ABSTRACT

Social media platforms are increasingly used during disasters. In the United States, users often consider these platforms to be reliable news sources and they believe first responders will see what they publicly post. While having ways to request help during disasters might save lives, this information is difficult to find because non-relevant content on social media completely overshadows content reflective of who needs help. To resolve this issue, we develop a framework for classifying hurricane-related images that have been human-annotated. Our approach uses transfer learning and classifies each image using the VGG-16 convolutional neural network and multi-layer perceptron classifiers according to the urgency, relevance, and time period, in addition to the presence of damage and relief motifs. We find that our framework not only successfully functions as an accurate method for hurricane-related image classification but also that real-time classification of social media images using a small training set is possible.

摘要

灾难期间越来越多地使用社交媒体平台。在美国，用户通常认为这些平台是可靠的新闻来源，他们相信急救人员会看到他们公开发布的内容。虽然在灾难期间寻求帮助的方法可能会挽救生命，但这些信息很难找到，因为社交媒体上不相关的内容完全掩盖了反映谁需要帮助的内容。为了解决这个问题，我们开发了一个框架，用于对人工注释的飓风相关图像进行分类。我们的方法使用迁移学习，并使用 VGG-16 卷积神经网络和多层感知器分类器根据紧急程度、相关性和时间段以及是否存在损坏和救济图案对每个图像进行分类。