<![CDATA[Natural disasters are events caused by natural phenomena that cause massive damage and pose a threat to human safety. Based on EM-DAT data (2000–2025), there have been more than 10,000 global disasters, resulting in millions of casualties and trillions of US dollars in losses. Notably, 2024 saw US$320 billion in losses due to extreme weather. This condition emphasizes the importance of an accurate disaster classification system for mitigation and rapid response. This study aims to develop a natural disaster image classification model using the Convolutional Neural Network (CNN) method with a Transfer Learning approach using the MobileNetV2 architecture, which is known to be efficient and lightweight. This study employs the SEMMA (Sample, Explore, Modify, Model, Assess) methodology, beginning with sampling, which involves collecting image data from various open sources, such as Kaggle and previous literature. The data is then processed through the selection, cleaning, and normalization stages. Data exploration is conducted to understand class distribution and detect data imbalance. To overcome this problem, the Synthetic Minority Over-sampling Technique (SMOTE) and image data augmentation (such as rotation, flipping, and contrast adjustment) were used to enrich the training data variation. The pre-trained MobileNetV2 model was then retrained using the modified data, with adjustments to hyperparameters to achieve optimal performance. The evaluation was conducted using various metrics, namely accuracy, precision, recall, F1-score, confusion matrix, and AUC-ROC curve. The results demonstrate that the combination of Transfer Learning, data augmentation, and SMOTE can enhance model performance, achieving an accuracy of up to 99% and an AUC-ROC above 0.99 on public test data. Additionally, testing on 26 private test images yielded an accuracy of 92.31%, with 24 of the 26 images classified correctly. These findings confirm that the combination of MobileNetV2, augmentation, and SMOTE effectively improves multi-class classification performance on natural disaster images. Furthermore, the use of a relatively lightweight model makes this system more efficient to implement on devices with limited resources, thereby supporting disaster mitigation and rapid response efforts.]]>
