The internship presentation summarized work done at EDQI on developing a parallel support vector machine model for forest fire prediction. The interns analyzed forest fire and weather data from India and Portugal, finding their model achieved a 63.45 RMSE on Portugal data compared to 63.5 for a standard SVM. They also discussed using their interface to predict forest fires in the Amazon rainforest based on historical data, obtaining 99% accuracy in predicting spot fire ranges. The goal of their extension project is predicting forest fires to enable better prevention through machine learning.
1. Review 1 Internship Presentation
1DT19IS131: Spoorthi R Prakash
1DT19IS142: Tanuja N G
1DT19IS171: Harshitha S
Internship Coordinator : Dr.Rashmi Amardeep
Associate Professor
Forest Fire Prediction
EDQI
EXTERNAL GUIDE NAME : Mr. Shree Harsha
Chief Executive Officer, EDQI
DAYANANDA SAGAR ACADEMY OF TECHNOLOGY & MANAGEMENT
Department of Information Science & Engineering
2. About EDQI
Overview
At Edqi they envision the role of technology in transforming education, we serve educational institutions with
hands-on and technology-based learning solutions across classrooms, libraries, innovation centers, science and
engineering labs in educational campuses and organizations across the globe.
Website
http://www.edqi.in
Industry
Education Administration Programs
Company size
11-50 employees
7 on LinkedIn Includes members with current employer listed as EDQI, including part-time roles.
Headquarters
Bengaluru North, Karnataka
3. INDRODUCTION
• Forest fire is considered as one of the main cause of the environmental hazard that provides many
negative effects. Effective Forest Fire prediction models help to take the necessary steps to
prevent forest fire and its negative effects. Existing methods of Cascade Correlation Network
(CCN), Radial Basis Function (RBF) and Support Vector Machine (SVM) were applied for the forest
fire prediction. Existing methods have the limitations of over fitting problems and lower efficiency
in prediction. Existing methods in forest fire prediction have lower efficiency in large dataset due
to overfitting problem in the models. The parallel SVM method is developed in this research for
reliable performance of the Forest Fire Prediction. Conventional SVM has a higher efficiency in
predicting the small fire and has lower efficiency in predicting large fire. The SPARK and PySpark
were applied to perform the data segmentation and feature selection in the prediction process. A
parallel SVM model is developed to train the meteorological data and predict the forest fire
effectively. The parallel SVM model reduces the computational time and high storage required for
the analysis. Parallel SVM considers the Forecast Weather Index (FWI) and some weather
parameters for the prediction of a forest fire. The parallel SVM model is evaluated on the Indian
and Portugal data to analyze the efficiency of the model. The parallel SVM model has the 63.45
RMSE and SVM method has 63.5 RMSE in the Portugal data.
4. Amazon Forest Context
• # 6.7 million square kilometers among 9
countries
or 28 times the size of the UK
• # 80K species of flora
and an estimated 400 billion trees standing
• # 6% of the world’s oxygen production
act as a carbon sink, absorbing large amounts of
carbon dioxide from the atmosphere
(Leman J., 2021),
• # home for 30 million people
including almost 3 million indigenous people
5. Problem
• The rate at which the Amazon is deforested is a global concern, as it plays a key
role in the environment and the atmosphere, not only in South America but also at a
global scale.
• Different studies indicate that deforestation has meant the loss of 20% of the
forest, and if it continued at this rate by 2050, 40% of it would have disappeared (H
Ter Steege, 2019). Between 36% and 57% of the Amazonian tree species could be in
danger.
6. Analysis
# Did the number of fires increased?
# Do carbon emissions correlate with the size of the burned areas?
# What is the capacity of the affected areas to retain Carbon?
# What are the most affected areas and how are they distributed
in the territory?
# How the Amazon fires behave throughout the months of the years?
7. Dataset information
# Characteristics of the fires in the Amazon forest between 1998 and 2013,
# Shape 1513059 rows × 15 columns,
# Some of the main features:
- C: Carbon emissions in g C/m2,
- NPP: Net primary production, carbon retained in g C/m2,
- Fire spots: number of spot fires ignited outside the perimeter of the main fire.
- Burned area: area affected by each fire in ha.
- Latitude/Longitude: geographical coordinates of the fires
8. OUTCOMES
To create a user interface for predicting forest fires using machine learning models.
Burned Areas: mostly concentrated in the surroundings of the
forest
- Southern and
eastern regions of
the amazon have the
greater burned areas
- Most affected areas
are in the states of
Tocantins and Mato
Grosso (Brazil)
- 2.115.433ha of area
burned according to
this dataset (0.3% of
the whole territory)
9. Evidence of “fire season”, where both spot fires and burned area
increase
- Season of fires in the
amazon occurs mainly
between july and
October
- Amount of spot fires
increase significantly
during fire season
- Both burned area and
spot fires follow a
similar pattern
10. Results: high accuracy prediction for the classes of spot fires
Confusion Matrix
on the test set
99.6%
test set accuracy
100%
train set accuracy
0.3032 - 0.3064
conf interval for fire_spots
(98% conf level)
11. Conclusion
- From 1998 to 2000 the number of the fires in the Amazon rapidly decreased, but stabilized
for the following 13 years, showing no clear trend for the future,
- There is an evident season of fires in the Amazon and the spot fires increase
significantly in that period, a similar pattern followed by the burned areas,
- Predicting the spot fires could be relevant to control the spread of the fire,
specially during the fire season,
- The multi class model high-performs with a 99% of accuracy on the test set to
predict the 3 different ranges of spot fires in the Amazon forest,
12. Timeframe of Extension Project
• Project most probably will be completed by 30th April.
• The goal of this project is to make prevention easier by predicting the
occurrence of forest fires using machine learning