tianma2000 / forestfiredetection

Comprehensive Analysis of Forest Fire Detection using Deep Learning Models and Conventional Machine Learning Algorithms

ForestFireDetection

Comprehensive Analysis of Forest Fire Detection using Deep Learning Models and Conventional Machine Learning Algorithms

Zeynep Hilal Kilimci and Süha Berk Kukuk

Paper : International Journal of Computational and Experimental Science and Engineering

Abstract : Forest fire detection is a very challenging problem in the field of object detection. Fire detection-based image analysis have advantages such as usage on wide open areas, the possibility for operator to visually confirm presence, intensity and the size of the hazards, lower cost for installation and further exploitation. To overcome the problem of fire detection in outdoors, deep learning and conventional machine learning based computer vision techniques are employed to determine the fire detection when indoor fire detection systems are not capable. In this work, we propose a comprehensive analysis of forest fire detection using conventional machine learning algorithms, object detection techniques, deep and hybrid deep learning models. The contribution of this work to the literature is to analyze different classification and object detection techniques in more details that is not addressed before in order to detect forest fire. Experiment results demonstrate that convolutional neural networks outperform other methods with 99.32% of accuracy result.

Dataset

You can Download Data On Google Drive. Link is Google Drive Link

Folders

1. Models Folder

• There are contain algorithms that we are used for fire detection

2. Prepare Data

• Own data's are collected using Beatifulsoup and Web Scariping. In this Folder contains codes are that we used

Result

The following abbreviations are used in the tables: AC: Accuracy, FM: F-measure, PR: Precision, RC: Recall, SVM: Support vector machine, RF: Random Forest, CNN: Convolutional neural network, CNN-GRU: Convolution Neural Network-Gated Recurrent Unit, CNN-LSTM: Convolutional neural network-long short-term memory, SSD: Single shot detector, Faster R-CNN: Faster recurrent-convolutional neural network, Avg: Average. The best results are obtained for each dataset in the Table 1 and Table 2 after experiments of hyperparameter tuning. The best performance results are also demonstrated in boldface in all tables. In Table 1, the performance results of all classification models according to evaluation metrics in the first dataset (DS1) are demonstrated.

In Table 2, the performance results of all classification models according to evaluation metrics in the second dataset (DS2) are demonstrated.

Packages Reqs:

• Keras
• TensorFlow
• Numpy
• Matplotlib
• Scipy
• ApiClient
• Imutils
• OpenCv
• Argparse
• MoviePy
• SimpleJson
• PyFcm
• Oauth2client
• Httplib2