Fire Detection Using Surface Vector Machine (SVM)

We know earth is covered with carbon rich vegetation, seasonal dry climate, widespread volcanoes and lightening. In this way we can say that there is something that is always burning on this earth, one of the possibilities that we are going to discuss here – wildfire/forest fire. Wildfire may be natural or manmade (accidental or deliberately). A large amount of greenhouses and smoke are released by these fires and cause severe degradation to our ecosystem. Based on MODIS, (Moderate Resolution Imaging Spectro-radiometer) of NASA’s Terrestrial and aqua satellites, several algorithm has designed and many approaches have been proposed to detect wildfire. The question on accuracy over existing method of fire detection method remains same. The director from school of management, karpagam university and students from IIM Bangalore have proposed a more robust and accurate approach using Surface Vector Machine for forest fire detection.

 

What causes wildfire?

As per NASA, the primary cause for wildfire are spontaneous combustion, lightening and heat waves caused by climate change and volcanic eruptions and the most common human factor includes arson, animal husbandry, land conversion burning, fireworks, campfire, power lines etc. Some fire burns for a days and weeks causing severe damage to the ecosystem and people.

Why it is important to heed?

With the rising temperatures wildfire is becoming a major concern wildfire on May 01, 2016 in McMurray, Alberta, Canada continue to spread for 2 days and on May 03, 2016 it swept about 590000 hectares of the forest before it declared to be under control on July 05, 2016 (about 1 month) and it was completely extinguished on the spring of 2017.

Similarly on the same day, wildfire destroyed about 3500 hectares on the land of Uttarakhand, India. According to US department of agriculture wildfire season in US has been increased by 78 days against the number in 1970. In 2016 approximate 2.6 million acres of land was destroyed in US. A paper published in 2016 by Alicia M, Kinoshita explains about wildfire and its adverse effects on environment.

What Steps To Be Taken?

Several algorithms have been developed for detecting wildfires. A simulation method was designed by  Boyd M Harnden (1973) which could be easily adapted for any forest fire detection while in 1993 a mathematical model based on infrared passive sensor – Lidar, was developed by F. Andreucc to calculate properties of smoke plume produced by wildfire and used it to detect fire.

In last decade a lot of study happening on fire detection, the study includes use of sensors and image (image processing). Frequency ratio model using remote sensing and geographical information system was developed in 2007. The NAUTEA in 2010 were some attention seeker methods to predict forest fire. S. Moroghati used MODIS data to detect forest fire using agent based algorithm. But the paper published was restricted to Iran. Recently in 2015 Yang Jia used adaptive flame segmentation and recognition algorithm to detect fire detection in spacious building while Emma Prema in 2016 developed an image processing approach for detecting smoke.

This article will focus on SVM (Surface Vector Machine) a popular machine learning algorithm proposed by Vapnik in 1998. SVM is a supervised learning model used for solving classification problem. The data used by the researchers for analysis was obtained from Earth Observing System Data & Information Center (EODIS), a wing of NASA’s Earth Observing System (EOS).

How SVM Will Work?

The fire maps show the locations of actively burning fires around the world on a monthly basis, based on observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA‟s Terra satellite. For the research, researchers used data from MODIS for the months of April 2014 and April 2015. Data collected in April 2014 is used for the training model while April 2015 is used for validating the model. In this paper we will be comparing the Logistic regression and Support Vector machines and also will try to show why SVM is better.

SVM functions by projecting feature space into kernel space and making classes linearly separable. Or to explain still simpler, SVM adds additional dimension to feature space in a way that makes classes linear separable. And this planar boundary when projected to original feature space emulate decision boundary which is non linear. The image below explains might give you better explanation. From the image-1 you can see that with third dimension is added to data, we can separate two classes with a linear separator, and when projected back to original 2D feature space, it becomes circular boundary. The image-2 shows, how SVM performs for our sample data. Now the difference might be clear. In SVM the decision boundary can be circular. But still the basic researcher question remains unanswered. That is when and which algorithm should be used when dealing with multi-dimensional data. And it will be explained in following section. 

Image-1

 

Image-2

 

The most important thing about support vector machines is that they rely on boundary cases for building separating curve. They can also handle decision boundaries which is nonlinear. Reliance on boundary cases helps them to handle missing data. Large feature space can be handled with support vector machines and which makes SVM most used algorithms in text analysis, whereas logistic regression is not a choice because it fails to handle large features. Result of SVM is not easier as logistic regression for layman. SVM is very costly to train huge data because of nonlinear kernel. To summarize pros and cons of SVM:

SVM Pros:

1. Large feature space can be handled

2. Non-linear feature interactions can be handled                         

3. Does not rely on entire data

 

SVM Cons:  

With large number of observations it’s not efficient  and is not easy to find appropriate kernel

 

Conclusion

There are many approaches for fire detection. While some models are complex and others are simple, and some are usually restricted to a particular demography. In this paper, we have developed a simple approach of using SVM and Logistic Regression for wildfire detection across the world. The Logistic Regression methodachieves an average detection rate of 88.0% whereas SVM method achieves an average detection rate of 94%. This proves that the SVM classifier has the best stability and the accuracy compared to most of the other approaches proposed till date. The result also shows that both the accuracy and robustness of the classification has improved using SVM and therefore SVM is appropriate for wildfire detection.

For Technical Details Click Here 

 

Follow us @ Facebook

Follow us @ Twitter