UAV Forest Fire Detection Model

This repository contains the necessary tools to setup, train, and test a neural network for forest fire recognition.

The project_results directory contains a report and poster discussing the design and performance of the final product.
In summary:

• 2596 aerial forest images were collected.
• 6 models were compared with an emphasis on accuracy and inference time.
• The best model uses transfer learning with ResNet50 to achieve an accuracy above 90% with an inference time of under 1 second.

Demo

Here is a demo of the model using ResNet50.

Setup

Two models are currently available for testing: ResNet50 and EfficientNetB7. They are run with their respective files using Python3. At the top of the file, you may specify parameters such as hidden layers, and whether or not you want a plot of the training performance over time.

Training the Model

Ensure you are using the correct Python3 environment. The server uses conda to manage environments, so you will want to run

conda activate tensorflow

If you're using the EfficientNetB7 model, it is not contained in the conda environment. EfficientNet is only support on TensorFlow nightly builds, so this has been installed with Pip. If running on the server, it should already by accessible to you.