Abstract:
"Understanding clouds and their patterns is crucial for meteorological research, but current groundbased cloud classification algorithms often fail to provide complete categorization. This study
presents a multi-stage classification technique specifically designed to capture the subtle nuances
of ground-based cloud formations, resulting in a more precise and detailed classification. Clouds
are divided into high-level and low-level classes that include patterns like clear skies, patterned
clouds, thick white or dark clouds, and veil clouds and then classified into 10 low-level
classification Ci, Cc, Cs, Ac, Ns, As, Cu, St, Sc, Cb. The proposed method addresses these
limitations by enhancing meteorological predictions and deepening our understanding of cloud
dynamics.
The research used a ground-based cloud categorization system that relied on Convolutional Neural
Networks (CNNs). This system was revolutionary in detecting cloud patterns. The CNN
architecture comprised convolutional, pooling, and fully connected layers, allowing automatic
training of features from raw cloud image data. The research employed transfer learning with pretrained CNN models such MobileNetv2 to extract relevant cloud features, which further improved
classification accuracy and efficiency. This reduced the need for manual feature engineering and
improved classification accuracy. The system was also capable of detecting complex cloud
patterns, textures, and structures by preprocessing the data for uniformity and augmenting the data
for diversity.
The ground-based cloud classification system has proven to be effective for both high-level and
low-level cloud classification according to the final results of the study. The system outperformed
known models in high-level cloud classification, with an accuracy rate of 98%, surpassing both
EfficientNet and ResNet50. In low-level cloud categorization, the system continued to perform
well with an accuracy rate of 57%, surpassing EfficientNet."
