t use binary colors for masks as it is hard to determine the edges of the smoke. To solve this, we used a range of values between 0 to 1, determining the thickness of the smoke. Moreover, the smoke will randomly change yellow and change brightness. Our project is to segment the smoke of forest fires. We used an AI synthetic dataset, which is beneficial due to the limited availability of forest fire datasets. We used FCN to segment the smoke.

What we learned

We learned how use AI to generate dataset. We learned how to use FCN as our neural network. We learned how to use LaTex to write papers.

What's next for SynthEcoWatch

Due to the limited time of the hackathon, our dataset quality was limited by the limited time, seasons, and weather for the forests. This can be improved easily by rendering and generating more images by AI if given more time.

Read our technical report at https://www.overleaf.com/read/qypxvptvrbym#0ac7ca