Oversampling Social Media-Sourced Image Datasets for Better Deep Learning Classification of Natural Disaster Damage Levels

People in areas affected by natural disasters and use social media websites such as Facebook, Twitter (also known as "X") a nd Instagram tend to post images of damage to their surroundings. These social media sites have become vital sources of immediate and highly available data for providing situational awareness and organisation for natural disaster response. A few previous attempts at classifying the level of natural disaster damage in these images using image processing techniques had noted the challenge in producing robust classification models due to the effect of overfitting caused by a lack of observations and data imbalance in annotated datasets. This article shows an attempt to improve a training strategy within the data level for deep learning models such as VGG16, ResNetV2 and EffecientNetV2, used to estimate the level of disaster damage in images by training them with data generated using image data augmentation with data balancing, oversampling up to eight times and combining the oversampled image data collections. The F-1 score achieved for classifying damage on earthquake images and images from the Hurricane Matthew data collection by training EfficientNetV2 on a generated dataset made with a combination of oversampled data surpassed previous benchmark results. These results show that using data balancing and oversampling on the dataset prior to training deep learning models on these datasets result in increased robustness.