Predicting wildfires from satellite images using deep learning
Loading...
Supplementary material
Other Title
Authors
Mammen, Blesson
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
Date
2024
Supervisors
Ardekani, Iman
Shakiba, Masoud
Shakiba, Masoud
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
wildfires
deep learning
bushfires
satellite image maps
deep learning
bushfires
satellite image maps
ANZSRC Field of Research Code (2020)
Citation
Mammen, B. (2024). Predicting wildfires from satellite images using deep learning. (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology https://hdl.handle.net/10652/6456
Abstract
Detecting the possibility of wildfires early can help individuals and organizations respond appropriately and minimise the potential damage caused. This report investigates the use of MobileNetV3 for predicting the occurrence of wildfires from satellite images that do not contain visible wildfire spots. The paper delves into current research and highlights the value of satellite imagery as a homogeneous data sources for wildfire prediction. A thorough evaluation and comparison of MobileNetV3’s performance with larger, more complicated models like ResNet50 and VGG19 is conducted. The findings demonstrate MobileNetV3’s efficacy in balancing computational efficiency with predictive power, offering a lightweight yet effective alternative to traditional models. By examining the potential of lightweight neural networks in managing complex and difficult environmental data, this paper advances wildfire prediction methodologies especially in resource constrained contexts.
This research focuses on a key challenge: developing a model designed to predict small wildfires from satellite images that do not contain any visible wildfire spots. This is achieved by training the model on a dataset comprising satellite images of areas where wildfires with a small wildfire spot size just greater an 0.01 acres have occurred. Our approach addresses a critical gap in wildfire management by focusing on predicting these small-scale fires without needing heterogeneous data collection. The ability to predict small wildfires from non-fire satellite images enhances the accuracy and utility of early warning systems. Using satellite images as a single source of data ingestion removes the need for heterogeneous data collection involving soil and atmospheric data as well as vegetative and geological data. It allows for the implementation of targeted preventive measures, such as controlled burns, the creation of firebreaks, and stricter fire bans during high-risk periods. Motivated by the capacity to generalize and the possibility to overcome challenges posed by cloud cover, haze, and diverse landscapes, this research uses MobileNet V3, a deep learning model based on transfer learning to predict wildfire occurrence from satellite images. MobileNetv3 model has given promising results, with a recall of 92 percent and an accuracy of over 82 percent. Despite being a lighter model, MobileNetv3 has shown robust results when evaluated alongside heavier models like Resnet50 and VGG19.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved