Application of Artificial Neural Network in Wildfire Early Prediction

Systems

HRISTINA NIKOVA, SNEJANA YORDANOVA, RADOSLAV DELIYSKI

Department of Automation,

Technical University of Sofia,

8 Kl. Ohridski blvd., 1000 Sofia,

BULGARIA

conditions and of an output of a prior computed probability for fire. The training and validation datasets consist

of 82 records of real measurement data. The developed and validated ANN can contribute to improvement of

the current forest fire prediction systems.

Key-Words: - wildfire, data processing, neural network, training, validation, prediction.

Received: March 25, 2023. Revised: October 21, 2023. Accepted: December 15, 2023. Published: December 31, 2023.

Forest ecosystems play a very important role for life

on Earth in general, covering nearly one-third of all

are home to up to 80% of terrestrial biodiversity on

Earth. Forests also help in the prevention of erosion,

and enrich and conserve soil.

ecosystem disaster. Since the twentieth century, the

increasing numbers of wildfires have turned them

into the main danger for forest destruction. The

latest research predicts increasing in wildfire

occurrence by 14% by the end of 2030 and even

30% by the end of 2050, [1]. That makes the need

knowledge of fire physics on one hand (theoretical

part) and fire modeling techniques on the other

(practical part), is expected. Despite the notable

advances in technology development in the last

to new and unseen examples. Another significant

advantage is their potential to find solutions to

difficult problems, which are rich in data but poor in

models.

The predicted variables are the burned forest fire

area, [6], [16], [17], [18], the fire danger index

defined by the daily number of forest fires, [19]; fire

spatial and temporal probability expressed as dates

and locations of fire events, etc., [19], [20], [21].

The ANN-based predictors are most often multi-

layer networks trained by BP. All developed

predictors are duly validated.

are suggested in [6]. They are based on Fuzzy

Weather Index System. Due to the limitations of the

fuzzy model, the most significant five variables are

selected - the first four of the basic and Fine Fuel

Moisture codes.

an ANN predictor of the total burned area [16]. The

ANN has one hidden layer of heuristically

determined 36 neurons.

occurrences, defined by a fire danger index on a

selected from six weather variables – the minimal

and the maximal temperatures, the average humidity

of the day, the solar radiation, the average wind

speed, and the cumulative precipitation. The ANN

has three layers with 4 neurons in each hidden layer

and 1 neuron in the output layer. All neuron

activation functions are hyperbolic tangent sigmoid.

The first ANN predictor is trained by Levenberg–

Marquardt BP while the second ANN predictor - by

SVM with a Gaussian kernel function.

images are fed respectively into the historical

records and the data from the previous n-weeks.

SVM is used for classifying the candidate region as

either real fire or non-fire.

specified area and data stations. However, a new

comparative analysis of the advantages and

drawbacks of various methods for wildfire

prediction including ANN, Binary regression as

well as normalized Burning index, Energy Release

Component, and Severe Fire Danger Index is

presented in [22], based on data from the same

period and geographical area.

general approach for predicting forest wildfire

Fig. 1: (a) Topographic location of the area; (b) Average high temperature per month; (c) Average wind

speed per month; (d) Average rainfall per month

moisture content depends on environmental

conditions and vegetation characteristics. Fuel

moisture measures the amount of water contained in

vegetation.

not spread rapidly.

grass, leaves, and small plants or roundwood. Also

the upper layer of litter on the forest floor. In the

Larger fuels (FM100, FM1000) classes are mainly

included roundwood with large sizes.

value equivalent to 0 and upper – to 800, estimating

the amount of precipitation that is needed to bring

the soil back to saturation (a value of 0 means

make the vector pnx10:

pnx10= [Temp, RH, SolR, Rain, maxT, Wind, FM1,

FM100, FM1000, KBDI].

and to ensure that there is no linear dependence

between the different vectors.

vectors (87%) in P72x10 used for predictor modeling

and NV=10 validation vectors (13%) in PV10x10 used

for predictor validation.

consist of the computed for each data vector fire

probability pf, [28].

a single independent variable of binomial

N or 6172 (here S2=1).

ANN weights and biases by gradient based

Levenberg-Marquardt BP.

the update B=[bj] of the biases B starts from the

second layer end moves to the first layer:

-Y1= [Y11…Y1S1] and Y2= [Y21…Y2S2] are the

output vectors of the first and the second layer

respectively with S1 and S2 the number of the

neurons in the first and the second layer;

- pl is the ith input variable (i=1R);

- k is the number of epochs;

concerning n2,

- the weights W1li and the biases b1 are determined

analogically, f1(n1) with

- α is a learning rate.

convergence of the tuning procedure for the weights

and the biases (2)-(5) by introducing second order

couples (P72x10, T72x1), (PV10x10, TV10x1), the number

of layers and neurons in each layer, the activation

functions, the training method, the cost function and

the end condition.

Fig. 5: General algorithm for development and

validation of multi-layer ANN predictor

neural network with 5 neurons in the hidden layer

with no presence of fire (values for pf close to 0).

The input data, the target, and the predicted output

probability are shown in Table 1.

line represent the target values TV for pf whereas the

starts, the time for simulation after training and for

validation.

4 Discussion and Analysis

execution of the algorithm. It is in the range [520]

minutes for the design of several ANN fire

predictors using a medium powerful computer and

MATLABTM Release 13.

However, the time for designing an ANN

predictor is of small importance since ANN

ANN design. A derivation of а new ANN predictor

may be required when thoroughly new conditions

arise.

5 Conclusions and Future Research

The novelty and the main contributions of the

presented research are concluded in the following.

and validation of a multi-layer ANN predictor is

and the fuel moisture condition.

with higher accuracy (error <5%) than the

regression prediction model from [28]. Besides, it

uses only a selected limited input data sample of 82

records from 10 years in the region of Whitmore,

North California, USA. The comprehensive and

low-complexity algorithm ensures fast processing.

To this, contribute the ANN simple architecture, the

fast-converging training method, and the small