A Forest Fire Detection System Based on Ensemble Learning

Abstract

Due to the various shapes, textures, and colors of fires, forest fire detection is a challenging task. The traditional image processing method relies heavily on manmade features, which is not universally applicable to all forest scenarios. In order to solve this problem, the deep learning technology is applied to learn and extract features of forest fires adaptively. However, the limited learning and perception ability of individual learners is not sufficient to make them perform well in complex tasks. Furthermore, learners tend to focus too much on local information, namely ground truth, but ignore global information, which may lead to false positives. In this paper, a novel ensemble learning method is proposed to detect forest fires in different scenarios. Firstly, two individual learners Yolov5 and EfficientDet are integrated to accomplish fire detection process. Secondly, another individual learner EfficientNet is responsible for learning global information to avoid false positives. Finally, detection results are made based on the decisions of three learners. Experiments on our dataset show that the proposed method improves detection performance by 2.5% to 10.9%, and decreases false positives by 51.3%, without any extra latency.

Full text

Forests 2021, 12, 217. https://doi.org/10.3390/f12020217 www.mdpi.com/journal/forests
Article
A Forest Fire Detection System Based on Ensemble Learning
Renjie Xu
1
, Haifeng Lin
1
, Kangjie Lu
1
, Lin Cao
2
and Yunfei Liu
1,
*
1
College of Information Science and Technology, Nanjing Forestry University, Nanjing 210037, China;
xurenjie@njfu.edu.cn (R.X.); haifeng.lin@njfu.edu.cn (H.L.); lukangjie@njfu.edu.cn (K.L.)
2
Co-Innovation Center for Sustainable Forestry in Southern China, Nanjing Forestry University,
Nanjing 210037, China; lincao@njfu.edu.cn
* Correspondence: lyf@njfu.com.cn; Tel.: +86-139-1389-5117
Abstract: Due to the various shapes, textures, and colors of fires, forest fire detection is a challenging
task. The traditional image processing method relies heavily on manmade features, which is not
universally applicable to all forest scenarios. In order to solve this problem, the deep learning tech-
nology is applied to learn and extract features of forest fires adaptively. However, the limited learn-
ing and perception ability of individual learners is not sufficient to make them perform well in com-
plex tasks. Furthermore, learners tend to focus too much on local information, namely ground truth,
but ignore global information, which may lead to false positives. In this paper, a novel ensemble
learning method is proposed to detect forest fires in different scenarios. Firstly, two individual learn-
ers Yolov5 and EfficientDet are integrated to accomplish fire detection process. Secondly, another
individual learner EfficientNet is responsible for learning global information to avoid false posi-
tives. Finally, detection results are made based on the decisions of three learners. Experiments on
our dataset show that the proposed method improves detection performance by 2.5% to 10.9%, and
decreases false positives by 51.3%, without any extra latency.
Keywords: forest fire detection; deep learning; ensemble learning; Yolov5; EfficientDet; EfficientNet
1. Introduction
With the change of the earth’s climate, forest fires occur frequently all over the world,
which not only cause serious economic losses and destroy the ecological environment, but
also pose a great threat to the safety of human life.
Forest fires usually spread quickly and are difficult to control in a short time. There-
fore, it is imperative to detect the early forest fire before it spreads out, but traditional
detection methods have obvious drawbacks in detecting it in open forest areas. Sensors-
based [1–3] detection systems have good performance in indoor space, but it is difficult to
install them outdoors, considering high coverage cost [4,5]. In addition, they cannot pro-
vide important visual information which can help firefighters promptly grasp the situa-
tion of the fire scene. Infrared or ultraviolet detectors [6,7] are easy to be interfered by the
environment, and considering their short detection distance, they are not suitable for large
open areas. Satellite remote sensing [8] is good at detecting large-scale forest fires, but it
cannot detect early regional fire.
Impressed by the rising computer vision technology, researchers start to seek an ef-
ficient and effective fire detection model based on image processing. Chen et al. [9] pro-
posed an RGB (red, green, blue) model based chromatic and disorder measurement for
extracting fire-pixels in the video. The color information is responsible for extracting fire-
pixels, and dynamic information is used to verify if it is a real fire. Töreyin et al. [10] used
1D temporal wavelet transform to detect flame flicker, and applied 2D spatial wavelet
transform to identify fire moving regions. This method, which integrated color and tem-
poral variation information, reduced false alarms in real-world scenes. Çelik et al. [11]
studied diverse video sequences and images, and proposed a fuzzy color model using
Citation: Xu, R.; Lin, H.; Lu, K.; Cao,
L.; Liu, Y. A Forest Fire Detection
System Based on Ensemble
Learning. Forests 2021, 12, 217.
https://doi.org/10.3390/f12020217
Academic Editor:
Stelian Alexandru Borz
Received: 4 January 2021
Accepted: 12 February 2021
Published: 13 February 2021
Publisher’s Note: MDPI stays neu-
tral with regard to jurisdictional
claims in published maps and institu-
tional affiliations.
Copyright: © 2021 by the authors. Li-
censee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and con-
ditions of the Creative Commons At-
tribution (CC BY) license (http://crea-
tivecommons.org/licenses/by/4.0/).

Forests 2021, 12, 217 2 of 17
statistical analysis. Combined with motion analysis, the model achieves a good discrimi-
nation between fire and fire-like objects. Teng et al. [12] analyzed fire characteristics and
proposed a real-time fire detection method based on hidden Markov models (HMMs),
which extracted candidate fire-pixels using moving pixel detection, fire-color inspection,
and pixel clustering. Chino et al. [13] found that most algorithms were designed for video,
which had obvious limitations. To solve this problem, a novel fire detection method
named BowFire was proposed. The method combined color features with superpixel tex-
ture discrimination to detect fire in still images. In conclusion, most traditional fire detec-
tion methods based on image processing focused on creating artificial features like color,
motion, and texture to detect fires.
However, powerful deep learners begin to replace human intelligence. They are bet-
ter at learning features than humans, and the features they extract contain much deeper
semantic information than manmade ones. Recently, deep learning has outperformed tra-
ditional manmade features in many fields, and have been widely used in fire detection.
Zhang et al. [14] created a forest fire benchmark, and used Faster R-CNN (region-based
convolutional neural network) [15], Yolo (you only look once) [16–19], and SSD (single
shot multibox detector) [20] to detect fire. They found that SSD was better regarding effi-
ciency, detection accuracy, and early fire detection ability. Moreover, they proposed an
improved tiny-Yolo by adjusting the network architecture. Kim et al. [21] employed faster
R-CNN to detect fire and non-fire regions based on their spatial features. In addition, long
short-term memory (LSTM) is used to verify the reliability of fire alarm. Lee et al. [22]
proposed a video-based fire detection model, which used faster R-CNN to generate a fire
candidate region for each frame. Then, structural similarity (SSIM) and mean square error
(MSE) were calculated to determine similarity between adjacent frames. Final fire regions
were determined based on spatial and temporal features. Pan et al. [23] proposed a cam-
era-based wildfire detection system via transfer learning, in which block-based analysis
strategy was used to improve fire detection accuracy. Redundant filters, which had low
energy impulse response, were removed to ensure the model’s efficiency on edge devices.
Wu et al. [24] applied principal component analysis (PCA) to process forest fire images,
and then fed them into the training network. The combination of two models was proved
to enhance location results. In conclusion, faced with fire detection task, most researchers
tend to only assign individual learners to perform object detection tasks, which is consid-
ered unreliable, since it may lead to false negatives.
In this paper, a novel method based on ensemble learning for forest fire detection is
proposed. First, forest fire detection is a complicated and difficult task, making it highly
impractical for individual learners to detect fires in diverse scenarios. Every individual
learner has its own expertise, and can extract different features from the image, so inte-
grating different individual learners can significantly improve the robustness of the model
and enhance detection performance. Therefore, two individual object detectors Yolov5
[25] and EfficientDet [26] are integrated to detect the fire in parallel. These two learners
work synergistically in detecting different types of forest fires, thereby improving the de-
tection accuracy. Second, the object detectors only care about what fire is like, so they do
not take the whole image into consideration. In this case, fire-like objects will absolutely
affect the detection results. To solve this problem, the EfficientNet image classifier [27] is
incorporated into our model, whose role is to enable the model to take full advantage of
the global information. Final detection results will be made through the decision strategy
according to results of these three learners, which will efficiently increase detection accu-
racy and decrease the false positives.
2. Materials and Methods
2.1. Datasets
To ensure our learners can handle different kinds of forest fires (ground fire, trunk
fire, and canopy fire), we collected images from multiple public fire datasets: BowFire [28],

Forests 2021, 12, 217 3 of 17
FD-dataset [29], ForestryImages [30], VisiFire [31], etc. After manual filtration, we created
a single integrated forest fire dataset containing 10,581 images, with 2976 forest fire images
and 7605 non-fire images. Representative samples of our dataset are shown in Figures 1–
3.
(a) (b)
(c) (d)
Figure 1. Representative forest fire images in the fire section of our dataset, including (a) ground fire 1, (b) ground fire 2,
(c) trunk fire, and (d) canopy fire.
(a) (b)
(c) (d)
Figure 2. Representative normal forest images in the non-fire section of our dataset, including (a) normal forest scene 1,
(b) normal forest scene 2, (c) normal forest scene 3, and (d) normal forest scene 4. (a–d) illustrate normal forest scenes
without fire objects.

Forests 2021, 12, 217 4 of 17
(a) (b)
(c) (d)
Figure 3. Representative images in the non-fire section of our dataset, including (a) wild scene with sun 1, (b) wild scene
with sun 2, (c) wild scene with sun 3, and (d) wild scene with sun 4. (a–d) illustrate normal wild scenes containing fire-
like object (e.g., sun).
2.2. Yolov5
Yolo is a state-of-the-art, real-time object detector, and Yolov5 is based on Yolov1-
Yolov4. Continuous improvements have made it achieve top performances on two official
object detection datasets: Pascal VOC (visual object classes) [32] and Microsoft COCO
(common objects in context) [33].
The network architecture of Yolov5 is shown in Figure 4. There are three reasons why
we choose Yolov5 as our first learner. Firstly, Yolov5 incorporated cross stage partial net-
work (CSPNet) [34] into Darknet, creating CSPDarknet as its backbone. CSPNet solves the
problems of repeated gradient information in large-scale backbones, and integrates the
gradient changes into the feature map, thereby decreasing the parameters and FLOPS
(floating-point operations per second) of model, which not only ensures the inference
speed and accuracy, but also reduces the model size. In forest fire detection task, detection
speed and accuracy is imperative, and compact model size also determines its inference
efficiency on resource-poor edge devices. Secondly, the Yolov5 applied path aggregation
network (PANet) [35] as its neck to boost information flow. PANet adopts a new feature
pyramid network (FPN) structure with enhanced bottom-up path, which improves the
propagation of low-level features. At the same time, adaptive feature pooling, which links
feature grid and all feature levels, is used to make useful information in each feature level
propagate directly to following subnetwork. PANet improves the utilization of accurate
localization signals in lower layers, which can obviously enhance the location accuracy of
the object. Thirdly, the head of Yolov5, namely the Yolo layer, generates 3 different sizes
(18 × 18, 36 × 36, 72 × 72) of feature maps to achieve multi-scale [18] prediction, ena-
bling the model to handle small, medium, and big objects. A forest fire usually develops
from small-scale fire (ground fire) to medium-scale fire (trunk fire), then to big-scale fire
(canopy fire). Multi-scale detection ensures that the model can follow size changes in the
process of fire evolution.

Forests 2021, 12, 217 5 of 17
Figure 4. The network architecture of Yolov5. It consists of three parts: (1) Backbone: CSPDarknet, (2) Neck: PANet, and
(3) Head: Yolo Layer. The data are first input to CSPDarknet for feature extraction, and then fed to PANet for feature
fusion. Finally, Yolo Layer outputs detection results (class, score, location, size).
2.3. EfficientDet
EfficientDet is a new family of object detectors developed by Google, and it consist-
ently achieves better efficiency than prior art across a wide spectrum of resource con-
straints. Similar to Yolov5, EfficientDet has also achieved remarkable performances in
Pascal VOC and Microsoft COCO tasks, and is widely used in real-world applications.
The network architecture of EfficientDet is shown in Figure 5. There are three reasons
why we choose EfficientDet as our second learner. Firstly, EfficientDet employed state-of-
the-art network EfficientNet [27] as its backbone, making that the model has sufficient
ability to learn the complex feature of diverse forest fires. Secondly, it applied an im-
proved PANet, named bi-directional feature pyramid network (Bi-FPN) as its neck, to al-
low easy and fast multi-scale feature fusion. Bi-FPN introduces learnable weights, ena-
bling network to learn the importance of different input features, and repeatedly applies
top-down and bottom-up multi-scale feature fusion. Compared with Yolov5′s neck
PANet, Bi-FPN has better performances with less parameters and FLOPS. Meanwhile,
different feature fusion strategy brings different semantic information, thereby bringing
different detection results. Thirdly, similar to EfficientNet, it integrates a compound scal-
ing method that uniformly scales the resolution, depth, and width for all backbone, fea-
ture network, and box/class prediction networks at the same time, which ensures the max-
imum accuracy and efficiency under the limited computing resources. With more availa-
ble resources, accuracy will be consistently improved. Our second learner, EfficientDet,
with different backbone, neck, and head, can learn different information that Yolov5 can-
not.

Forests 2021, 12, 217 6 of 17
Figure 5. The network architecture of EfficientDet. It consists of three parts: (1) Backbone: EfficientNet, (2) Neck: Bi-FPN,
(3) Head. Similar to Yolov5, the data are first input to EfficientNet for feature extraction, and then fed to Bi-FPN for feature
fusion. Finally, head outputs detection results (class, score, location, size).
2.4. EfficientNet
EfficientNet is a new efficient network proposed by Google. It applied a novel model
scaling strategy, namely compound scaling method, to balance network depth, network
width, and image resolution for better accuracy at a fixed resource budget. With this, Ef-
ficientNet outperformed other hot networks like ResNet [36], DenseNet [37], ResNeXt [38]
with the highest Top-1 accuracy in ImageNet image classification task.
The network architecture of EfficientNet is shown in Figure 6. The reason why we
choose EfficientNet as our third learner is that it achieves a superior trade-off between
accuracy and efficiency. In our model, the third learner plays the most important role. It
is responsible for learning the whole image to guide the detection, meaning that its deci-
sions directly determine the final results. Meanwhile, it must be highly efficient, otherwise
it will slow down the speed of the entire model.
Figure 6. The network architecture of EfficientNet. It can output a feature map with deep semantic information after the
input data flows through the multi-layer network.
2.5. Our Model
In real-world forest fire detection task, we need to handle different types of forest
fires like ground fire, trunk fire, canopy fire. These fires, influenced by the environment,
are diverse in shape, texture, or even color, bringing great difficulty for individual learner
to extract effective features. By careful observations, we find that Yolov5 is better at learn-
ing long-area fires (Figure 7), but it sometimes misses objects (Figure 8). Meanwhile, even
though EfficientDet is not sensitive to long-area fires (Figure 7), it is more careful than

Forests 2021, 12, 217 7 of 17
Yolov5, meaning that EfficientDet can make a complementary detection (Figure 8). There-
fore, we consider that integrating these two efficient learners with different specialties to
make detection together can improve detection accuracy.
(a) (b)
(c) (d)
Figure 7. Yolov5 is better at detecting long-area fires than EfficientDet. (a) True positive of Yolov5; (b) true positive of
Yolov5; (c) false negative of EfficientDet; (d) false negative of EfficientDet. (a,b) illustrate that Yolov5 detect long-area fires
successfully, while (c,d) show that EfficientDet fails to detect them.
(a) (b)
(c) (d)
Figure 8. EfficientDet is a more careful object detector than Yolov5, meaning that it seldom losses potential objects easily.
(a) Yolov5 fails to cover all fire areas; (b) Yolov5 misses two fire objects; (c) EfficientDet covers all fire areas; (d) EfficientDet
detects four fire objects.

Forests 2021, 12, 217 8 of 17
Another issue is that the ability of the object detector is limited. It only learns the fire
region, which is just a local pattern of the whole image, but ignores the other information
like background. As a result, the object detector may treat fire-like objects (e.g., sun) as
fires (Figure 9), thereby making false alarms. Therefore, a good leader EfficientNet that
has a full understanding of the whole image is needed to guide the detection process.
(a) (b)
(c) (d)
Figure 9. Object detectors Yolov5 and EfficientDet are easy to be deceived by fire-like objects (e.g., sun). (a) False positive
of Yolov5 (confidence score: 0.63); (b) false positive of Yolov5 (confidence score: 0.59); (c) false positive of EfficientDet
(confidence score: 0.84); (d) false positive of EfficientDet (confidence score: 0.71).
To address the above two issues and make sure our model is robust to diverse sce-
narios, three deep learners are integrated to make decisions together (Figure 10). The first
and second learners Yolov5 and EfficientDet act as object detectors, to detect fire locations
in images by generating candidate boxes, respectively. Then, the non-maximum suppres-
sion algorithm [39] (Algorithm 1) is employed to eliminate redundant boxes, preserving
boxes with top confidence. The third learner EfficientNet acts as a binary classifier, re-
sponsible for learning the whole image to determine whether the image contains fire ob-
jects. Finally, the object detection results, and image classification results are sent into a
decision strategy module, in which if the image is considered to contain fire objects, re-
taining object detection results, otherwise ignoring them.
In addition, integrating multiple learners will not affect the overall efficiency of
model, because the three learners are structurally independent, and the whole model is
executed by multi processes, meaning that each learner has a separate process responsible
for it.

Forests 2021, 12, 217 9 of 17
Figure 10. Structure of the proposed model in this paper. Three deep learners are ensembled in parallel. Two object detec-
tors Yolov5 and EfficientDet are integrated to perform object detection task, and the classifier EfficientNet is in charge of
discriminating whether the image contains fire objects. Final detection results are made based on the decisions of three
learners.
Algorithm 1. Non-Maximum Suppression (NMS)
INPUT: B={b,…,b
}, S={s,…,s
}, N
B is the list of initial detection boxes
S contains corresponding detection scores
N is the NMS threshold
Begin:
D←{ }
while B≠empt
y
do
m ← argmax S
M ← b
D ← D ∪ M; B ← B − M
for b in B do
if iouM, b≥N
 then
B←B−b
; S ← S − s
end
end
end
Return D, S
End
2.6. Model Evaluation
We evaluate models using Microsoft COCO criteria (Table 1), which is widely used
in object detection tasks. However, fire is a special object, which is diverse in shape, tex-
ture, and color. Bounding box generated by object detectors may slightly differ from
ground truth (Figure 11), thereby influencing the calculation of average precision, but de-
tectors do identify the fire areas successfully. Therefore, to evaluate models more compre-
hensively, we introduce two additional evaluation metrics, namely frame accuracy (FA)
and false positive rate (FPR). For one image, if the detector misses any fire object, we call
it is a frame false (FF), otherwise frame true (FT). If the detector treats any fire-like object
as fire, we call it is a false positive (FP), otherwise true positive (TP). Note that FA is cal-
culated on the test set containing 476 forest images, and FPR is calculated on our challeng-
ing non-fire dataset containing 641 images with fire-like objects (e.g., sun). The FA and
FPR can be calculated as Equation (1) and Equation (2), respectively:
FA = FT
FT + FF × 100, (1)
FPR = FP
FP + TP × 100. (2)

Forests 2021, 12, 217 10 of 17
(a) (b)
(c) (d)
Figure 11. Bounding boxes generated by (a) Yolov5, (b) EfficientDet, and (c) our model (3 learners) are different from (d)
ground truth, but still has good detection performance.
Table 1. Microsoft COCO criteria—commonly used in object detection task for evaluating the
model precision and recall across multiple scales.
Average Precision (AP)
AP
. AP at IoU = 0.5
AP Across Scales:
AP
 AP
. for small objects: area < 32
AP
 AP
. for medium objects: 32<area<96

AP
 AP
. for big objects: area > 96
Average Recall (AR)
AR. AR at IoU = 0.5
AR Across Scales:
AR AR. for small objects: area < 32
AR AR. for medium objects: 32<area<
96
AR AR. for big objects: area > 96
3. Results
3.1. Training
We applied different strategies to train our three learners: Yolov5, EfficientDet, and
EfficientNet. Object detectors, namely Yolov5 and EfficientDet, are trained with 2381 for-
est fire images, and tested with 476 forest fire images. The image classifier, namely Effi-
cientNet, is trained with 2381 forest fire images and 5804 non-fire images, and tested with
476 forest fire images and 1160 non-fire images. Note that non-fire images contain normal
images, and images with fire-like objects (e.g., sun). Each model is built up by Pytorch [40]

Forests 2021, 12, 217 11 of 17
and trained on NVIDIA GTX 2080TI. The details of our training strategy are shown in
Table 2.
Table 2. Detailed training strategies of models.
Model Train Test Optimizer LR Batch Size Epoch
Yolov5 2381 476 SGD [41,42]
1×10
 8 300
EfficientDet 2381 476 AdamW [43]
1×10
 4 300
EfficientNet 8185 1636 SGD 1×10
 8 300
LR: learning rate, SGD: stochastic gradient descent, AdamW: Adam with decoupled weight decay.
3.2. Comparison
We compare our model with typical one-stage object detectors. As is shown in Table
3, even though Yolov5 and EfficientDet are the most powerful detectors in this task, the
high false positive rate and missing detections cannot be ignored. By integrating them (2
learners), all evaluation metrics are significantly improved, but the false positive rate is
increased to 51.6%, since the false positives come from both Yolov5 and EfficientDet. Un-
der the guide of our third learner EfficientNet, the false positive rate is reduced to 0.3%.
What is also worth mentioning is that, after introducing the third learner, some metrics
are slightly decreased. It is because that EfficientNet wrongly treats some fire images as
non-fire ones, and then ignores the object detection results, but we consider it is worth-
while to sacrifice a tiny decrease in average precision and recall for substantial improve-
ment in the false positive rate. To sum up, our model (3 learners) is superior in AP
., AP
,
AP
, AP
, AR., AR, AR, AR, FPR, and FA compared with other typical object detec-
tors. Comprehensive improvements make the model have better performance in detecting
different types of forest fires: small-scale fires, medium-scale fires, big-scale fires, ground
fires, trunk fires, canopy fires, and fires at night (Figures 12 and 13). Faced with fire-like
objects (e.g., sun), our model will not be interfered. (Figure 14).
Table 3. Experiments on our dataset—evaluating models using Microsoft COCO criteria, FPR, FA, and latency.
Model 𝐀𝐏𝟎.𝟓 𝐀𝐏𝐒 𝐀𝐏𝐌 𝐀𝐏𝐋 𝐀𝐑𝟎.𝟓 𝐀𝐑𝐒 𝐀𝐑𝐌 𝐀𝐑𝐋 FPR FA Latency(ms)
SSD 66.8 37.8 42.4 78.6 70.1 39.1 45.7 82.7 45.6 92.6 88.8
Yolov3 66.4 26.0 44.6 78.1 71.1 26.1 52.5 82.5 22.9 88.0 15.6
Yolov3-SPP 68.3 56.3 49.9 76.7 73.9 60.9 56.6 81.9 30.7 93.3 15.6
Yolov4 69.6 53.7 48.9 78.4 75.5 60.9 57.5 83.9 61.9 94.1 20.5
Yolov5 70.5 51.9 53.7 79.2 75.6 56.5 61.2 83.0 22.6 94.7 28.0
EfficientDet 75.7 63.7 58.5 83.0 79.2 65.2 63.9 86.5 41.8 95.5 65.6
Ours (2 learners) 79.7 72.2 65.6 85.5 84.1 76.1 73.1 89.3 51.6 99.4 66.8
Ours (3 learners) 79.0 72.2 64.9 84.7 83.8 76.1 72.6 88.9 0.3 98.9 66.8
Note that AP
., AP
, AP
, AP
, AR., AR, AR, AR, FPR, and FA are all percentages. The best figure of each metric
are highlighted in bold.

Forests 2021, 12, 217 12 of 17
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 12. Our ensemble model (3 learners) has better performance on ground fires, trunk fires, and canopy fires. (a) Four
ground fires detected by Yolov5; (b) Yolov5 fails to detect the trunk fire; (c) three canopy fires detected by Yolov5; (d) four
ground fires detected by EfficientDet; (e) the trunk fire detected by EfficientDet; (f) two canopy fires detected by Effi-
cientDet; (g) six ground fires detected by our model; (h) the trunk fire detected by our model; (i) three canopy fires detected
by our model.
(a) (b) (c)

Forests 2021, 12, 217 13 of 17
(d) (e) (f)
(g) (h) (i)
Figure 13. Our improved model has better performance on small-scale, medium-scale, and big-scale fires at night. (a)
Medium-scale and big-scale fires detected by Yolov5; (b) medium-scale and big scale fires detected by Yolov5; (c) small-
scale, medium-scale and big-scale fires detected by Yolov5; (d) medium-scale and big-scale fires detected by EfficientDet;
(e) medium-scale and big scale fires detected by EfficientDet; (f) small-scale, medium-scale, and big-scale fires detected
by EfficientDet; (g) medium-scale and big-scale fires detected by our model; (h) medium-scale and big scale fires detected
by our model; (i) small-scale, medium-scale, and big-scale fires detected by our model.
(a) (b)
(c) (d)

Forests 2021, 12, 217 14 of 17
(e) (f)
Figure 14. Under the guide of EfficientNet, our ensemble model has a good discriminability between fire and fire-like
objects (e.g., sun). (a) True negative of Yolov5; (b) false positive of Yolov5 (confidence score: 0.59); (c) false positive of
EfficientDet (confidence score 0.71); (d) true negative of EfficientDet; (e) true negative of our model; (f) true negative of
our model.
4. Discussion
Compared with other common objects that have fixed form, forest fire is a dynamic
object [44]. In the real-world scenario, a forest fire usually starts from small-scale fire, de-
velops to medium-scale fire, and then to big-scale fire [45]. In terms of types, it starts from
ground fire, then spreads to the trunk, and finally to the canopy [46]. The various shapes,
sizes, textures, and colors of forest fires make the fire evolution a complex process, and
bring great difficulty in fire detection.
Therefore, it is highly imperative for detectors to be sensitive to different types of
fires. Through careful experimental comparisons, we find that no single detector that can
handle all kinds of fires. They have respective advantages and disadvantages. Yolov5 is
excellent at detecting long-area fires (Figure 7), but it frequently misses objects (Figure 8).
EfficientDet is a more careful detector, compared to Yolov5; even though it has a bad per-
formance on long-area fires (Figure 7), it can detect fires that Yolov5 cannot (Figure 8),
meaning that it is a good partner for Yolov5. Our model, which efficiently integrates de-
cisions of these two powerful learners, boost detection performance by 2.5–10.9%, in terms
of AP
., AP
, AP
, AP
, AR. , AR, AR, AR. The significant improvements of aver-
age precision and average recall for small, medium, and big objects make the model more
sensitive to the size changes of fires, thereby enhancing detection performance on differ-
ent types of forest fires: ground fire, trunk fire, canopy fire, and fires at night in the fire
evolution (Figures 12 and 13).
Another problem is that the false positive rate of the improved model (2 learners)
becomes higher: 22.6% to 51.6% since the model also integrates wrong detection results
from both learners. To address this issue, we use 8185 images containing 2381 forest fire
images and 5804 non-fire images (containing fire-like images and normal forest images)
to train our third learner EfficientNet. Sufficient training sets enabled EfficientNet to show
a good discriminability between fire objects and fire-like objects, with 99.6% accuracy on
476 fire images, and 99.7% accuracy on 676 fire-like images. With the help of the leader
learner EfficientNet, wrong detection results are eliminated, and the false positive rate is
significantly decreased to 0.3% (Figure 14). Noticeably, the join of EfficientNet reduces
AP
. , AP
, AP
, AR. , AR, AR by roughly 1%, which is because that EfficientNet
wrongly ignores 2 fire images containing medium-scale and big-scale fire objects.
In terms of latency, the Yolo family is superior compared to EfficientDet and SSD.
Excellent inference speed makes Yolo family widely used in real-world applications, but
experimental results show that they are not able to have a satisfactory performance on
forest fire detection tasks. The latency of EfficientDet is 65.6 ms, which is over twice that
of Yolov5 (28.0 ms), but EfficientDet outperforms Yolov5 by over 5% regarding detection
performance. We ensemble these three learners Yolov5 (28.0 ms), EfficientDet (65.6 ms),
EfficientNet (31.3 ms) in parallel to make sure that our model can achieve the best perfor-
mance without any extra latency. The final latency of our model (3 learners) is 66.8 ms,

Forests 2021, 12, 217 15 of 17
which shows that an excellent trade-off between detection performance and efficiency has
been achieved, and the model is applicable for real-time detection task.
For further improvement, we plan to study the labeling strategy for forest fires, since
the quality of training data directly determines the detection performance. Another inter-
esting extension is to investigate the network architecture of backbones, and modify them
to make sure that they are specially designed for forest fire detection task. Additionally,
we will work on developing a forest fire tracking system, which can classify different
types of forest fires: ground fire, trunk fire and canopy fire, to track the evolution and
spread of forest fires.
5. Conclusions
The successful application of convolutional neural networks significantly improves
the performance of object detection. However, forest fire is a dynamic object with no fixed
form, which the individual object detector cannot handle. In addition, object detectors are
easy to be deceived by fire-like objects and generate false positives due to their limited
visual field. To address these problems, a novel ensemble learning method for real-time
forest fire detection is proposed in this paper. Two powerful object detectors (Yolov5 and
EfficientDet) with different expertise are integrated to make the whole model more robust
to diverse forest fire scenarios. Then, a leader (EfficientNet) is introduced to guide the
detection process to reduce false positives. Experimental results show that, compared
with other popular object detectors, our model achieves a superior trade-off among aver-
age precision, average recall, false positive rate, frame accuracy, and latency. The signifi-
cant improvements make it possible for the model to perform well in real-world forestry
applications.
Author Contributions: R.X. devised the programs and drafted the initial manuscript. H.L. and K.L.
helped with data collection, data analysis, and figures and tables. L.C. contributed to fund acquisi-
tion and writing embellishment. Y.L. designed the project and revised the manuscript. All authors
have read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Key R&D Program of China (grant number
2017YFD0600904) and the Priority Academic Program Development of Jiangsu Higher Education
Institutions (PAPD).
Data Availability Statement: Publicly available datasets were analyzed in this study. The data can
be found here: BowFire [28], FD-dataset [29], ForestryImages [30], VisiFire [31].
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Zhang, J.; Li, W.; Yin, Z.; Liu, S.; Guo, X. Forest fire detection system based on wireless sensor network. In Proceedings of the
4th IEEE Conference on Industrial Electronics and Applications (ICIEA 2009), Xi’an, China, 25–27 May 2009; pp. 520–523.
2. Yu, L.; Wang, N.; Meng, X. Real-time forest fire detection with wireless sensor networks. In Proceedings of the International
Conference on Wireless Communications, Networking and Mobile Computing (WiCOM 2005), Wuhan, China, 26 September
2005; pp. 1214–1217.
3. Chen, S.J.; Hovde, D.C.; Peterson, K.A.; Marshall, A.W. Fire detection using smoke and gas sensors. Fire Saf. J. 2007, 42, 507–515,
doi:10.1016/j.firesaf.2007.01.006.
4. Zhang, F.; Zhao, P.; Xu, S.; Wu, Y.; Yang, X.; Zhang, Y. Integrating multiple factors to optimize watchtower deployment for
wildfire detection. Sci. Total Environ. 2020, 737, 139561, doi:10.1016/j.scitotenv.2020.139561.
5. Zhang, F.; Zhao, P.; Thiyagalingam, J.; Kirubarajan, T. Terrain-influenced incremental watchtower expansion for wildfire
detection. Sci. Total Environ. 2018, 654, 164–176, doi:10.1016/j.scitotenv.2018.11.038.
6. Lee, B.; Kwon, O.; Jung, C.; Park, S. The development of UV/IR combination flame detector. J. KIIS 2001, 16, 1–8.
7. Kang, D.; Kim, E.; Moon, P.; Sin, W.; Kang, M. Design and analysis of flame signal detection with the combination of UV/IR
sensors. J. Korean Soc. Int. Inf. 2013, 14, 45–51, doi:10.7472/jksii.2013.14.2.45.
8. Fernandes, A.M.; Utkin, A.B.; Lavrov, A.V.; Vilar, R.M. Development of neural network committee machines for automatic
forest fire detection using lidar. Pattern Recognit. 2004, 37, 2039–2047, doi:10.1016/j.patcog.2004.04.002.

Forests 2021, 12, 217 16 of 17
9. Chen, T.H.; Wu, P.H.; Chiou, Y.C. An early fire-detection method based on image processing. In Proceedings of the IEEE
International Conference on Image Processing (ICIP 2004), Singapore, 24–27 October 2004; pp. 1707–1710.
10. Töreyin, B.U.; Dedeoğlu, Y.; Güdükbay, U.; Cetin, A.E. Computer vision based method for real-time fire and flame detection.
Pattern Recognit. Lett. 2006, 27, 49–58, doi:10.1016/j.patrec.2005.06.015.
11. Çelik, T.; Özkaramanlı, H.; Demirel, H. Fire and smoke detection without sensors: Image processing based approach. In
Proceedings of the IEEE 15th European Signal Processing Conference (EUSIPCO 2007), Poznan, Poland, 3–7 September 2007;
pp. 1794-1798.
12. Teng, Z.; Kim, J.H.; Kang, D.J. Fire detection based on hidden Markov models. Int. J. Control Autom. Syst. 2010, 8, 822–830,
doi:10.1007/s12555-010-0414-2.
13. Chino, D.Y.; Avalhais, L.P.; Rodrigues, J.F.; Traina, A.J. Bowfire: Detection of fire in still images by integrating pixel color and
texture analysis. In Proceedings of the 28th SIBGRAPI Conference on Graphics, Patterns and Images, Salvador, Brazil, 26–29
August 2015; pp. 95–102.
14. Wu, S.; Zhang, L. Using popular object detection methods for real time forest fire detection. In Proceedings of the 11th
International Symposium on Computational Intelligence and Design (ISCID 2018), Hangzhou, China, 8–9 December 2018; pp.
280–284.
15. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Trans.
Pattern Anal. Mach. Intell. 2016, 39, 1137–1149, doi:10.1109/TPAMI.2016.2577031.
16. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), Las Vegas, NV, USA, 26 June–1 July 2016; pp. 779–
788.
17. Redmon, J.; Farhadi, A. YOLO9000: Better, faster, stronger. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR 2017), Honolulu, Hawaii, 21–26 July 2017; pp. 7263–7271.
18. Redmon, J.; Farhadi, A. Yolov3: An incremental improvement. arXiv 2018, arXiv: 1804.02767.
19. Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv 2020, arXiv:
2004.10934.
20. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings
of the European Conference on Computer Vsion (ECCV 2016), Amsterdam, the Netherlands, 8–16 October 2016; pp. 21–37.
21. Kim, B.; Lee, J. A video-based fire detection using deep learning models. Appl. Sci. 2019, 9, 2862, doi:10.3390/app9142862.
22. Lee, Y.; Shim, J. False Positive Decremented Research for Fire and Smoke Detection in Surveillance Camera using Spatial and
Temporal Features Based on Deep Learning. Electronics 2019, 8, 1167, doi:10.3390/electronics8101167.
23. Pan, H.; Badawi, D.; Cetin, A.E. Computationally Efficient Wildfire Detection Method Using a Deep Convolutional Network
Pruned via Fourier Analysis. Sensors 2020, 20, 2891, doi:10.3390/s20102891.
24. Wu, S.; Guo, C.; Yang, J. Using PCAand one-stage detectors for real-time forest fire detection. J. Eng. 2020, 2020, 383–387,
doi:10.1049/joe.2019.1145.
25. Ultralytics-Yolov5. Availabe online: https://github.com/ultralytics/yolov5 (accessed on 1 Januray 2021).
26. Tan, M.; Pang, R.; Le, Q.V. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR 2020), Washington, DC, USA, 14–19 June 2020; pp. 10781–10790.
27. Tan, M.; Le, Q. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In Proceedings of the International
Conference on Machine Learning (ICML 2019), Long Beach, CA, USA, 9–15 June 2019; pp. 6105–6114.
28. BoWFire Dataset. Availabe online: https://bitbucket.org/gbdi/bowfire-dataset/downloads/ (accessed on 1 Januray 2021).
29. FD-Dataset. Availabe online: http://www.nnmtl.cn/EFDNet/ (accessed on 1 Januray 2021).
30. ForestryImages. Availabe online: https://www.forestryimages.org/browse/subthumb.cfm?sub=740 (accessed on 1 Januray
2021).
31. VisiFire. Availabe online: http://signal.ee.bilkent.edu.tr/VisiFire/ (accessed on 1 Januray 2021).
32. Everingham, M.; Eslami, S.A.; Van Gool, L.; Williams, C.K.; Winn, J.; Zisserman, A. The pascal visual object classes challenge:
A retrospective. Int. J. Comput. Vis. 2015, 111, 98–136, doi:10.1007/s11263-014-0733-5.
33. Lin, T.Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Dollár, P.; Zitnick, C.L. Microsoft coco: Common objects in
context. In Proceedings of the 13th European Conference on Computer Cision (ECCV 2014), Zurich, Switzerland, 6–12
September 2014; pp. 740–755.
34. Wang, C.Y.; Mark Liao, H.Y.; Wu, Y.H.; Chen, P.Y.; Hsieh, J.W.; Yeh, I.H. CSPNet: A new backbone that can enhance learning
capability of cnn. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2020),
Washington, DC, USA, 14–19 June 2020; pp. 390–391.
35. Wang, K.; Liew, J.H.; Zou, Y.; Zhou, D.; Feng, J. Panet: Few-shot image semantic segmentation with prototype alignment. In
Proceedings of the IEEE International Conference on Computer Vision (ICCV 2019), Seoul, Korea, 20–26 October 2019; pp. 9197–
9206.
36. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR 2016), Las Vegas, NV, USA, 26 June–1 July 2016; pp. 770–778.

Forests 2021, 12, 217 17 of 17
37. Huang, G.; Liu, Z.; Van Der Maaten, L.; Weinberger, K.Q. Densely connected convolutional networks. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2017), Honolulu, Hawaii, 21–26 July 2017; pp. 4700–4708.
38. Xie, S.; Girshick, R.; Dollár, P.; Tu, Z.; He, K. Aggregated residual transformations for deep neural networks. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), Las Vegas, NV, USA, 26 June–1 July 2016; pp.
1492-–1500.
39. Neubeck, A.; Van Gool, L. Efficient non-maximum suppression. In Proceedings of the 18th International Conference on Pattern
Recognition (ICPR 2006), Hong Kong, China, 20–24 August 2006; pp. 850–855.
40. Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L. Pytorch: An
imperative style, high-performance deep learning library. In Proceedings of the Neural Information Processing Systems (NIPS
2019), Vancouver, BC, Canada, 8–14 December 2019; pp. 8026–8037.
41. Bottou, L. Large-scale machine learning with stochastic gradient descent. In Proceedings of the 19th International Conference
on Computational Statistics (COMPSTAT 2010), Paris, France, 22–27 August 2010; pp. 177–186.
42. Zinkevich, M.; Weimer, M.; Li, L.; Smola, A.J. Parallelized stochastic gradient descent. In Proceedings of the Neural Information
Processing Systems (NIPS 2010), Vancouver, BC, Canada, 6–11 December 2010; pp. 2595–2603.
43. Loshchilov, I.; Hutter, F. Decoupled weight decay regularization. arXiv 2017, arXiv: 1711.05101.
44. Merino, L.; Caballero, F.; Martínez-de-Dios, J.R.; Maza, I.; Ollero, A. An unmanned aircraft system for automatic forest fire
monitoring and measurement. J. Intell. Robot. Syst. 2012, 65, 533–548, doi:10.1007/s10846-011-9560-x.
45. Serón, F.J.; Gutiérrez, D.; Magallón, J.; Ferragut, L.; Asensio, M.I. The evolution of a wildland forest fire front. Vis. Comput. 2005,
21, 152–169, doi:10.1007/s00371-004-0278-7.
46. Pimont, F.; Dupuy, J.L.; Linn, R.R.; Dupont, S. Impacts of tree canopy structure on wind flows and fire propagation simulated
with FIRETEC. Ann. For. Sci. 2011, 68, 523–530, doi:10.1007/s13595-011-0061-7.