Forest fire prediction using satellite images
Forest fires occur naturally by the environment or by human intervention where both have a huge impact on the environment and natural resources etc. According to NFS, there are many causes of forest fire in the world out of which 3% are caused by natural phenomena and the other 97% is caused by human activities. The information needed for forest fire prediction is pre-forest fire data, post-forest fire data, and effective predicting techniques using remote sensing and GIS. The results that are produced here are timely and cost-effective. In the forest, fire prediction satellite sensor imagery in combination with the earth is the main source of data and these different types of sensors take the surface imagery with different pixel resolutions which also impacts the prediction of the model. The study area that we have taken for prediction is from India which is the northern part. In India, there are few areas that are prone to forest fires out of them we took a few villages in Uttarakhand state and few areas of New Delhi were the areas affected with forest fires in recent years. There are also other forests in that region which are affected by the wildfire are also taken into consideration
Our approach
Dataset(GEE)
The idea of Forest fire detection using satellite images is taken from a few research papers and by also understanding the current requirement in these fields. Here in the project, we considered the northern part of India where places like some parts of Uttarkhand and New Delhi are taken as a study area mainly these places are considered using a tool called” Bhuvan app” which plots the forest fire region that has occurred in the year 2019 out of all the areas in India these places are heavily affected.
For training, the model the dataset considered was the post-forest fire dataset in these regions. The dataset is taken from the tool called Google earth engine where this tool consists of many kinds of datasets out of which I have chosen Modis dataset for both burned layer and satellite layer the reason for taking this satellite imagery is it consists of pixels with a higher resolution which is useful in training the model. So to access the modes dataset the code has to be written for accessing the dataset and changing it to the way we want the satellite imagery after getting access to the dataset it is imported to google drive the dataset that we got now is taken to QGIS software for applying operations like slicing and resizing the images and converting the formate that we want.
The images that we got are further preprocessed and the dataset we obtained here is given to training the model. The model here we used is U net Neural Network this is basically used for the image classification the model here we used is able to distinguish and localize the borders for each pixel. The data we got after pre-processing is being trained with the model and the output predicted as the location of the forest fire. The average accuracy of the model we obtained is 82%.
Experiences and Future works
It was a great experience for us working on satellite imagery and prediction of forest fires as this is is the first time working on satellite imagery I have learned many new technologies and explored and learned solving new problems, during the development of the project we have faced a lot of challenges like identifying different layers of modis and combining them to one single data importing this large chunk and processing took a lot of efforts in the development of the project. There are a few changes and improvements are to be done in adding the features etc.
Acknowledgments
I would like to thank Dr. Kuldeep who guided me and helped me in each step of the development of the project
Thank you!
Abhishek Veeravelli
Btech 4th year CSE student
Bennett Univesity
