2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI)

Design and Implementation of a Cloud Particle Shape

Recognition Algorithm Based on MATLAB
2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) | 978-1-6654-8887-7/22/$31.00 ©2022 IEEE | DOI: 10.1109/CISP-BMEI56279.2022.9980263

Junyao Huo Minsong Huang

School of Information and Communication Engineering Key Laboratory of Cloud-Precipitation Physics and Severe
Beijing Information Science and Technology University Storms
Beijing, China Institute of Atmospheric Physics, CAS
[email protected] Beijing, China
[email protected]
Ruili Jiao
School of Information and Communication Engineering
Beijing Information Science and Technology University
Beijing, China
[email protected]

Abstract—Clouds are an important factor affecting climate particles to improve cloud particles recognition accuracy [7].
change and the cloud microphysical characteristics are an Xiao et al. (2019) used the ice crystal images observed from
important aspect to describe the cloud. Airborne two- the Cloud Particle Imager (CPI) to build an ice crystal dataset
dimensional stereo probe detector (2D-S) is widely used in named "China Ice Crystal Database", which consists of 10
domestic meteorological observation for the cloud microphysics shape categories and contains more than 7,000 ice crystal
research. Aiming at the situation that cloud particle shape particle images [8]. Huang et al. (2022) analyzed the
classification based on artificial naked eye is not only subjective distribution characteristics of ice particle habits in different
and different, but also time-consuming and poor in classification precipitation stages, along with temperature and diameter
effect, a cloud particle shape recognition method based on
ranges [9]. The convolution neural network (CNN) has now
convolution neural network (CNN) is proposed. A 21-layer
convolutional neural network model is built by using a
widely and successfully used in image detection area, Justin
lightweight convolution module to realize the automatic shape Ker introduces a lot of examples of CNN applied to medical
recognition of cloud particle shapes. The results show that the image analysis field [10]. Milica M. Badža present a new CNN
model accuracy rate reaches 96%, which is better than architecture for brain tumor classification of three tumor types
traditional manual shape recognition method. and in that case, the accuracy was 96.56% [11]. At present,
there are few researches on automatic recognition and
Keywords-component; cloud particles; shape recognition; classification of cloud particle images from two-dimensional
convolutional neural network; 2D-S stereo probe detector (2D-S). It is of great significance to apply
the CNN model to the recognition of cloud particle images.
I. INTRODUCTION
Clouds are an important factor affecting climate change and II. CLOUD PARTICLE DATASET PARTITION
play an important role in regulating the Earth’s radiation The data source is the cloud particle image data captured
balance [1]. Clouds can affect short-term and local weather by the airborne 2D-S provided by the Weather Modification
changes, and have an impact on global climate change, which Office of Hebei Province and Beijing Weather Modification
is one of the important objects of atmospheric scientific Office. Cloud particle image data collected from two different
research [2]. the cloud microphysical characteristics are an units at different times need to be treated differently in the
important aspect to describe the cloud, and identifying the image preprocessing process. The connected domain labeling
shape of ice crystal particles in clouds is of great significance method is used to segment the particle shapes in the original
for the study of clouds microphysical changes [3]. According image, and 9 types of datasets are manually constructed and
to the method proposed by Holroyd [4], Korolev and Sussman divided into training set and test set.
[5], Wang et al. (2014) proposed a particle shape recognition
method suitable for grayscale probes. The two-dimensional A. Image preprocessing
images of particles are divided into spherical, hexagonal, The original cloud particle image contains multiple particle
irregular, microscopic, aggregated, linear, graupel, and images captured at different times. To perform shape
branched shapes [2]. Huang et al. (2020) proposed a Holroyd recognition on cloud particles, it is first necessary to segment a
threshold algorithm suitable for shape recognition of cloud single cloud particle image from the original image. The
particles measured by cloud imaging probe (CIP) [6]. Huang original images of both data sources contain labels such as time
et al. (2020) further divided complete particles and partial information, it is necessary to convert the original image into a

This work is supported by the National Natural Science Foundation of

China (41775166), Beijing Information Science & Technology University
Project (20190091).

978-1-6654-8887-7/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on September 28,2023 at 13:17:39 UTC from IEEE Xplore. Restrictions apply.
binary image and split the original image according to the pixel III. SHAPE CLASSIFICATION BASED ON CONVOLUTIONAL
ratio to obtain narrow strips, and perform morphological NEURAL NETWORKS
operations on the cloud particles in each narrow strip to make
In recent years, deep CNNs have been widely used in the
the cloud particle shape plump and facilitate subsequent
field of image recognition. The CNN extracts image fractures,
connected domain marking. Through the connected domain
and its process is similar to the processing of images by the
identified the rectangular area to realize the extraction of a
visual nervous system of the human brain, which realize
single cloud particle image, which is input to the CNN model
automatic learning of image content. In this paper, a 21-layer
training. Figure 1 shows the processes of cloud particle image
convolutional neural network model is built to realize
preprocessing.
automatic shape recognition of 9 types of cloud particle shapes.
The GUI interface is designed for the realized program
function, which is convenient to display the main work content,
Original cloud Convert into Segmentation and also provides convenience for later transplantation and
particle image binary image narrow strips expansion.

A. CNN Model Structure

Uniform image Connected Morphological
size Domain manipulation
The convolutional neural network model in this paper has
Marker a total of 21 layers, including an input layer, five convolutional
layers, three pooling layers, four activation functions, and two
fully connected layers. The input image size is 32x32 pixels,
Save single and the convolution kernels used are 1x1, 3x3, 5x5 and 7x7.
particle images
Table 1 shows the details of the number of layers of the
convolutional neural network built in the paper.
Figure 1. The cloud particle image preprocessing procedures
TABLE I. CNN MODEL STRUCTURE

B. Cloud particle image dataset creation

According to the cloud particle image samples in the
experimental data and the studies in the cloud particle habit
identification at home and abroad, the cloud particle shape
images are manually divided into 9 types: Plate (Plt), Irregular
(Irg), Aggregate (Agt), Rosette (Ros), Sphere (Sph), Cap
column (Cap), Linear (Lin), Graupel (Grp), Dendrite (Dte).
Figure 2 is an example of the shape of a dataset sample, filter
about 180 cloud particle images for each type of shape set, and
divide training set and test set for each type of cloud particle
dataset according to the ratio of 5:1.

After passing the segmented single-particle image through

the input layer, the convolution kernel of the first convolution
layer is set to 5x5, which extract as many image features as
possible, the number of convolution kernel layers is set to 32,
and the weight of the first convolution layer is set to 0.0001,
and bias learning factor is set to 2. In order to ensure that the
edge data is not lost, the facture map will be filled with a length
of 2 on the edge after each convolution; the second convolution
layer sets two convolutions the kernel is a 3x3 lightweight
convolution module, which is convenient for extracting
Figure 2. Example of dataset shape
detailed features. The number of convolution kernel layers is
64, and the bias learning factor is 2; the third convolution layer

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on September 28,2023 at 13:17:39 UTC from IEEE Xplore. Restrictions apply.
sets the convolution kernel to 7x7 and 1x1, and weight factor B. GUI interface design
of the last convolution layer is set to 0.1, the number of Relying on MATLAB software, a simple GUI interface is
convolution kernel layers is 128, and the 1x1 convolution built according to the realized functions, which is convenient
kernel plays the role of reducing the dimension and saving the
for the display of experimental functions. According to the
operation time.
structure of the experiment, this experiment displays the main
Set a three-layer pooling layer. The first two layers select functions and key steps realized by the MATLAB program on
maximum pooling to highlight detailed features, and the last the GUI user interface. The GUI interface of this paper is set
layer selects average pooling to select the average value of the with 6 interactive buttons and 2 display windows, which
image to retain the global features. The kernel of the pooling respectively realize the select an original image, Image source
layer is set to 3 and the step is 2. The activation function adopts selection, image segmentation and size unification, input to
the batch normalization operation and the Rectified Linear the trained convolutional neural network model, display the
Unit (ReLU) function to speed up the convergence. The classification result, and close the application; two display
positive part of the ReLU function is not 0, and the negative windows respectively display the process of image processing
part is always 0, so using the ReLU activation function in the and the output of the final classification result. The above 8
deep network will not cause the gradient to be small and controls can simply and comprehensively display the main
explosion problem. Two fully connected layers are set, the first work content of this experiment, and also provide convenience
fully connected layer is set to 1048 neurons, the second fully
for later transplantation, expansion and display to others.
connected layer is set to 9 neurons, and 9 neurons represent the
final convolutional neural network output 9 classification Figure 4 shows the GUI program interface.
results.
After the convolutional neural network model is built, set
the relevant parameters. the number of training times are set to
16, and the minimum batch of each iteration are 15. The
minimum learning rate for gradient descent is 0.001. Set the
learning rate to decrease by 0.1 every 8 epochs of training.
Perform out-of-order processing before each batch of training.
By returning the confusion matrix of the classification results,
the performance of the classification results of each type of
dataset can be evaluated. The performance of the network
model is evaluated by parameters such as accuracy and time
consumption, finally saved the trained CNN model. The
operation process is shown in Figure 3, the flow chart of
building a convolutional neural network model.

Figure 4. GUI program interface

Its main operation process is:

• Click Select Image to select an original 2D-S cloud
particle image to be imported.
• Click on the data source, and select the image from the
Weather Modification Office of Hebei Province or
Beijing Weather Modification Office.
• Click to segment the image to realize single particle
recognition of the original cloud particle image and
segment the image to unify the image size.
• Click the input model, input the segmented single
particle image into the trained CNN model for
classification, and save the classification result in an
excel sheet.
• Click Show Results to display the saved classification
results on the result output windows.

Figure 3. Flow chart of building a CNN model • Click the close button to close the program.

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on September 28,2023 at 13:17:39 UTC from IEEE Xplore. Restrictions apply.
 IV. ANALYSIS OF RESULTS analysis shows that with the increase of the number of
This experiment relies on the MATLAB software convolution kernel layers, the time consumption also increases.
platform version R2018b to study the parameters affecting the At the same time, it can be found that the influence of the
classification results of the convolutional neural network number of layers of the second layer of convolution kernels is
model. The experimental data are 9 types of training sets and not significant, and the change of the third layer of convolution
test sets that have been manually divided. In this simulation layers has a greater impact on the accuracy of the results.
experiment, two experiments are set up to analysis the
influence of different initial learning rates and the number of TABLE III. THE EFFECT OF THE NUMBER OF CONVOLUTION KERNEL
convolution kernel layers on the model classification results, LAYERS ON ACCURACY

finally explore the optimal parameter settings under the Convolution

Layers Accuracy Time-consuming
experimental data. The operating system used in the kernel
experiment is Windows 10, the processor is Intel Core i5 3x3 32 91.21% 05′21″
5200U, and the computer graphics card with 32GB of memory 3x3 64 96.36% 05′42″
is AMD Radeon R7 M260.
3x3 96 92.43% 05′56″
A. The effect of initial learning rate 7x7 96 94.32% 05′23″
The initial learning rate can set the global initial learning 7x7 128 96.36% 05′42″
rate and the initial learning rate of each layer. The initial 7x7 160 89.01% 06′52″
learning rate represents the ability of the neural network to
learn the characteristics of the sample object in one batch C. Optimal model
learning. Generally, if the initial learning rate is set low, the By exploring the performance of the classification results
training results will be more reliable, but the training time will of the CNN model under different parameters, the paper
increase at the same time. On the other hand, if the initial obtained the optimal convolutional neural network model
learning rate is set high, the training process may not converge, parameters under the experimental data, and the model
and the loss function will deteriorate due to the large learning classification accuracy under the current parameters can reach
span each time it may exceed the optimal solution. This 96%, according to the confusion matrix returned by the test
experiment explores the impact of different initial learning
results of the convolutional neural network model in Figure 5,
rates on the accuracy of the convolutional neural network
the experimental results show that the classification effect of
model. The conventional initial learning rate is set to 0.001 and
then expanded to both sides with 10-fold changes each time, irregular shapes and aggregate shapes is poor. The reason is
three groups value are observed which are 0.01, 0.001, and that irregular shapes and aggregate shapes do not have obvious
0.0001. After the experiment, when the initial learning rate is regularity in shape characteristics, and it is difficult to
0.001, the accuracy rate of the network model reaches the distinguish them in many cases, there will be a situation where
maximum value of 96.36%. At this time, the learning rate an image can be considered either aggregate or irregular.
increase or decrease, which will reduce the accuracy of the
model. The accuracy rate changes under the change of the
initial learning rate are shown in Table 2. It can be observed
that the time-consuming increases as the initial learning rate
decreases.

TABLE II. THE EFFECT OF INITIAL LEARNING RATE ON ACCURACY

Initial learning rate Accuracy Time-consuming

0.01 91.52% 05′31″
0.001 96.36% 05′42″
0.0001 89.02% 06′33″

B. The effect of the number of convolution kernel layers

The number of convolutional layers determines the feature
extraction ability, but it does not mean that the more
convolutional layers, the better the training effect of the neural Figure 5. Confusion matrix of recognition result
network model. In this paper, we set up the number of 32, 64
and 96 three setting scenarios to analysis classification The accuracy rate is defined as the labels output by the
accuracy research of the second layer convolution kernel layer, classification results of the convolutional neural network
and set the third layer convolution kernel layer number to 96, model are equal to the number of labels in the test set divide
128, 160 respectively. Experimental research methods all use by the total number of labels in the test set, and the effect of
the control variable method, and only one parameter is changed model training is evaluated by the classification accuracy. The
at a time. The research results are shown in Table 3. The highest accuracy rate of the model reaches 96%. Figure 6

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on September 28,2023 at 13:17:39 UTC from IEEE Xplore. Restrictions apply.
shows the accuracy and loss changes in the training process of thesis. At the same time, I would like to think Mr. Huang
the convolutional neural network. Minsong, due to the epidemic, although we have never been
masked, you have given me great help in guiding me in terms
of professional knowledge.
REFERENCES
[1] V. Ramanathan, R. D. Cess, E. F. Harrison, et al. Cloud radiative forcing
and climate: Results from the earth radiation budget experiment.
Science, 1988, 243(4887): 57-63.
[2] Wang Lei, Li Chengcai, Zhao Zengliang, etc. Application of two-
dimensional particle shape classification technology in cloud
microphysical feature analysis. Atmospheric Science, 2014, 38(02):201-
212.
[3] H. Letu, H. Ishimoto, J. Riedi, et al. Investigation of ice particle habits to
be used for ice cloud remote sensing for the GCOM-C satellite mission.
Atmospheric Chemistry and Physics,2015,15(21),31,665–31,703.
Figure 6. The accuracy and loss changes in the training process [4] E. W. Holroyd. Some techniques and uses of 2D-C habit classification
software for snow particles. Journal of Atmospheric and Oceanic
Technology, 1987, 4:498-511.
V. SUMMARY [5] A. Korolev, B. Sussman. A technique for habit classification of cloud
particles. Journal of Atmospheric and Oceanic Technology, 2000,
The classification of cloud particle shape is of great 17:1048-1057.
significance to meteorological observation and atmospheric [6] Huang Minsong, Lei Hengchi, Wang Xiujuan, Ice particle habit
cloud microphysics research. Previously, the classification of classification method with improved thresholds and its application.
cloud particle shape mostly relied on artificial naked-eye Climate and Environment Research,2020,25(04):419-428.
classification, which is subjectivity and differently. In this [7] Huang Minsong，Lei Hengchi. An improved Holroyd cloud particle
habit identification method and its application. Acta Meteorologica
paper, the original cloud particle image is preprocessed to Sinica,2020,78(02): 289-300
realize the segmentation of single particle image, 9 types of [8] Xiao Haixia, Zhang Feng. Classification of ICE Crystal Habits Observed
cloud particle shape datasets are created manually, divided From Airborne Cloud Particle Imager by Deep Transfer Learning. Earth
into training set and test set, and build a CNN model. The and Space Science, 2019, 6:36-46.
highest accuracy rate of the model reaches 96%, which [9] Huang Minsong, Lei Hengchi. Ice particle habit distribution in stratiform
clouds at different precipitation stages. Transactions of Atmospheric
realizes the automatic classification and recognition of cloud Sciences, 2020,45(02):247-256.
particle shapes with high accuracy. [10] J. Ker, L. Wang, J. Rao and T. Lim, "Deep Learning Applications in
Medical Image Analysis," in IEEE Access, vol. 6, pp. 9375-9389, 2018,
ACKNOWLEDGMENT doi: 10.1109/ACCESS.2017.2788044.
I would like to think my tutor, Mrs. Jiao Ruili. You are [11] Badža MM, Barjaktarović MČ. Classification of Brain Tumors from
not only my class teacher when I first entered the university, MRI Images Using a Convolutional Neural Network. Applied Sciences.
2020; 10(6):1999. https://doi.org/10.3390/app10061999.
but also a teacher who imparts my subject knowledge, thank
you for your guidance from the beginning to the end of my

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on September 28,2023 at 13:17:39 UTC from IEEE Xplore. Restrictions apply.