Building Systems to Enable Spatial Data Science at Scale
Date and Time of the talk: April 13 2021, 9:30 AM EDT
Information of the Speaker
Mohamed Sarwat, Arizona State University
Mohamed Sarwat is an assistant professor of computer science at Arizona State University. Dr. Sarwat is a recipient of the 2019 National Science Foundation CAREER award and was named an Early Career Distinguished Lecturer by the IEEE Mobile Data Management community in June 2019. His general research interest lies in developing robust and scalable data systems. The outcome of his research has been recognized by two best research paper awards in the IEEE International Conference on Mobile Data Management (MDM 2015) and the International Symposium on Spatial and Temporal Databases (SSTD 2011), a best of conference citation in the IEEE International Conference on Data Engineering (ICDE 2012) as well as a best vision paper award (3rd place) in SSTD 2017. Besides impact through scientific publications, Mohamed is also the co-architect of several software artifacts, which include GeoSpark (a scalable system for processing big geospatial data) that is being used by major tech companies such as Uber, Facebook and MoBike. Dr. Sarwat spent the summers of 2011 and 2012 at NEC laboratories and Microsoft Research Redmond, respectively. He is an associate editor for the GeoInformatica journal and has served as an organizer / reviewer / program committee member for major data management and spatial computing venues.
Abstract
In the last 20 years, geospatial data (extracted from GPS traces, geo-tagged social media, weather maps, natural disasters, satellites imagery, and epidemic situations) has become wildly ubiquitous. That led to the rise of spatial data science as a field, which usually refers to extracting meaningful information from geospatial data. However, the lack of scalability and interactivity in state-of-the-art spatial data systems makes it extremely difficult for a data scientist to store, retrieve, explore, analyze, visualize and learn from large-scale geospatial data.
In this talk, I will first shed light on Apache Sedona (formerly named GeoSpark), an open source data system that builds upon the core engine of Apache Spark to efficiently process large-scale geospatial data in a cluster computing environment. Internally, Sedona represents geospatial data as a SpatialRDD, which is tailored for Apache Spark in-memory data processing paradigm. Sedona allows users to write their spatial data processing tasks in Spatial SQL, compiles the input SQL into a set of optimized SpatialRDD operations, and finally executes such operations in the cluster. Since a data scientist many times applies the analysis to only a subset of the entire database, I will give an overview of Hippo a lightweight indexing scheme that outperforms de-facto database indexes such B-tree and R-tree in terms of storage and maintenance overhead, while still executing range queries (to retrieve a subset of the data) at a comparative performance to such in dexes. Furthermore, a data scientist may sometimes allow for a slight trade-off between the accuracy and scalability of the analysis. To allow for such trade-off, I will present a sampling middleware system called Tabula, which sits between the data system and the data science tool (such as Tableau) to make the inherently iterative human-in-the-loop analysis process more seamless and interactive. Finally, I will switch gears a bit to present my future plan related to spatial data systems support for the Internet of Things (IoT) and the data science for cities initiative.