Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Wildforest fire detection using deep learning and fusion techniques on aerial image datasets

Zachariadis Emmanouil

Full record

---

Summary

Greece, like other countries, suffers every year from intense fires which inevitably cause losses. In the best-case scenario, these losses are materialistic only, but there are many cases where the consequences are not limited to those losses. This kind of destruction has a crucial impact on firefighters and civilian lives, animal extinction, and forest degradation. Considering scientists' predictions that fires will increase every year due to climate change, many prevention systems have been set up to avoid this kind of situation. This thesis focuses on developing a detection approach that can classify images of forests in “Fire” and “non-Fire” cases with input from both RGB and Infrared (IR) cameras. Along with this work, an architecture is proposed that increases the possibility to prevent these previously mentioned consequences. The new deep learning architecture is called ShRe-Xception (Short Recursive), inspired by already existing Xception network architectures (i.e., small-arch Xception and original Xception Network). In this study, an experimental process takes place around how to train a neural network with two different datasets and the relevance of its architecture. First, transfer learning is performed on the small-arch Xception network with input from either RGB or IR images, and then, the same structure is trained with all the RGB images from the very beginning. Also, the proposed ShRe-Xception network is trained by using all RGB images from the very beginning and then retrained with infrared frames. Datasets used for training and testing purposes contain “Fire” and “non-Fire” images of forests that are captured from UAVs and uploaded to the IEEE portal. When testing the above models, the evaluation accuracy was RGB: 90.10% and IR: 99.31% after initial training and ShRe-Xception model, RGB: 77.13% and IR: 94.47% after transfer learning on small-arch Xception, and RGB: 84.86% and IR: 29.19% when initially training the small-arch Xception. The above results are described and analyzed within the present thesis, as many other operations were performed in terms of experimentation and results. In general, the latter model came to be the most accurate one for images of both spectrums in our experiments. The ShRe-Xception can potentially play a vital and efficient role in real-time fire detection during aerial surveillance of wild forests.

Available Files

Services