Wildfire-Detection-using-Deep-Neural-Networks-IntelONEAPI

Darpan Deb

Bengaluru, Karnataka

0 0

Forest fires wreaking havoc and destroying several irreplaceable ecosystems is common news in the past few years owing to climate change. The best solution to the issue is having a reliable detection system, which can pick up on the early signs of a forest fire and inform the authorities immediately ...learn more

Project status: Concept

oneAPI, Artificial Intelligence

Intel Technologies
oneAPI

Code Samples [1]Links [2]

Overview / Usage

The idea we are proposing will definitely help in preventing wildfires from reaching exorbitant levels. It can also in some cases help in preventing the forest-fire from breaking out at all and upholds the environmental concerns by predicting the air quality in affected areas. In the worst case scenarios we can have an early warning for a forest fire and the authorities can contain and put the fire out before it destroys a large area.

Methodology / Approach

We conducted frame-wise classification to identify instances of forest fires within frames. Additionally, we applied feature extraction techniques to our numeric data and implemented data pre-processing methods.
Conducted preliminary data analysis on numerical data and applied data augmentation techniques to images in order to train our model effectively
Applied various model building alogorithms on numeric data like Logistic Regression, Random Forest, KNN, XGBoost Classifiers and along with these we implemented Hyperparameter testing on numeric data. We also trained and validated the Deep Neural Network model on imagery dataset for fire or no_fire image classification.
For testing, the accuracy of respective models are given as

Accuracy of model on numeric dataset

KNN --> 0.9315
Logistic Regression --> 0.91
RF --> 0.97
XGBoost --> 0.9726

Accuracy of classification model on Image dataset

DNN ---> 0.54

Conclusions

Among all the models, we got XGBoost Classifier with highest accuracy, so we deployed the XGBoost Classifier model to create our Web Application and for image classfication the DNN produced an accuracy of 54%.
The entire procedure is completed with the help of Intel oneDAL and AIToolkit and Intel oneDNN (TensorFlow) to get better results and faster computation(Intel oneAPI Data Analytics Library (oneDAL) and oneAPI Deep Neural Network Library (oneDNN))