2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

Climate Change Forecast for Forest Fire Risk

2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS) | 979-8-3503-9737-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICACCS57279.2023.10112983

Prediction using Deep Learning

R. Shanmuga Priya, 2K. Vani.
1

Department of Information Science and Technology,

College of Engineering, Guindy.
Anna University.
Chennai 600025, India
1
[email protected] , [email protected]

Abstract—Climate change is a hazard to the environment, Forest fires are a result of such alarming temperature rises
generating serious problems such as global warming, which can observed which prove to be catastrophic to wildlife and
be catastrophic. To predict the severity of this situation, it is humans. They lead to deforestation and loss of lives which
necessary to examine climate change patterns over the course of hider the eco cycle thereby, wreaking havoc to mankind.
several decades. This forecast can help us predict the magnitude According to some experts, if we keep using fossil fuels at the
of the impact and, as a result, give us an insight on taking the same rate, sea levels could rise by several meters during the
required precautions to address this pressing issue. It has been next 50 to 150 years, devastating coastal communities all over
observed through multiple studies that climate change has led the world. With the use of machine learning methods such as
to a dramatic increase in forest fires and risks associated with it.
regression models, our project tries to forecast the severity of
Temperature peaks can cause forest fires because of the
scorching heat. This work intends to aid in the prediction of
climate change patterns and thereby, give an insight on
forest fire risk based on climate change patterns based on probable forest fire risk that can be observed in the future.
temperature records from the last few decades, and to give us Deep Learning Techniques, notably neural networks, help
an indication of how the climate will be in the future, thereby, to improve the model's accuracy far more. This climate
giving an insight into probable forest fire risks. For the prediction or forecast will assist us in taking the required
prediction, we use traditional Machine Learning methods, with preventative and precautionary actions to mitigate the impacts
Deep learning techniques included to enhance forecast
of global warming and forest fires and bring it under control.
accuracy.
II. RELATED WORKS
Keywords—Machine learning, climate change, forest fire,
greenhouse gas, temperature forecast. Development of a classification method using statistical
data on forest fires and meteorological data collected from
I. INTRODUCTION 1996 to 2008 according to the Canadian Forest Fire Weather
The progressive heating of the Earth's surface, oceans, and Index system (CFFWIS) is discussed in [1]. A new fire danger
atmosphere known as global warming is a result of human classification method was developed by means of the
activity, primarily the burning of fossil fuels, which releases statistical data on forest fires and meteorological data
greenhouse gases like carbon dioxide (CO2) and methane into collected from 1996 to 2008 according to the Canadian Forest
the atmosphere. The Earth's atmosphere is covered with Fire Weather Index system (CFFWIS) in [2].
greenhouse gas emissions, trapping the sun's heat. Global An overview of advancements in deep learning and
warming and climatic change occur as a consequence. The machine learning is provided using a new category of
rate at which the Earth is currently warming is unprecedented methodologies. The report claims that research is currently
of in recorded history. The increase in global temperatures is being done on deep learning, which is still in its early phases
one of the most evident and direct effects of global warming. of development. In order to find correlations between short-
Climate change is the long-term alteration of the regular term and long-term temperature responses to various climate
weather patterns that characterize local, regional, and global forcing scenarios, L. A. Mansfield and colleagues [3] devised
climates on Earth. These changes have a variety of effects. As a machine learning method that makes use of a special dataset
the average global temperature increases, weather patterns are of recent climate model simulations. This approach can not
changing. Global warming directly contributes to extreme only reduce the cost of scenario calculations, which could
weather. Winters that are colder than usual may be one effect speed up climate change forecasts, but it also helps in the
of climate change in some regions. Numerous people have lost search for indications of simulated long-term climatic
their lives because of severe heat waves in recent years. An reactions, which is significant for the detection, predictability,
unsettling sign of things to come is the roughly four trillion and attribution of climate change. Their findings draw
metric tons of ice that Antarctica has lost since the 1990s. attention to the difficulties and opportunities associated with
data-driven climate modeling, particularly when more
979-8-3503-9737-6/23/$31.00 ©2023 IEEE
1065

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

extensive model datasets are added in the future. Although grids, could bridge existing gaps. A Systematic Review based
many different Earth System variables have been examined on the effects of Climate Change and Earth Sciences using
using ML approaches, more general applications to various artificial neural network concepts, a survey to
comprehend the complete climate system have not yet been forecast climate changes [14].
created.
III. METHODOLOGY
For instance, ML could be useful in teleconnection
identification, where difficult feedbacks make it difficult to A. ARIMA
perform direct equation analysis, visualize observations, or A class of models known as ARIMA, or "Auto Regressive
diagnose Earth System model (ESM) issues. Artificial Integrated Moving Average," employs a time series' own prior
intelligence (AI) can then exploit the discovered climate values to explain the time series and predict future values.
connections to provide better forecasts of impending weather ARIMA (p, d, q), where g stands for the order of the moving-
phenomena, including extreme events. As suggested and average model, d for the degree of differencing, and p for the
presented by Chris Huntingford and colleagues in their paper order of the autoregressive model, is the common shorthand
[4], while ESM development is crucial, a parallel focus on for ARIMA models. To forecast future values and extrapolate
employing machine learning and artificial intelligence is present values into the future ARIMA models use auto-
advised to better comprehend and profit from existing data correlations and moving averages, which also transforms non-
and simulations. Sushmitha Kothapalli and her colleagues [5] stationary time series into a stationary one.
introduced an ARIMA model for predicting future values that
delivers superior outcomes for time-series data. The B. Logistic Regression
parameter correlation analysis aids in forecasting future Climate dataset that has been collected and pre-processed
values. Data mining techniques and AI algorithms are utilized was split into train and test set and trained with a Logistic
in this paper Regression Classifier Model. The C_ space parameter, known
as regularization strength inversion is fine-tuned, which
[6] to anticipate future precipitation based on historical controls the amount of regularization being performed on the
data. To foresee future forecasting and possible climate model. A GridSearch CV, a Brute Force approach that
change, data mining and statistical tools are applied. This chooses the best model from all the possible models has been
paper [7] focuses on weather forecasting prediction and included to find the best fitting model.
analysis, and it uses machine learning in its entirety. We now
have many better predictive systems, but no adequate LogisticFunction(x) = 1/1 + e−x (1)
assessment for future damage exists because the time span
between the present moment and the time for which A base linear regression function is incorporated into
forecasting is generated and issued is variable. deter- mining the probability of the prediction of Logistic
Regression.
In their paper [8], P. Chandrashaker Reddy and colleagues
determined that by gathering past datasets from datacenters
such as India Meteorological Departments in a certain place
and using regression analysis, we may forecast numerical (2)
damage values. The research on weather forecasting using a Where P(y) is the probability of prediction where
historical database is presented in this paper [9]. The usual x1,x2,..xp are feature and β0, β1,… are constant parameters.
method is ineffective because the atmospheric pattern is a Weak learners are estimators that yield wrong outcomes in
dynamic, non- linear system. A powerful solution for handling rare cases. The idea behind this approach to make a series of
this problem is the multi-layer neural network. The weak learners to produce a strong rule which eventually
atmospheric pattern is separated into five patterns using back boosts the accuracy. The observations are assigned weights
propagation neural networks based on rainfall. according to the level of difficulty by the algorithm.
This paper [10] offers a classifier strategy for predicting Sequentially, the weak learners are added to train the high
weather conditions and demonstrates the use of Naive Bayes weighted observations. Once they are trained successfully,
and Chi square algorithms for classification. The prediction of their weights are reduced and loss is minimized with equating
global temperature and carbon emissions of India from the actual and predicted value.
datacollected previous year was presented by M. C. Random Forest Classification
Purushotham Reddy et al. in their paper [11]. The predictions
This model creates a set of estimators for the selected
were done using Linear Regression Model. An Ensemble
subset of random observations from training dataset
Approach for Global Land Temperature (EAGTL) was
containing symptoms and result. The final prediction is made
proposed by Himika; Shubhdeep Kaur; Sukhchandan
from the decisions contributed by individual estimators all
Randhawa [12]. Machine learning has the ability to help
together as a whole.
civilization adapt to a changing environment and cut
emissions of greenhouse gases, according to a study [13] by
David Rolnick and colleagues. They identify high-impact
problems where machine learning in collaboration with other
disciplines, ranging from emergency preparedness to smart (3)

1066

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

Where, RFfi sub(i) denotes feature importance i deduced • The dataset is cleaned and made devoid of null and
from all trees in the model. normfi sub(ij) represents the redundant values.The cleaned dataset is subject to
normalized feature importance for i in tree j. T refers to total normalization.
number of trees.
• Model fitting is done alongside parameter tuning and
D. Polynomial Regression application of deep learning when necessary.
In this regression model, the relationship between the • Model training is carried out for the suitable number of
independent variable x and the dependent variable y is epochs.
denoted as an nth degree polynomial in x. The slope of the
regression model (regression line or curve) is used the • The model is testing and utilized for prediction or fore-
prediction. cast.
y = a0 + a1x + a2x2 + ... + anxn (4) A. Temperature Forecast
where a0, a1, ..., an - coefficients n - degree of the model. The temperature dataset for the states of India was
obtained from Kaggle and subject to exploratory data
E. Linear Regression analysis after the removal of missing values and dropping
It models relationship between dependent and independent null values. This forecast used the ARIMA model, and the 'd'
variables using a linear method. The slope of the regression value was chosen accordingly. A grid search was conducted
model (regression line or curve) is used as the prediction. over a range of p and q values, and as a result, the ARIMA
model with the lowest AIC and BIC was chosen. The values
corresponding to the same were found to be (2,2). Predictions
(5) were made using these parameters applied to the ARIMA
model and the model accuracy was calibrated. Plots were
Where, Yi is the dependent variable f is the function made to find deviation of predicted values from actual values.
Xi is the independent variable ei represents the error terms Temperature forecast was done for the input number of years
under consideration as represented in Fig. 5. The predictions
The above methods were used to analyze climate change were also made using conventional ML models like SVM,
patterns and peaks to determine forest fire risks consequently. Random Forest, Naive Bayes, Logistic Regression, etc. and
compared for a similar dataset. The model implementation
IV.ALGORITHM allows choosing any state of the country and is capable of
Exploratory data analysis of the datasets resulted in the forecasting the temperature patterns over a decade. The
following visual aided conclusions: Fig. 2 represents the temperature peaks can be observed to yield an insight for the
deviation from the trend of temperatures. The general occurrence of forest fires.
procedural steps for forest fire risk prediction are elaborated
B. Weather Prediction based on Rainfall
below as depicted in Fig. 1.
The weather dataset for Australia was obtained from
Kaggle and subject to exploratory data analysis after the
removal of missing values and dropping null values. The date
column was split into date month and year columns
separately via feature extraction. The cleaned dataset was
saved as a separate .csv file. Necessary modules were
imported and Date column of object type was changed to
numeric data type. The rain and non-rain records were
separated.
The following conclusions were made:
1) Maximum Temperature is highly correlated with
Temperature at 3pm.
2) Similar correlation is observed for Temperature at
9amand Minimum Temperature.
3) Pressure and Temperature are negatively correlated.
4) Humidity and Temperature are negatively
correlated.
Rainfall occurrence for the next day i.e., tomorrow is
Fig. 1. Model Architecture predicted with this model. This was done by performing

1067

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

regression with manipulation on the newly added “Rain 1) Carbon Dioxide

tomorrow” column of the dataset.
2) Methane
3) Nitrous Oxide
4) HydroChloroFluoroCarbons (HCFCs)

Correlation between the features were studied and visually

represented using a heatmap as shown in Fig. 4 that led to the
conclusion as to which gas had a major impact.

Fig. 2. Observed deviations

A regression model using Deep learning was used to make

the prediction. The deep learning model was implemented to
observe the model growth whose architecture is illustrated in
Fig. 3. The model has three layers with the hidden layer
having 'ReLu' activation function and the output layer
consisting of one dense layer. The model was compiled and
optimized using Adam optimize. The model was then fitted
and run for 15 epochs. The same was implemented by altering
the number of hidden layers to 3 of 10 dense layers with
'ReLu' activation function and run for 50 epochs. The loss and
accuracy trends were graphically observed. A 0.25 increase
was observed on the usage of deep learning technique to
enhance performance and this temperature and greenhouse Fig. 4. Heatmap for Greenhouse Gas Correlation
gas forecast over the years leads to forest fire risk.
V. RESULTS AND DISCUSSION
The models were used for temperature and greenhouse gas
emission forecast via regression using ARIMA model and
rainfall prediction as mentioned in the previous section. The
results of the temperature rise peak patterns shows the
associated forest fire risk as elaborated below. Table I depicts
the improvement of accuracy on the application of deep
learning techniques. Table II shows the various ML
algorithms employed and a comparison of their metrics. Table
III depicts the metrics observed with ARIMA model and
Table IV depicts the comparison of ML models used for
forecasting.

TABLE I. APPLICATION OF D E E P L E A R N I N G
Fig. 3. Deep Learning Model
Models Accuracy Improved
C. Greenhouse Gas Emission Analysis
Accuracy
The Global Land Temperature dataset from Kaggle was Logistic 0.65 0.87
used which consists information of greenhouse gas emission. Regression
It was subject to pre-processing and feature engineering with
Random 0.62 0.85
respect to the Date column. Null values and non-crucial data
Forest
were dropped. Yearly emission for the 4 major greenhouse
gases was obtained. The four main gases were identified to be:

1068

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

TABLE II. CLASSIFICATION RESULTS COMPARISON OF ML & DL

MODELS

Model Precision Recall F1-Score Accuracy

0 1 0 1 0 1
Logistic 0.88 0.72 0.94 0.52 0.91 0.60 0.85
Regression

Random 0.84 0.83 0.98 0.31 0.90 0.45 0.84

Forest

TABLE III. ARIMA MODEL RESULTS

Fig. 6. Loss Visualization
Metrics Value

MAE 0.6481533654375918
MSE 0.6590247754749318
RMSE 0.6481533654375918

Graphical representation of accuracy boost is observed in

the below graphs: Accuracy trend over subsequent epochs
were analyzed as shown in Fig. 5. Similarly, variation or
reduction inloss is observed in Fig. 6. Deviation of predicted
values from actual values is represented in Fig. 7. where we
can observe that the prediction is quite accurate. Fig. 8.
represents the temperature trend analysis to observe peaks
for the prediction of forest fire occurrences. Based on the
observed temperature peak patterns, conclusions are made on
forest fire risks inherent from the observations. This can help Fig. 7. Predicted Vs Actual Values of temperature
in taking necessary preventative measures to avoid or
mitigate the probable forest fire.

TABLE IV. CLASSIFICATION RESULTS

Metrics Linear Polynomial

Regression Regression
MAE 3.71 3.17
MSE 21.38 21.78
Median 3.19 3.55
Absolute Error
Explain 0.56 0.59
Variance Score
R2 Score 0.56 0.57

Fig. 8. Temperature trend analysis for forest fire risk prediction

VI.CONCLUSION
Global warming is a critical issue that must be addressed
immediately. Assessing global climate change patterns can
help us gain a better understanding of the current situation and
take appropriate action. With the machine learning model, our
project assists with the same and has attained an MSE of
0.659. The accuracy was increased by using a deep learning
model. This climate prediction or forecast will aid us in
adopting the necessary preventative and precautionary
Fig. 5. Accuracy Trend measures to curb and manage the effects of global warming.
The analysis of the Greenhouse gas emission dataset led to the

1069

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

conclusion that HCFCs are the major contributors towards (2021). Global Warming Analysis and Prediction Using Data
increased land temperatures. The extrapolated futuristic Science. [online] IEEE Xplore. https://doi:10.1109/I-
SMAC52330.2021.9640944
predicted temperature and rain fall patterns were a key
[11] Himika, Kaur, S. and Randhawa, S. (2018). Global and Temperature
contributing element in deducing the expected pattern or trend
Prediction by Machine earning Combo Approach. [online]
in global warming giving us an idea of the intensity of the IEEExplore. https://doi:10.1109/ICCCNT.2018.8494173
issue thereby, enabling us to take necessary preventive [12] Sreehari, E. and Srivastava, S. (2018). Prediction of Climate
measures. Based on the observed temperature peak patterns Variable using Multiple Linear Regression. 2018 4th International
from the regression models and forecasts, conclusions can be Conference on Computing Communication and Automation
drawn on forest fire risks inherent from the observations. This (ICCCA).[online]Availabl at: https://www.semanticscholar.org
can help in taking necessary preventative measures to avoid /paper/Prediction-of-Climate-Variable-using-Multiple-Sreehari- Sri
vasta v a/7ba10d831f4281f9907aa598fd11855c8a6f1e35
or mitigate the probable forest fire.
[13] Ardabili, S. et al. (2020) “Deep learning and machine learning in
Future works could involve a better dataset for all three hydrological processes climate change and Earth systems asystematic
aspects. Inclusion of other contributing factors of global review,” Lecture Notes in Networks and Systems, pp. 52–62.
warming that may become significant in the future or remain Available at: https://doi.org/10.1007/978-3-030-36841-8_5.
latent currently may be incorporated. Usage of [14] Mansfield, L.A. et al. (2020) “Predicting global patterns of long- term
climate change from short-term simulations using machine learning,”
meteorological techniques for prediction can be considered. npj Climate and Atmospheric Science, 3(1). Available at:
Boosting accuracy using various methods on the availability https://doi.org/10.1038/s41612-020-00148-5
of proper isolated raw data can yieldbetter results and suggest
appropriate preventive measures basedon the intensity of the
issue in hand.
REFERENCES
[1] G. Yang and X. Di, "Adaptation of Canadian Forest Fire Weather
Index system and it's application," 2011 IEEE International
Conference on Computer Science and Automation Engineering,
Shanghai, China, 2011, pp. 55-58, https://doi:
10.1109/CSAE.2011.5952422
[2] J. S. S. Júnior, J. Pãulo, J. Mendes, D. Alves and L. M. Ribeiro,
"Automatic Calibration of Forest Fire Weather Index For Independent
Customizable Regions Based on Historical Records," 2020 IEEE
Third International Conference on Artificial Intelligence and
Knowledge Engineering (AIKE), Laguna Hills, CA, USA, 2020, pp.
1-8, https://doi: 10.1109/AIKE48582.2020.00011
[3] Huntingford, C. et al. (2019) “Machine Learning and Artificial
Intelligence to aid climate change research and Preparedness,”
Environmental Research Letters, 14(12), p. 124007. Available at:
https://doi.org/10.1088/1748-9326/ab4e55
[4] Kothapalli, S. and Totad, S.G. (2017) “A real-time weather
forecasting and analysis,” 2017 IEEE International Conference on
Power, Control, Signals and Instrumentation Engineering (ICPCSI).
Available at: https://doi.org/10.1109/icpcsi.2017.8391974
[5] Omary, A. et al. (2012) “An interactive predictive system for weather
forecasting,” 2012 International Conference on Computer,
Information and Telecommunication Systems (CITS). Available at:
https://doi.org/10.1109/cits.2012.6220375
[6] Mishra, D. and Joshi, P. (2021) “,” 2021 9th International Conference
on Reliability, Infocom Technologies and Optimization (Trends and
Future Directions) (ICRITO). Available at:
https://doi.org/10.1109/icrito51393.2021.9596117
[7] Reddy, P.C. and Babu, A.S. (2017) “Survey on weather prediction
using Big Data analystics,” 2017 Second International Conference
on Electrical, Computer and Communication Technologies
(ICECCT).Availableat:https://doi.org/10.1109/icecct.2017.8117883
[8] Yuan Quan and Lu Yuchang “Research on weather forecast based on
Neural Networks,” Proceedings of the 3rd World Congress on
Intelligent Control and Automation (Cat. No.00EX393) [Preprint].
Available at: https://doi.org/10.1109/wcica.2000.863402
[9] Biswas, M., Dhoom, T. and Barua, S. (2018) “Weather forecast
prediction: An integrated approach for analyzing and measuring
weather data,” International Journal of Computer Applications,
182(34),pp.20–24.Availableat: https://doi.org/10.5120/ijca 2018918 2
65
[10] Purushotham Reddy, M., Aneesh, A., Praneetha, K. and Vijay, S.

1070

Authorized licensed use limited to: VTU Consortium. Downloaded on November 23,2024 at 07:21:54 UTC from IEEE Xplore. Restrictions apply.