Abstract: Wildfire events are increasingly threatening our lands, cities, and lives. To contrast this phenomenon and to limit its damages, governments around the globe are trying to find proper counter-measures, identifying prevention and monitoring as two key factors to reduce wildfires impact worldwide. In this work, we propose two deep convolutional neural networks to automatically detect and delineate burned areas from satellite acquisitions, assessing their performances at scale using validated maps of burned areas of historical wildfires. We demonstrate that the proposed networks substantially improve the burned area delineation accuracy over conventional methods.

Abstract: Multimedia content in social media platforms provides significant information during disaster events. The types of information shared include reports of injured or deceased people, infrastructure damage, and missing or found people, among others. Although many studies have shown the usefulness of both text and image content for disaster response purposes, the research has been mostly focused on analyzing only the text modality in the past. In this paper, we propose to use both text and image modalities of social media data to learn a joint representation using state-of-the-art deep learning techniques. Specifically, we utilize convolutional neural networks to define a multimodal deep learning architecture with a modality-agnostic shared representation. Extensive experiments on real-world disaster datasets show that the proposed multimodal architecture yields better performance than models trained using a single modality (e.g., either text or image).

Abstract: Innovative technologies powered by Computer Vision algorithms can aid first responders, increasing their situ ational awareness. However, adverse conditions, such as smoke, can reduce the efficacy of such algorithms by degrading the input images. This paper presents a pipeline of image de-smoking, object detection, and augmented reality display that aims to enhance situational awareness in smoky conditions. A novel smoke-reducing deep learning algorithm is applied as a preprocessing step, before state-of-the-art object detection. The detected objects and persons are highlighted in the user’s augmented reality display. The proposed method is shown to increase detection accuracy and confidence. Testing in realistic environments provides an initial evaluation of the method, both in terms of image processing and of usefulness to first responders.

Abstract: We report on the development of a classifier to identify Twitter users contributing first-hand information during a disaster. Identifying such users helps social media monitoring teams identify critical information that might otherwise slip through the cracks. A parallel study (St. Denis et al., 2020) demonstrates that Twitter user filtering creates an information-rich stream of content, but the best way to approach this task is unexplored. A user's profile contains many different “modalities” of data, including numbers, text, and images. To integrate these different data types, we constructed a multimodal neural network that combines the loss function of all modalities, and we compared the results to many individual unimodal models and a decision-level fusion approach. Analysis of the results suggests that unimodal models acting on Twitter users' recent tweets are sufficient for accurate classification. We demonstrate promising classification of Twitter users for crisis response with methods that are (1) easy to implement and (2) quick to both optimize and infer.

Abstract: Flooding is a natural hazard that causes a lot of deaths every year and the number of flood events is increasing worldwide because of climate change effects. Detecting and monitoring floods is of paramount importance in order to reduce their impacts both in terms of affected people and economic losses. Automated image analysis techniques capable to extract the amount of water from a picture can be used to create novel services aimed to detect floods from fixed surveillance cameras, drones, crowdsourced in-field observations, as well as to extract meaningful data from social media streams. In this work we compare the accuracy and the prediction performances of recent Deep Learning algorithms for the pixel-wise water segmentation task. Moreover, we release a new dataset that enhances well-know benchmark datasets used for multi-class segmentation with specific flood-related images taken from drones, in-field observations and social media.

Abstract: In emergency management, identifying disaster information accurately and promptly out of numerous documents like news articles, announcements, and reports is important for decision makers to accomplish their mission efficiently. This paper studies the application of the question answering system which can automatically locate answers in the documents by natural language processing to improve the efficiency and accuracy of disaster knowledge extraction. Firstly, an improved question answering model was constructed based on the advantages of the existing neural network models. Secondly, the English question answering dataset pertinent to disasters and the Chinese question answering dataset were constructed. Finally, the improved neural network model was trained on the datasets and tested by calculating the F1 and EM scores which indicated that a higher question answering accuracy was achieved. The improved system has a deeper understanding of the semantic information and can be used to construct the disaster knowledge graph.

Abstract: Nowadays, micro-blogging sites such as Twitter have become powerful tools for communicating with others in
various situations. Especially in disaster events, these sites can be the best platforms for seeking or providing social
support, of which informational support and emotional support are the most important types. Sympathy, a sub-type
of emotional support, is an expression of one?s compassion or sorrow for a difficult situation that another person
is facing. Providing sympathy to people affected by a disaster can help change people?s emotional states from
negative to positive emotions, and hence, help them feel better. Moreover, detecting sympathy contents in Twitter
can potentially be used for finding candidate donors since the emotion ?sympathy? is closely related to people who
may be willing to donate. Thus, in this paper, as a starting point, we focus on detecting sympathy-related tweets.
We address this task using Convolutional Neural Networks (CNNs) with refined word embeddings. Specifically, we
propose a refined word embedding technique in terms of various pre-trained word vector models and show great
performance of CNNs that use these refined embeddings in the sympathy tweet classification task. We also report
experimental results showing that the CNNs with the refined word embeddings outperform not only traditional
machine learning techniques, such as Naïve Bayes, Support Vector Machines and AdaBoost with conventional
feature sets as bags of words, but also Long Short-Term Memory Networks.

Abstract: The user-base of social media platforms, like Twitter, has grown dramatically around the world over the last decade. As people post everything they experience on social media, large volumes of valuable multimedia content are being recorded online, which can be analysed to help for a range of tasks. Here we specifically focus on crisis response. The majority of prior works in this space focus on using machine learning to categorize single-modality content (e.g. text of the posts, or images shared), with few works jointly utilizing multiple modalities. Hence, in this paper, we examine to what extent integrating multiple modalities is important for crisis content categorization. In particular, we design a pipeline for multi-modal learning that fuses textual and visual inputs, leverages both, and then classifies that content based on the specified task. Through evaluation using the CrisisMMD dataset, we demonstrate that effective automatic labelling for this task is possible, with an average of 88.31% F1 performance across two significant tasks (relevance and humanitarian category classification). while also analysing cases that unimodal models and multi-modal models success and fail.