{"id":21793,"date":"2018-09-11T11:18:08","date_gmt":"2018-09-11T15:18:08","guid":{"rendered":"https:\/\/www.crim.ca\/blogue\/ce-que-jai-lu-cette-semaine-les-donnees-spatio-temporelles\/"},"modified":"2023-05-25T12:19:25","modified_gmt":"2023-05-25T16:19:25","slug":"what-i-read-this-week-spatiotemporal-data","status":"publish","type":"blogue","link":"https:\/\/www.crim.ca\/en\/blogue\/what-i-read-this-week-spatiotemporal-data\/","title":{"rendered":"What I read this week: Spatiotemporal data"},"content":{"rendered":"

This week, I took an interest in spatiotemporal data<\/strong>. There are several reasons why I have been looking into this subject. First, the City of Montreal makes available open data related to different spheres of activity: economy and business, education, health, society and culture. The sector that is of particular interest to me for Montreal is transportation<\/strong>.<\/p>\n

CRIM<\/a> has worked with spatiotemporal data to predict the response time of firefighters at the Service de s\u00e9curit\u00e9 incendie de Montr\u00e9al<\/a>. The article on this can be read on Medium<\/a>. Given the experience gained during this project, it seemed logical to continue studying the subject. This new knowledge could eventually apply to resolving other mobility-related<\/strong> problems, such as travel in Montreal<\/strong>, thanks to the data collected by the MTL Trajet<\/a> application or the cameras and detectors scattered around the Island of Montreal.<\/p>\n

Spatiotemporal data also offers interesting potential from a local perspective for flood prevention in Quebec<\/strong>. In 2018, public security officials at the Minist\u00e8re de la S\u00e9curit\u00e9 publique<\/em> unveiled a civil security plan of action in cases of flooding (the Quebec government release, in French, is found here<\/a>) as a follow-up to the major spring floods of the previous year. Over five years, an envelope of more than $30M was provided to update the mapping of flood-prone areas on Quebec territory. In climatology<\/strong>, spatiotemporal data have become such an important tool that many data science projects are being developed to organize, analyze and interpret them.<\/p>\n

Spatiotemporal data<\/h2>\n

Spatiotemporal data are collected in several fields: in climate science, to predict the onset of extreme weather events; in Earth science, to detect anomalous areas within marine environments; or in epidemiology, to study the spread of diseases. Other fields, such as neuroscience, environmental science, social media and traffic dynamics, also use such information in various ways.<\/p>\n

\n

But what is special about spatiotemporal data? They are characterized by spatial<\/strong> (distance, direction, position) and temporal<\/strong> (number of occurrences, changes in time, duration) attributes. In other words, they are data, or measurements, that change in time and space.<\/p>\n<\/blockquote>\n

For example, for predicting meteorological events, the spatial representation is defined by a grid of the land area under study, and various parameters and atmospheric measurements characterize it. The temporal representation is then defined by the evolution of these parameters over time, which could allow, for instance, to design a prediction model of the evolution of hurricanes.<\/p>\n

There is ample literature on the subject: the diversity of models, approaches and applications is practically countless. To begin with, I have limited myself to four articles of interest to give me an idea of how to manage and use spatiotemporal data.<\/p>\n

The first paper presents an approach using a 3D convolutional neural network to learn spatiotemporal features. The second paper presents data mining applications using association rules in the context of hurricane analysis. The third deals with a learning model of hierarchical structures to predict the onset of extreme weather events. The fourth defines a spatiotemporal data mining model analyzing anomalous association structures in a marine environment. The following sections will summarize the essential and relevant ideas and concepts from each paper.<\/p>\n

<\/div>\n
\n

Learning Spatiotemporal Features with 3D Convolutional Networks<\/h2>\n
\n

Tran, D., Bourdev, L., Fergus, R., Torresani, L., & Paluri, M. (2015, December). Learning spatiotemporal features with 3d convolutional networks. In Computer Vision (ICCV), 2015 IEEE International Conference on (pp. 4489\u20134497). IEEE.<\/p>\n<\/blockquote>\n

Objective<\/h2>\n

Design of an approach for learning spatiotemporal features using a deep 3D convolutional neural network<\/strong> (3D ConvNets) trained on a large-scale supervised video dataset. The goal is to recognize different types of actions (101 actions) and objects (42 types) and to classify pairs of similar actions (432 actions) or individual scenes (14 scenes)<\/p>\n

Summary<\/h2>\n
\n
\n
\"\"<\/div>\n<\/div>\n<\/figure>\n

The idea behind 3D convolution operations is to preserve the temporal information of an input signal. Indeed, 2D convolutions applied on one or multiple images generate an image, thus a two-dimensional output. Only 3D convolutions preserve temporal information and generate a volumetric output. The same phenomenon is observed for pooling stages in 2D and 3D, respectively. Thus, using 3D convolutions to encapsulate the information related to objects, scenes and actions.<\/p>\n

Method<\/h2>\n

Architecture<\/strong><\/p>\n

The basic architecture of the convolution kernel is a fixed receptive field of dimension {d x 3 x 3} where only the time dimension is modified for experimentation purposes. This limitation is due to the considerable training time that would result from numerous experiments. The notation for the dimensions of the video clips and kernels is as follows:<\/p>\n

Video clip\u00a0(c\u00a0x\u00a0l\u00a0x\u00a0h\u00a0x\u00a0w)<\/em><\/p>\n