Abstract:
The continued rise of global temperatures is causing a major climate crisis and this is leading to devastating and deadly
natural disasters. People use social media platforms to capture and share real-time incidents in the form of images, videos and text.
However, sharing too much information at once makes it harder for first responders to determine where exactly individuals are in
need and whether they require immediate assistance. In the past, machine learning techniques were used to automatically identify
and infer disaster response from images, as manually identifying disaster types is currently challenging. Therefore, in this work,
deep learning models are used to investigate how well they can classify the images according to their disaster type by learning
the features extracted from the input images on their own. Seven categories of disaster were considered in this study. These were:
cyclones, earthquakes, floods, droughts, landslides, wildfires and urban fires. Two existing datasets namely the Comprehensive Disaster
Dataset (CDD) and the Natural Disaster Dataset (NDD) were customised into a single dataset which we named as the Customised
Disaster Dataset (CDD). The Customised Disaster Dataset comprises of a total of ten classes, three of which are images which
are not related to disaster. These three classes are: regular images of buildings and streets, wild forest and sea. Three pre-trained
deep learning models such MobileNetV2, VGG16 and InceptionV3 were used to train the datasets to allow for further comparison
with existing studies. Along with that, a customised neural network model was created and trained on the datasets. Different
scenarios were devised to assess the top performing models. The InceptionV3 had the best classification accuracy of 96.86% when
the trainable layers were set to false. We also obtained an accuracy of 96.86% with the MobileNetV2 model but this time the
trainable layers were set to true. In this study, we have demonstrated the effectiveness of CNN models as a tool for the automatic
classification of disaster-related images. Most studies have used only two categories (disaster-related and non-disaster related images)
or are restricted to only one type of disaster (water-related, land-related, etc.) while in our studies we have used seven categories
of disasters. However, the accuracy of the models may be less if the images are taken at night or when the weather conditions are very bad.