applied

sciences
Article
A Highly Accurate Forest Fire Prediction Model Based on an
Improved Dynamic Convolutional Neural Network
Shaoxiong Zheng 1 , Peng Gao 1 , Weixing Wang 1,2, * and Xiangjun Zou 3,4, *

1 College of Electronic Engineering, South China Agricultural University, Guangzhou 510642, China;
[email protected] (S.Z.); [email protected] (P.G.)
2 Guangdong Engineering Research Center for Monitoring Agricultural Information, Guangzhou 510642, China
3 Guangdong Laboratory for Lingnan Modern Agriculture, College of Engineering, South China Agricultural
University, Guangzhou 510642, China
4 Foshan-Zhongke Innovation Research Institute of Intelligent Agriculture and Robotics, Foshan 528000, China
* Correspondence: [email protected] (W.W.); [email protected] (X.Z.)

Abstract: In this work, an improved dynamic convolutional neural network (DCNN) model to
accurately identify the risk of a forest fire was established based on the traditional DCNN model.
First, the DCNN network model was trained in combination with transfer learning, and multiple
pre-trained DCNN models were used to extract features from forest fire images. Second, principal
component analysis (PCA) reconstruction technology was used in the appropriate subspace. The
constructed 15-layer forest fire risk identification DCNN model named “DCN_Fire” could accurately
identify core fire insurance areas. Moreover, the original and enhanced image data sets were used to
evaluate the impact of data enhancement on the model’s accuracy. The traditional DCNN model was
improved and the recognition speed and accuracy were compared and analyzed with the other three
DCNN model algorithms with different architectures. The difficulty of using DCNN to monitor forest
fire risk was solved, and the model’s detection accuracy was further improved. The true positive
Citation: Zheng, S.; Gao, P.; Wang, rate was 7.41% and the false positive rate was 4.8%. When verifying the impact of different batch
W.; Zou, X. A Highly Accurate Forest sizes and loss rates on verification accuracy, the loss rate of the DCN_Fire model of 0.5 and the batch
Fire Prediction Model Based on an size of 50 provided the optimal value for verification accuracy (0.983). The analysis results showed
Improved Dynamic Convolutional that the improved DCNN model had excellent recognition speed and accuracy and could accurately
Neural Network. Appl. Sci. 2022, 12, recognize and classify the risk of a forest fire under natural light conditions, thereby providing a
6721. https://doi.org/10.3390/ technical reference for preventing and tackling forest fires.
app12136721

Academic Editor: Antonella Keywords: forest fire; deep learning; transfer learning; convolutional neural network
D’Alessandro

Received: 2 June 2022

Accepted: 1 July 2022
1. Introduction
Published: 2 July 2022
Forests are an indispensable link in the ecological chain and an important ecosystem in
Publisher’s Note: MDPI stays neutral
nature, and damaging them can seriously affect the Earth’s ecological environment. Forest
with regard to jurisdictional claims in
protection plays an important role in sustainable human development, and preventing
published maps and institutional affil-
forest fires is of the utmost importance [1].
iations.
Research shows that forest fire risk monitoring plays a key role in preventing forest
fires [2,3]. The steps for monitoring them are (1) identifying hot spots, (2) assessing fire risks,
(3) identifying areas vulnerable to forest fires, (4) recognizing manmade infrastructure, and
Copyright: © 2022 by the authors.
(5) examining the meteorological conditions. These are considered to be influential param-
Licensee MDPI, Basel, Switzerland. eters that play an important role when establishing a model to assess the risk of a forest
This article is an open access article fire [4–7]. Generally speaking, the current identification method of forest fire insurance is
distributed under the terms and traditional and singular. At present, geographic information systems (GIS), artificial intelli-
conditions of the Creative Commons gence (AI), and machine learning (ML) can be used to predict the risk of a forest fire [8,9].
Attribution (CC BY) license (https:// Stula M et al. [10] proposed a forest fire recognition algorithm based on a convolutional
creativecommons.org/licenses/by/ neural network (CNN) that can extract and recognize image features, thereby providing
4.0/). excellent recognition accuracy. It was implemented in a forest fire video monitoring system

Appl. Sci. 2022, 12, 6721. https://doi.org/10.3390/app12136721 https://www.mdpi.com/journal/applsci

Appl. Sci. 2022, 12, 6721 2 of 15

that could achieve good recognition results. Ciprián-Sánchez et al. [11] extracted and
combined information from various imaging modes for detecting and extracting features,
and predicting the spread of wildfires via a state-of-the-art (SOTA) model based on deep
learning (DL); the architecture, loss function, and different types of image combinations
of the SOTA model were evaluated to identify the parameters related to improving the
segmentation results and comparing them with traditional fire segmentation technology,
thereby improving the ability to identify the risk of a forest fire. Moayedi H et al. [12] used
a hybrid evolutionary algorithm to build a forest fire risk prediction model; a forest fire
sensitivity map of fire-prone areas in Iran was generated based on combining an adaptive
neuro-fuzzy inference system, a genetic algorithm, particle swarm optimization, and differ-
ential evolution algorithm with reliable accuracy. Vikram R et al. [13] used a support vector
machine (SVM) for a semi-supervised classification model to divide the forest area into
different subareas: high activity (HA), medium activity (MA), and low activity (LA). Due
to energy limitations, a sensor node that only monitored one parameter could predict the
fire risk in different areas with 90% accuracy. Moreover, greedy forwarding technology was
used to continuously and periodically send data packets from HA and MA areas to the base
station, respectively, but prevented them being sent from LA areas. This data forwarding
technology improved the network life and reduced congestion during data transmission
from forest areas to the base station. Achu A L et al. [14] proposed an automatic fire
monitoring system to identify forest fires in their early stages that could predict the size of
the flame areas. A deep CNN was used for forest fire risk image classification to extract
descriptors from images and was then applied to a logistic regression classifier. After using
882 images to form an ordered data set and related image metadata (such as flame, smoke,
fog, cloud, and human elements) during testing, the classification accuracy of 695 daytime
and 187 night-time scene images was 94.1% and 94.8%, respectively, with good accuracy in
estimating the flame area. Pham B T et al. [15] compared image frames via a segmentation
method based on distance measurements in a color channel histogram and a fast frame
comparison and extraction algorithm based on time. The module uses key advances in
video frame technology and processes the frames through a deep CNN that was trained
via normalization prior to segmentation to assist in detecting fire and smoke. The fire risk
recognition accuracy was 90%. Michael Y et al. [16] used two pre-trained DL CNN models
(vgg16 and resnet50) for fire detection and enhanced the depth of network learning by
fine-tuning the network on the basis of a fully connected layer. Although this improved the
detection accuracy, it also increased the network training time. However, the deep CNN
performed well on complex data sets.
At present, research on forest fire risk detection based on ML mainly involves flame
detection and smoke detection [17–21]. The forest fire risk monitoring methods proposed
by most scholars are for specific scenes, with several problems that need to be solved.
How to recognize the color, form, and texture of fire in different chaotic backgrounds is
the key to forest fire risk recognition [22–26]. The purpose of this study was to effectively
extract the complex upper layer features of fire risk images and improve the robustness of
input conversion by constructing an unmanned aerial vehicle (UAV) image-based forest
fire risk prediction model based on a deep learning back-propagation neural network. The
problems to be solved in the process of fire risk identification are feature extraction from
training image data sets and training with a large number of fire risk image data sets [27,28].
In this study, we attempted to solve the above problems as follows. First, the dynamic
CNN (DCNN) network was trained in combination with transfer learning, and multiple
pre-trained DCNNs were used to extract features from forest fire images. The traditional
DCNN model was improved and the recognition speed and accuracy were compared and
analyzed with the other three DCNN model algorithms with different architectures. The
difficulty of using DCNN to monitor forest fire risk was solved and the model’s detection
accuracy was further improved. Therefore, the forest fire risk model based on the improved
DCNN could accurately identify and classify the risk of a forest fire under natural light
conditions. The main research contents of this paper were as follows:
Appl. Sci. 2022, 12, 6721 3 of 15

(1) The DCNN network was trained in combination with migration learning, and multiple
pre-trained DCNN models were used to extract the features of forest fire images.
(2) Principal component analysis (PCA) reconstruction technology was used to enhance
image category differentiation.
(3) A 15-layer DCNN model called “DCN_Fire”, as an improvement on the traditional
DCNN model, was constructed and analyzed.
(4) The recognition speed and accuracy of the improved DCNN model were compared
with three other DCNN models with different architectures.
(5) The DCNN model could accurately recognize the risk of a forest fire risk in natural
light; thus it was suitable as an early warning system for forest fires.
In Section 2, we discuss the extraction and enhancement of pertinent image features
using the DCNN. Section 3 presents the improved DCNN model and its construction
process. Section 4 provides an analysis of the model verification and experimental results.
Finally, conclusions based on the study findings are presented in Section 5.

2. Model Construction and Analysis

2.1. Model Construction
Here, we comprehensively evaluate the self-learning ability of a DL CNN. A DCNN
model can be used to determine whether there is a fire risk from forest images collected by
an UAV by evaluating the influence of different model structural parameters. The proposed
improved DCNN model was used to describe the DL CNN model accurately and in detail.

2.1.1. Overall Architecture

Figure 1 shows the architecture of the improved DCNN model. The model architecture
consisted of 15 layers of DCNNs collectively called “DCN_Fire”. The structural parameters
are described below. The first 12 layers of the DCNN model were composed of eight
convolutions (Conv 1–8) in the convolutional layer and four maximum pooling layers (Pool
1–4). Forest fire features were extracted from a convolutional layer on a lower level and
transferred to a higher level. The maximum pooling layers could extract the changeable part
of the image and reduce the output size of the convolutional layer. The last classification
layer recognized images of fires according to the previously extracted high-level features.
Appl. Sci. 2022, 12, 6721 4 of 16
The model structure was suitable for extracting different forest flame features from the
training data set.

Figure1.1.The
Figure Thearchitecture
architectureof
ofthe
theDCN_Fire
DCN_Firemodel.
model.

2.1.2. Details of the DCN_Fire Model

Input layer: input data comprised a segmentation database of significant images
with a pixel size of 128 × 128 in lightning memory-mapped database format.
Convolutional layer: eight convolutions (Conv 1–8) were defined in the convolu-
tional layer. The convolutional kernel size was 3 × 3, the sliding step was one, and Xavier
Appl. Sci. 2022, 12, 6721 4 of 15

2.1.2. Details of the DCN_Fire Model

Input layer: input data comprised a segmentation database of significant images with
a pixel size of 128 × 128 in lightning memory-mapped database format.
Convolutional layer: eight convolutions (Conv 1–8) were defined in the convolutional
layer. The convolutional kernel size was 3 × 3, the sliding step was one, and Xavier was
used to initially configure the parameter weights. A rectified linear unit (relu) was used
on this layer as the activation function and the output was sent to the next volume or
pooling layer.
Pooling layers: these reduced the data dimensions to avoid over-fitting problems. The
kernel size of this layer was set to 2 × 2 by using the maximum pooling method, and the
step size was two. Each fully connected layer (FC1 and FC2) had 256 neurons and Xavier
mode was used for weight initialization. In the hidden layer, the exit technique was used to
reduce the dependence between neurons, and the activation function in the fully connected
layer was a relu.
Classification layer: the Softmax function is calculated as

exp( ai )
So f tmax = (1)
∑ j exp a j

Pixels of size 128 × 128 were input into the DCNN, after which the error between the
actual output value and the required value was calculated. The stochastic gradient descent
method was used on the training-depth CNN. After passing the back-propagation error in
the previous layer, the method calculated the derivative on the optimization coefficient and
weight. The learning process of the model was iteratively optimized until the training error
was small enough to be ignored.

2.2. Transfer Learning

Due to the depth of training, neural networks needed labeled data and a large number
of special hardware devices, such as high-powered graphics processors. Training a forest
fire risk identification model takes a long time and using transfer learning can solve this
problem [29–31].
As the latest technology for visual classification, a DCNN pre-trained on large data
sets can be reused for new tasks. A DCNN with learning weights can be directly applied
to various computer vision situations [32,33]. In this study, the pre-trained DCNN model
was used to extract features from forest fire images, which provided effective help for
multi-feature fusion; thus, it provided robust class identification.
In the three training structures (AlexNet, VGG-16, and Inception-v3), the class predic-
tion layer of each DCNN was used as a feature extractor. Each layer provided the output
for 1000 neurons, which were the neglected top layer in the network.
We used a computer with an Intel® CoreTM i7-3770 CPU with Inception-v3 to obtain
bottleneck functions for the published benchmark cifar10 data set [34]. Table 1 summarizes
the various architectures of the models.

Table 1. The architectures of the models.

Structure Name Input Dimensions The Penultimate Floor Feature Size

AlexNet 227 × 227 × 3 FC7 4096
VGG16 224 × 224 × 3 FC2 4096
Inception-V3 any × any × 3 Pool_3 2048

2.3. Feature Extraction and Classification

For the classification and recognition of forest fire risk images, single multimodal
features cannot deal with large changes in the statistical data of images [35–37]. We
Appl. Sci. 2022, 12, 6721 5 of 15

combined classification technology with various features [38] to improve the accuracy of
recognizing forest fire risk from images. These features could be expressed as
n om
Fi = f ik (2)
k =1

for i = 1, 2, · · · · · · , D0 and k = 1, 2, · · · · · · , m. Multi-feature fusion is mainly used to

determine a generalized subspace ∏ ∈ D0 ×d for distribution d << D0 , d = l, or d < l.
A single feature is normalized according to its energy level as follows:

F 0 i = Ωi Fi (3)

where   m
1 1
Ωi = /∑ (4)
EFi i =1
EFi

and Fi is the original feature of f ( x ) = max (0, x ). The value of the series is calculated
by using

T
FFV = F 0 1 , · · · · · · , F 0 m (5)
The quantity product of the characteristics is calculated as follows:

FFV = ∏ im=1 F0 i (6)

The summation formula for the characteristics can be written as

m
FFV = ∑ F0 i (7)
i =1

Calculating the average pooling value can be achieved by using

 
T T
FFV = mean F 0 1 , · · · , F 0 m (8)

The maximum pooling value is calculated by using

 
T T
FFV = max F 0 1 , · · · , F 0 m (9)

The influence of the features is eliminated and the richness of the fusion feature sub-
space is enhanced in the above feature fusion methods [38]. We used reconstruction tech-
nology based on PCA to enhance the distinction between specific categories. The variables
x and y are combined as ( x1 , y1 ), ( x2 , y2 ), · · · , ( xk , yk ), where xi ∈ <n ,yi ∈ {1, 2, · · · , M } is
the real class label of xi estimated according to its density pi ( x ).

ŷ = argmaxpi ( x ),
(10)
i = 1, · · · , M

where pi ( x ) = ∑ kj=1 y j λij C x, x j + bi when x is positive (the greater the value, the higher

the confidence level), where λ is the coefficient term, C represents the penalty function, and
b is the deviation parameter.
We adopted a linear SVM based on scikit-learn in Python. The bottleneck feature of
DCNN is usually linearly separable, and so the features of the fused DCNN model could
be processed well by using a linear SVM.

3. Comparative Analysis of the Algorithm Implementation Performance

3.1. Data Source
We used more than 4000 forest fire risk images as the original data that ranged in size
from 200 × 200 to 4000 × 4000 pixels and were taken in the Guangdong Longshan and
 We adopted a linear SVM based on scikit-learn in Python. The bottleneck feature of
DCNN is usually linearly separable, and so the features of the fused DCNN model could
be processed well by using a linear SVM.

3. Comparative Analysis of the Algorithm Implementation Performance

Appl. Sci. 2022, 12, 6721 6 of 15
3.1. Data Source
We used more than 4000 forest fire risk images as the original data that ranged in size
from 200 ×Sihui
Jiangmen 200 to 4000farms.
forest × 4000Thepixels andimages
input were taken
were in the Guangdong
normalized Longshan
to produce and Jiang-
a standard data
men Sihui forest farms. The input images were normalized to produce
set. UAV images usually contain cluttered information, and so some important features a standard data set.
UAV
can imagessubmerged
become usually containwhencluttered
their size information, and so some important features can be-
is directly adjusted.
comeWe submerged whenlocate
could quickly their size
the is directly
core adjusted.
fire area in a forest fire risk image despite a complex
We could
background byquickly
dividing locate the core fire
the detection area
area intoinaarecommended
forest fire risk area
image anddespite a complex
a selection area
background
by by dividing
using significance the detection
detection and a area intoregression
logistic a recommended areaThe
classifier. andspecific
a selection area
steps by
were
using significance
as follows. detection and a logistic regression classifier. The specific steps were as
follows.
Step 1 The region of interest (ROI) proposal. Bayesian-based saliency detection was
Step 1 The region
applied to of
aninterest
original(ROI)
image proposal. Bayesian-based
I; the saliency was mapped saliency
to thedetection was applied
planarization in the
to an original image I; the saliency was mapped to the planarization
image mask IM; and the original image was masked to areas R1–Rn. in the image mask IM;
and
Stepthe original
2 The colorimage was masked
components, image toenergy,
areas R1–Rn.
and entropy (vectors X1–Xn) of each ROI were
Step 2 The color components,
determined image energy,
to decide whether and entropyvector
the characteristics (vectorsin X1–Xn)
the ROIof each ROI were
corresponded to
determined to decide whether the characteristics vector in the ROI
flame or smoke, after which MBR and a fixed size boundary were used to dividecorresponded to flame or
smoke, after which
the fire area.MBR and a fixed size boundary were used to divide the fire area.
We used the
We used the color
color matrix
matrixvaluevalueasasthethecolor
colorfeature
featuredescriptor
descriptor andandthethe
energy andand
energy en-
tropy of the angular second-order matrix of an image as the texture
entropy of the angular second-order matrix of an image as the texture descriptor. descriptor.

3.2. Significance
SignificanceSegmentation
Segmentationand
andData
DataEnhancement
Enhancement
We used
We used significance
significance detection
detection to
to segment
segment the
the flame
flame image
image to
to standard
standard sizes
sizes and
and train
train
the model.
model.The
Themodel
model was used
was to process
used images
to process withwith
images flames, smoke,
flames, or flames
smoke, and smoke.
or flames and
Figure 2Figure
smoke. shows2the processing
shows results. results.
the processing

Appl. Sci. 2022, 12, 6721 7 of 16

(a)

(b)

(c)

(d)
Figure2.2. Image
Figure Image segmentation
segmentation based
based on
on saliency
saliency detection.
detection. (a)
(a) Original
Original images;
images;(b)
(b)significance
significancetest
test
results;(c)
results; (c)ROI
ROIselection
selectionresults;
results;and
and(d)
(d)standard
standardsegmentation
segmentationresults.
results.

Figure 2a shows the original forest fire images while Figure 2b exhibits the results of
significance detection. As shown in Figure 2c, the proposed significance method effec-
tively located the core flame area in the aerial images and even very small ignition areas.
In addition, Figure 2d shows that the original image was segmented into independent
 (d)
Appl. Sci. 2022, 12, 6721 7 of 15
Figure 2. Image segmentation based on saliency detection. (a) Original images; (b) significance test
results; (c) ROI selection results; and (d) standard segmentation results.

Figure2a
Figure 2ashows
showsthetheoriginal
originalforest
forestfire
fireimages
imageswhile
whileFigure
Figure2b2bexhibits
exhibitsthe
theresults
resultsof
of
significancedetection.
significance detection.AsAsshown
shown in in Figure
Figure 2c, 2c,
the the proposed
proposed significance
significance methodmethod effec-
effectively
tively located
located theflame
the core core flame
area inarea
thein the aerial
aerial images images and even
and even veryvery
smallsmall ignition
ignition areas.
areas. In
In addition,
addition, Figure
Figure 2d shows
2d shows thatoriginal
that the the original
imageimage was segmented
was segmented into independent
into independent flame
flame images
images containing
containing complete complete flame using
flame features featurestheusing the proposed
proposed method.
method. Data Data en-
enhancement
hancement
could could
effectively effectively
solve solve the
the over-fitting over-fitting
problem causedproblem
by usingcaused
a smallby using data
training a small
set.
Figure 3 shows
training examples
data set. Figure 3ofshows
the test results. of the test results.
examples

(a)

(b)

Appl. Sci. 2022, 12, 6721 8 of 16

(c)

(d)

(e)

(f)
Figure 3.
Figure 3. Examples
Examples of
of fire
fire location
location and
and segmentation
segmentationscenes.
scenes. (a–f)
(a–f) Examples
Examples1–6.
1–6.

We tested the performance of the fire risk location and segmentation algorithm with
550 images of flames selected from the original data set. The test results are reported in
Table 2. From the test results, the true positive rate was 7.41% and the false positive rate
was 4.8%. Since the images were small, the characteristics of flames and smoke covered
Appl. Sci. 2022, 12, 6721 8 of 15

We tested the performance of the fire risk location and segmentation algorithm with
550 images of flames selected from the original data set. The test results are reported in
Table 2. From the test results, the true positive rate was 7.41% and the false positive rate
was 4.8%. Since the images were small, the characteristics of flames and smoke covered
most areas. At this time, segmentation was not necessary. To enhance the recognition of
fire risk images, the data set was enhanced. The new enhanced image database named
“DB_Fire” contained 3845 images. Table 3 shows the parameters of the original image
database and enhanced image database.

Table 2. The performances of the location and segmentation methods.

Image
Total ROI Detected from 240 Aerial Fire Images 492 Total ROI Detected from 310 Normal-View Images 683
Type
TP FN FP TN TPR FPR TP FN FP TN TPR FPR
Performance
329 19 8 136 96.1% 4.8% 463 47 16 157 92.2% 7.41%
TP, true positive; TN, true negative; FP, false positive; FN, false negative; TPR, the true positive rate; FPR, the false
positive rate.

Table 3. The image databases.

Original Image Database Enhanced Image Database

Image Type UAV Image Ordinary UAV Image Ordinary
Total Total
Acquisition Image Acquisition Image
No flames 354 762 1116 748 1476 2224
With flames 247 497 744 457 1164 1621
Total 601 1259 1860 1205 2640 3845

3.3. Optimization of the DCN_Fire Parameters

The DCN_Fire model parameters included the convolutional core size, number of
layers, and loss rate. Some empirical values were determined (the input image size was
128 × 128 × 128) after the full-parameter tuning test. Good results were obtained in terms
of verification accuracy for a learning rate of 0.0001, a momentum of 0.9, and 20,000 itera-
tions (Table 4).

Table 4. The influence of batch size and loss rate on the verification accuracy of DCN_Fire.

Batch Size Verification Accuracy Loss Rate Verification Accuracy

50 0.968 0.5 0.983
64 0.971 0.6 0.967
48 0.972 0.4 0.964
128 0.965 0.7 0.952
32 0.961 0.3 0.949
16 0.949 0.2 0.947

4. Model Validation
4.1. Model Performance Verification
To verify the model’s accuracy, we explored several forest fire risk monitoring DCNN
models with various hierarchical structures and calculated the image processing time and
verification accuracy of each one using the original and enhanced image data sets. The
models involved in the comparison included the DCN_Fire model and 13-layer, 15-layer,
and 20-layer DCNN models. Figure 4 shows the comparison of flame image verification
accuracy. Figure 5 shows the comparison of flame image processing times by the models.
 To verify the model’s accuracy, we explored several forest fire risk monitoring
DCNN models with various hierarchical structures and calculated the image processing
time and verification accuracy of each one using the original and enhanced image data
sets. The models involved in the comparison included the DCN_Fire model and
13-layer, 15-layer, and 20-layer DCNN models. Figure 4 shows the comparison of flame
Appl. Sci. 2022, 12, 6721 9 of 15
image verification accuracy. Figure 5 shows the comparison of flame image processing
times by the models.

Appl. Sci. 2022, 12, 6721 10 of 16

Figure 4.4. Flame
Figure Flame image
imageverification
verificationaccuracy
accuracycomparison of of
comparison thethe
DCN_Fire model
DCN_Fire withwith
model 13-layer,
13-layer,
15-layer,and
15-layer, and20-layer
20-layerDCNN
DCNN models.
models.

Figure
Figure 5.5.Flame
Flameimage
image processing
processing time
time comparison
comparison ofDCN_Fire
of the the DCN_Fire
modelmodel and 13-layer,
and 13-layer, 15-layer, 15-layer
and 20-layerDCNN
and 20-layer DCNN models.
models.

Although
Althoughthe 20-layer
the DCNN
20-layer DCNN model achieved
model a verification
achieved accuracyaccuracy
a verification similar to similar
that of to that
the DCN_Fire model (0.983), its processing time was 1.67 longer due to
of the DCN_Fire model (0.983), its processing time was 1.67 longer due to the larger the larger number
of layers, increasing the calculation time and cost. Therefore, the DCN_Fire model was
number of layers, increasing the calculation time and cost. Therefore, the DCN_Fire
superior in terms of processing time and verification accuracy. Meanwhile, the original
model was superior
and enhanced image datain terms
sets wereof used
processing timethe
to evaluate and verification
impact accuracy. Meanwhile
of data enhancement on
the original and enhanced image data sets were used to evaluate
model accuracy. The data revealed that the verification accuracy of the same model the impact of data en-
using
hancement
the enhancedon model
image dataaccuracy. Thethan
set was better data revealed
that using thethat the verification
original image data set.accuracy of the
sameAfter
model using the enhanced image data set was better than that
dividing the image samples into training and verification sets, the DCN_Fire using the original
model and
image datathe 13-layer and 15-layer DCNN models were trained and their loss values were
set.
calculated. Figure
After dividing 6 shows the training
the image and verification
samples into traininglossand
curves of the foursets,
verification models.
the DCN_Fire
The loss curves is showen in Figure 6, it mark the training
model and the 13-layer and 15-layer DCNN models were trained and their loss once every 100 iterations
loss values
and the validity loss once every 500 iterations.
were calculated. Figure 6 shows the training and verification loss curves of the four
The data from Figures 4–6 show that the DCN_Fire model had better recognition
models.
accuracy than the other models with minimal training and verification losses.
 hancement on model accuracy. The data revealed that the verification accuracy of the
same model using the enhanced image data set was better than that using the original
image data set.
After dividing the image samples into training and verification sets, the DCN_Fire
model and the 13-layer and 15-layer DCNN models were trained and their loss values
Appl. Sci. 2022, 12, 6721 10 of 15
were calculated. Figure 6 shows the training and verification loss curves of the four
models.

Appl. Sci. 2022, 12, 6721 11 of 16

(a)

(b)

(c)
Figure 6.
Figure 6. Training and validation
Training and validation loss
loss curve
curve of
of (a) the DCN_Fire
(a) the DCN_Fire model,
model, (b)
(b) the 13-layer DCNN
the 13-layer DCNN
model, and (c) the 13-layer DCNN model.
model, and (c) the 13-layer DCNN model.

The lossofcurves
4.2. Analysis is showenby
Fire Recognition inthe
Figure 6, it mark
DCN_Fire Modelthe training loss once every 100 itera-
tionsFigure
and the validity loss once every 500 iterations.
7 shows the visual characteristics after the input images had passed through
The data
the Conv 2 andfrom
ConvFigures 4–6 Itshow
3 layers. couldthat
be the DCN_Fire
observed model
that the lowerhadlayer
better recognition
captured low-
accuracy than the other models with minimal training and verification losses.

4.2. Analysis of Fire Recognition by the DCN_Fire Model

Figure 7 shows the visual characteristics after the input images had passed through
the Conv 2 and Conv 3 layers. It could be observed that the lower layer captured
Appl. Sci. 2022, 12, 6721 11 of 15

Appl. Sci. 2022, 12, 6721 12 of 16

level texture, edges, and color features while the higher layer captured sparse higher-level
features by eliminating irrelevant content.

(a) (b) (c)

Figure 7.
Figure 7. The
The feature
featuremapping
mappingprocess
processafter
afterthe Conv
the 2 and
Conv Conv
2 and Conv3 layers in the
3 layers in DCN_Fire model.
the DCN_Fire model.
(a) The original input image, (b) the Conv 2 feature map, and (c) the Conv 3 feature
(a) The original input image, (b) the Conv 2 feature map, and (c) the Conv 3 feature map. map.

The DCN_Fire
The DCN_Fire model
modelwaswasusedusedtotoidentify
identifyforest firefire
forest images
imagesandand
verify the effective-
verify the effective-
ness and rationality of the classification of the forest fire risk from aerial
ness and rationality of the classification of the forest fire risk from aerial images images (Figure 8). 8).
(Figure
Since the
Since thesmoke
smokefeatures
featuresininthe
thesmoke
smoke(Figure
(Figure8c)8c)
andandmistmist (Figure
(Figure 8d)8d) images
images were
were highly
highly similar, DCN_Fire mistakenly classified forest mist images as fire
similar, DCN_Fire mistakenly classified forest mist images as fire risk images. This was not risk images.
This was not surprising as humans can also mistakenly identify mist as smoke.
surprising as humans can also mistakenly identify mist as smoke.
The test set consists of 1200 random images with and without flames from the database.
From the results in Table 5, it could be seen see that the false negative rate with the DCN_Fire
model (0.13%) was very low.

Table 5. The performance of the DCN_Fire model on the test data set.

TP FN TN FP TPR TNR FNR General Accuracy

(a)
592 8 588 12 98.7% 98.0% 0.13% 98.3%
TP, true positive; TN, true negative; FP, false positive; FN, false negative; TPR, the true positive rate; TNR, the true
negative rate; FNR, the false negative rate.

The same training and test data sets were used to verify and compare the recognition
accuracy of the DCN_Fire model and other mainstream ML models (Kim’s CNN, AlexNet,
8-layer CNN + Fisher vector, HOG + (b)SVM, and deep belief net + neural net) for fire risk
recognition. The training results showed that DCN_Fire was more accurate than the other
models and performed well (Figure 9, Table 6).

Table 6. DCN_ Fire compared with the accuracy of other models.

Model Classifier Accuracy

DCN_ Fire (c) Softmax 98.3%
Kim’s CNN model Softmax 93.2%
AlexNet Softmax 97.3%
Eight-layer CNN + Fisher vector SVM 95.1%
HOG + SVM SVM 43.3%
Deep belief net + neural net BPNN 87.7%

(d)
 (a) The original input image, (b) the Conv 2 feature map, and (c) the Conv 3 feature map.

The DCN_Fire model was used to identify forest fire images and verify the effective-
ness and rationality of the classification of the forest fire risk from aerial images (Figure 8).
Appl. Sci. 2022, 12, 6721
Since the smoke features in the smoke (Figure 8c) and mist (Figure 8d) images12 were of 15
highly similar, DCN_Fire mistakenly classified forest mist images as fire risk images.
This was not surprising as humans can also mistakenly identify mist as smoke.

(a)

Appl. Sci. 2022, 12, 6721 13 of 16

(b)DCN_Fire model. (a) Images without fire charac-

Figure 8. Fire risk image classification using the
teristics; (b) images with flames; (c) images with smoke; and (d) misclassified images with mist.

The test set consists of 1200 random images with and without flames from the da-
tabase. From the results in Table 5, it could be seen see that the false negative rate with
the DCN_Fire model (0.13%) was very low.

Table 5. The performance of the DCN_Fire model on the test data set.
(c)
TP FN TN FP TPR TNR FNR General Accuracy
592 8 588 12 98.7% 98.0% 0.13% 98.3%
TP, true positive; TN, true negative; FP, false positive; FN, false negative; TPR, the true positive
rate; TNR, the true negative rate; FNR, The false negative rate.

The same training and test data sets were used to verify and compare the recogni-
tion accuracy of the DCN_Fire model and other mainstream ML models (Kim’s CNN,
(d)
AlexNet, 8-layer CNN + Fisher vector, HOG + SVM, and deep belief net + neural net) for
fire risk8.recognition.
Figure Theclassification
Fire risk image training results
usingshowed that DCN_Fire
the DCN_Fire model. (a)was more
Images accurate
without fire than
character-
the other models and performed well (Figure 9, Table 6).
istics; (b) images with flames; (c) images with smoke; and (d) misclassified images with mist.

Figure
Figure 9.
9. Accuracy
Accuracycomparison
comparisonofof
the DCN_Fire
the model
DCN_Fire andand
model other mainstream
other ML ML
mainstream models.
models.
Appl. Sci. 2022, 12, 6721 13 of 15

5. Conclusions
The detection and recognition of forest fire risk images is of great significance for the
detection and prevention of forest fires. Combined with migration learning, the work in this
paper trained a DCNN network, used multiple pre-trained DCNN models to extract the
features of forest fire images, used PCA reconstruction technology to convert features into
a shared feature subspace, and established the forest fire risk recognition model. According
to the experimental results, the main research conclusions of this paper were as follows:
(1) A DCNN network combined with migration learning was constructed and trained
using multiple pre-trained DCNN models to extract features from forest fire images
and used PCA reconstruction technology to convert features into a shared feature
subspace to establish a forest fire risk recognition model. We used 550 flame images
to test the fire risk location performance and the segmentation algorithm. The true
positive rate was 7.41% and the false positive rate was 4.8%. When verifying the
impact of different batch sizes and loss rates on verification accuracy, the loss rate of
the DCN_Fire model of 0.5 and the batch size of 50 provided the optimal value for
verification accuracy (0.983).
(2) Comparing the test results showed that the performance of the improved DCNN
model was comparable with that of other models. In terms of processing time and
verification accuracy, the DCN_Fire model was considered a better DCNN architecture
in fire risk identification. When calculating the training and verification losses of
several models, those of the DCN_Fire model were minimized.
Although the improved DCNN model achieved good results in the detection of forest
fire risk, the detection accuracy still warrants improvement. In future research, we will
further optimize the network model structure and improve the network performance of
the forest fire risk identification model.

Author Contributions: Conceptualization, S.Z. and W.W.; methodology, S.Z., P.G., W.W. and X.Z.;
software, S.Z. and X.Z.; validation, S.Z., W.W. and X.Z.; formal analysis, S.Z., P.G. and X.Z.; investiga-
tion, S.Z., W.W. and X.Z.; resources, S.Z., W.W. and X.Z.; data curation, X.Z. and P.G.; writing—original
draft preparation, S.Z.; writing—review and editing, W.W. and X.Z.; visualization, S.Z. and X.Z.;
supervision, W.W., X.Z. and S.Z.; project administration, W.W. and X.Z. All authors have read and
agreed to the published version of the manuscript.
Funding: This work was supported by the Key Field R&D Program Project of Guangdong Province,
China (Grant No. 2019B020223003), Guangzhou Science and Technology Plan Project Innovation
Platform Construction and Sharing (Grant No. 201605030013), Guangdong Laboratory of Lingnan
Modern Agriculture Project under Grant NT2021009 and the No. 03 Special Project and the 5G Project
of Jiangxi Province under Grant 20212ABC03A27.
Acknowledgments: We would like to thank our anonymous reviewers for their critical comments
and suggestions for how to improve our manuscript.
Conflicts of Interest: The authors declare that there are no conflict of interest.

References
1. Tariq, A.; Shu, H.; Siddiqui, S.; Mousa, B.G.; Munir, I.; Nasri, A.; Waqas, H.; Lu, L.; Baqa, M.F. Forest fire monitoring using
spatial-statistical and Geo-spatial analysis of factors determining forest fire in Margalla Hills, Islamabad, Pakistan. Geomat. Nat.
Hazards Risk 2021, 12, 1212–1233. [CrossRef]
2. Vigna, I.; Besana, A.; Comino, E.; Pezzoli, A. Application of the socio-ecological system framework to forest fire risk management:
A systematic literature review. Sustainability 2021, 13, 2121. [CrossRef]
3. Naderpour, M.; Rizeei, H.M.; Khakzad, N.; Pezzoli, A. Forest fire induced Natech risk assessment: A survey of geospatial
technologies. Reliab. Eng. Syst. Saf. 2019, 191, 106558. [CrossRef]
4. Çolak, E.; Sunar, F. Evaluation of forest fire risk in the Mediterranean Turkish forests: A case study of Menderes region, Izmir.
Int. J. Disaster Risk Reduct. 2020, 45, 101479. [CrossRef]
5. Van Hoang, T.; Chou, T.Y.; Fang, Y.M.; Nguyen, N.T.; Nguyen, Q.H.; Xuan Canh, P.; Ngo Bao Toan, D.; Nguyen, X.L.; Meadows,
M.E. Mapping forest fire risk and development of early warning system for NW Vietnam using AHP and MCA/GIS methods.
Appl. Sci. 2020, 10, 4348. [CrossRef]
Appl. Sci. 2022, 12, 6721 14 of 15

6. Janiec, P.; Gadal, S. A comparison of two machine learning classification methods for remote sensing predictive modeling of the
forest fire in the North-Eastern Siberia. Remote Sens. 2020, 12, 4157. [CrossRef]
7. Gigovic, L.; Jakovljevic, G.; Sekulović, D.; Regodić, M. GIS multi-criteria analysis for identifying and mapping forest fire hazard:
Nevesinje, Bosnia and Herzegovina. Teh. Vjesn. 2018, 25, 891–897.
8. Mohajane, M.; Costache, R.; Karimi, F.; Pham, Q.B.; Essahlaoui, A.; Nguyen, H.; Laneve, G.; Oudija, F. Application of remote
sensing and machine learning algorithms for forest fire mapping in a Mediterranean area. Ecol. Indic. 2021, 129, 107869. [CrossRef]
9. Kalantar, B.; Ueda, N.; Idrees, M.; Janizadeh, S.; Ahmadi, K.; Shabani, F. Forest fire susceptibility prediction based on machine
learning models with resampling algorithms on remote sensing data. Remote Sens. 2020, 12, 3682. [CrossRef]
10. Stula, M.; Krstinic, D.; Seric, L. Intelligent forest fire monitoring system. Inf. Syst. Front. 2012, 14, 725–739. [CrossRef]
11. Ciprián-Sánchez, J.F.; Ochoa-Ruiz, G.; Rossi, L.; Morandini, F. Assessing the impact of the loss function, architecture and image
type for Deep Learning-based wildfire segmentation. Appl. Sci. 2021, 11, 7046. [CrossRef]
12. Moayedi, H.; Mehrabi, M.; Bui, D.T.; Pradhan, B.; Foong, L.K. Fuzzy-metaheuristic ensembles for spatial assessment of forest fire
susceptibility. J. Environ. Manag. 2020, 260, 109867. [CrossRef] [PubMed]
13. Vikram, R.; Sinha, D.; De, D.; Das, A.K. EEFFL: Energy efficient data forwarding for forest fire detection using localization
technique in wireless sensor network. Wirel. Netw. 2020, 26, 5177–5205. [CrossRef]
14. Achu, A.L.; Thomas, J.; Aju, C.D.; Gopinath, G.; Kumar, S.; Reghunath, R. Machine-learning modelling of fire susceptibility in a
forest-agriculture mosaic landscape of southern India. Ecol. Inform. 2021, 64, 101348. [CrossRef]
15. Pham, B.T.; Jaafari, A.; Avand, M.; Al-Ansari, N.; Dinh Du, T.; Yen, H.P.H.; Phong, T.V.; Nguyen, D.H.; Le, H.V.; Mafi-Gholami,
D.; et al. Performance evaluation of machine learning methods for forest fire modeling and prediction. Symmetry 2020, 12, 1022.
[CrossRef]
16. Michael, Y.; Helman, D.; Glickman, O.; Gabay, D.; Brenner, S.; Lensky, I.M. Forecasting fire risk with machine learning and
dynamic information derived from satellite vegetation index time-series. Sci. Total Environ. 2021, 764, 142844. [CrossRef]
17. Naderpour, M.; Rizeei, H.M.; Ramezani, F. Forest fire risk prediction: A spatial deep neural network-based framework. Remote
Sens. 2021, 13, 2513. [CrossRef]
18. Wang, C.; Luo, T.; Zhao, L.; Tang, Y.; Zou, X. Window zooming–based localization algorithm of fruit and vegetable for harvesting
robot. IEEE Access 2019, 7, 103639–103649. [CrossRef]
19. Resco de Dios, V.; Nolan, R.H. Some challenges for forest fire risk predictions in the 21st century. Forests 2021, 12, 469. [CrossRef]
20. Bowman, D.M.; Williamson, G.J. River flows are a reliable index of forest fire risk in the temperate Tasmanian Wilderness World
Heritage Area, Australia. Fire 2021, 4, 22. [CrossRef]
21. Razavi-Termeh, S.V.; Sadeghi-Niaraki, A.; Choi, S.M. Ubiquitous GIS-based forest fire susceptibility mapping using artificial
intelligence methods. Remote Sens. 2020, 12, 1689. [CrossRef]
22. Salazar, L.G.F.; Romão, X.; Paupério, E. Review of vulnerability indicators for fire risk assessment in cultural heritage.
Int. J. Disaster Risk Reduct. 2021, 60, 102286. [CrossRef]
23. Son, B.H.; Kang, K.H.; Ryu, J.R.; Roh, S.J. Analysis of Spatial Characteristics of Old Building Districts to Evaluate Fire Risk Factors.
J. Korea Inst. Build. Constr. 2022, 22, 69–80.
24. Wang, H.; Dong, L.; Zhou, H.; Luo, L.; Lin, G.; Wu, J.; Tang, Y. YOLOv3-Litchi detection method of densely distributed litchi in
large vision scenes. Math. Probl. Eng. 2021, 2021, 8883015. [CrossRef]
25. Anderson-Bell, J.; Schillaci, C.; Lipani, A. Predicting non-residential building fire risk using geospatial information and convolu-
tional neural networks. Remote Sens. Appl. Soc. Environ. 2021, 21, 100470. [CrossRef]
26. Maffei, C.; Lindenbergh, R.; Menenti, M. Combining multi-spectral and thermal remote sensing to predict forest fire characteristics.
ISPRS J. Photogramm. Remote Sens. 2021, 181, 400–412. [CrossRef]
27. Hansen, R. The Flame Characteristics of a Tyre Fire on a Mining Vehicle. Min. Metall. Explor. 2022, 39, 317–334. [CrossRef]
28. Tomar, J.S.; Kranjčić, N.; Ðurin, B.; Kanga, S.; Singh, S.K. Forest fire hazards vulnerability and risk assessment in Sirmaur district
forest of Himachal Pradesh (India): A geospatial approach. ISPRS Int. J. Geo-Inf. 2021, 10, 447. [CrossRef]
29. Ozenen Kavlak, M.; Cabuk, S.N.; Cetin, M. Development of forest fire risk map using geographical information systems and
remote sensing capabilities: Ören case. Environ. Sci. Pollut. Res. 2021, 28, 33265–33291. [CrossRef]
30. Lin, G.; Zhu, L.; Li, J.; Zou, X.; Tang, Y. Collision-free path planning for a guava-harvesting robot based on recurrent deep
reinforcement learning. Comput. Electron. Agric. 2021, 188, 106350. [CrossRef]
31. Park, M.; Tran, D.Q.; Lee, S.; Park, S. Multilabel Image Classification with Deep Transfer Learning for Decision Support on
Wildfire Response. Remote Sens. 2021, 13, 3985. [CrossRef]
32. Šerić, L.; Pinjušić, T.; Topić, K.; Blažević, T. Lost person search area prediction based on regression and transfer learning models.
ISPRS Int. J. Geo-Inf. 2021, 10, 80. [CrossRef]
33. Lawal, M.O. Tomato detection based on modified YOLOv3 framework. Sci. Rep. 2021, 11, 1447. [CrossRef] [PubMed]
34. Tang, Y.; Zhu, M.; Chen, Z.; Wu, C.; Chen, B.; Li, C.; Li, L. Seismic performance evaluation of recycled aggregate concrete-filled
steel tubular columns with field strain detected via a novel mark-free vision method. In Structures; Elsevier: Amsterdam,
The Netherlands, 2022; Volume 37, pp. 426–441.
35. Arif, M.; Alghamdi, K.K.; Sahel, S.A.; Alosaimi, S.O.; Alsahaft, M.E.; Alharthi, M.A.; Arif, M. Role of machine learning algorithms
in forest fire management: A literature review. J. Robot. Autom. 2021, 5, 212–226.
Appl. Sci. 2022, 12, 6721 15 of 15

36. Quintero, N.; Viedma, O.; Urbieta, I.R.; Moreno, J.M. Assessing landscape fire hazard by multitemporal automatic classification
of landsat time series using the Google Earth Engine in West-Central Spain. Forests 2019, 10, 518. [CrossRef]
37. Akilan, T.; Wu, Q.J.; Zhang, H. Effect of fusing features from multiple DCNN architectures in image classification. IET Image
Processing 2018, 12, 1102–1110. [CrossRef]
38. Roy, A.M.; Bhaduri, J. Real-time growth stage detection model for high degree of occultation using DenseNet-fused YOLOv4.
Comput. Electron. Agric. 2022, 193, 106694. [CrossRef]