Page 175 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Tran Quy Nam and Phi Cong Huy 159

people before going outside, which help to give good guidance to people for wearing,
traveling in the appropriate weathers. In addition, the high accurate recognition of
weather image can help people to avoid negative effects or damages of natural
disasters.
For weather forecasting, the correct recognition of image of weather phenomena
that occurred the day before will also affect weather conditions for the next few days,
thus lead to the exact assessment of environmental quality on weather. Furthermore,
the high accuracy classifications of different weather phenomena have positive
impacts on agriculture. The accurate recognition of weather phenomena can improve
agricultural production. In transportation, the trustful assessment of weather
phenomena has much influence on moving, transportation, and vehicle assistant
driving systems. Therefore, it is becoming a significant issue to soon recognize the
weather image in our daily lives. The automatic methodology of weather image
classification through AI technology can help people to achieve sustainable
development. The accurate processing and identification of weather images taken
from drone or camera observation stations is an important method in social and
economic development and our lives in the real world.
In this paper, we implement the combined models of CNN neural network with
XGBoost algorithm to apply into a weather image phenomenon classification, which
is a deep learning algorithm using a data set containing 6,862 images with 11 types of
weather phenomena. Due to limited dataset of images and computing resources, this
study does not aim to find the better and higher rate of accuracy of the deep learning
models in classification problem. Instead, this study aim to prove that the proposed
hybrid model, namely CNN-XGBoost can provide higher performance in term of
accuracy measurement in comparisons with other models, such as simple CNN with
softmax probability, and other hybrid models, namely CNN-SVC, CNN-Decision
Tree, CNN-AdaBoost, CNN-Multi-layer Perceptron Classifier. The following
paragraphs will describe the related works, methodology, experimental results and
conclusion.

2 Related works

In fact, there have been many studies using machine learning models, deep learning
models to identify weather images. In their papers, Xiao et al. [1] implemented a
novel deep CNN that was named as MeteCNN for weather phenomena classification
to provide good results. Their MeteCNN used VGG16 as the framework to build the
proposed MeteCNN model. In fact, the MeteCNN discarded the fully connected
layers (FCL) and then it added a global average pooling layer instead of max pooling
layer before the softmax layer for classification task. Mohammad & Selvia [2] studied
to classify weather images using CNN with Transfer Learning with four CNN
architectures, namely MobileNetV2, VGG16, DenseNet201, and Xception to perform
weather image classification. They used the transfer learning aimed to speed up the
process of training models to get better and faster performance. They applied those 4
CNN architectures into the weather image which consists of six classes, namely

ISBN: 978-604-80-8083-9 CITA 2023