Abstract:
Geo-crowdsourcing is an important emerging technique for geospatial data collection, where contributors identify the position and attributes of features to be included in geographic information systems. The geo-crowdsourcing movement shares many similarities to other well-known citizen science projects, including the use of training to improve the accuracy of the information being collected. This thesis describes an ongoing research project where participants are being trained to contribute data to a crowdsourcing system used to identify accessibility problems associated with transient obstacles and navigation hazards. The training includes information about determining and establishing location of obstacles and defining attributes for obstacles. In order to assess the effectiveness of this training program, student participants were divided into two treatment groups, with one group receiving the standard training program and the other receiving no training other than a short introduction to the project. The students in the trained treatment group were shown images of obstacles along with a characterization of each obstacle, followed by a series of sample images of obstacles that they could practice assessing and characterizing. Student participants in both treatment groups were then led one-by-one on a pre-determined path around campus and asked to report at least 4 obstacles encountered, using a crowdsourcing web-map. Concurrently, several trained project moderators independently characterized all obstacles along the path to provide ground truth for this study, and compiled a master list of obstacles. The moderators then evaluated the reports submitted by study participants and provided a comprehensive quality assessment of each report. Analysis of this quality assessment data using t-tests and Mann-Whitney tests finds that with the small sample size used in this study (n=23) there was no significant difference between the trained participants and the untrained participants. Both groups achieved quality scores in the 60-65% range, with the trained group having a higher (but insignificantly higher) average total quality score, obstacle type quality score, and moderator-assessed final quality score. The untrained group had a slightly higher (but insignificantly higher) average score for accuracy of report location. These findings and their implications are discussed in this thesis, along with future suggested research directions.
