<![CDATA[Cloud image classification plays an important role in supporting atmospheric analysis and the development of artificial intelligence-based weather prediction systems. However, visual similarities among cloud types and variations in lighting conditions remain major challenges in automatic classification tasks. This study aims to compare the performance of three Convolutional Neural Network (CNN) architectures, namely DenseNet121, MobileNetV2, and ResNet50, for cloud image classification using a transfer learning approach. The dataset used in this study was the TJNU Ground-based Cloud Dataset, consisting of seven cloud classes with a total of 19.000 images; after removing the Mixed Clouds category with the total of 18.047 images; after removing. The research stages included pre-processing, data augmentation using the Color Jitter technique, model training through feature extraction and fine-tuning, and evaluation using Accuracy, Precision, Recall, and F1-Score metrics. Experimental results showed that ResNet50 achieved the best performance with an accuracy of 90,58% and an F1-score of 91,38%, followed by DenseNet121 with 89,53% accuracy and MobileNetV2 with 85,10% accuracy. In addition to obtaining the highest classification performance, ResNet50 also demonstrated good computational efficiency with the fastest training time of 20.9 minutes. These findings indicate that architectures based on residual learning are more effective in capturing the visual characteristics of cloud images compared to the other architectures evaluated in this study.]]>
Copyrights © 2026
