Geographically weighted evidence combination approaches for combining discordant and inconsistent volunteered geographical information

There is much interest in being able to combine crowdsourced data. One of the critical issues in information sciences is how to combine data or information that are discordant or inconsistent in some way. Many previous approaches have taken a majority rules approach under the assumption that most people are correct most of the time. This paper analyses crowdsourced land cover data generated by the Geo-Wiki initiative in order to infer the land cover present at locations on a 50 km grid. It compares four evidence combination approaches (Dempster-Shafer, Bayes, Fuzzy Sets and Possibility) applied under a geographically weighted kernel with the geographically weighted average approach applied in many current Geo-Wiki analyses. A geographically weighted approach uses a moving kernel under which local analyses are undertaken. The contribution (or salience) of each data point to the analysis is weighted by its distance to the kernel centre, reflecting Tobler’s 1st law of geography. A series of analyses were undertaken using different kernel sizes (or bandwidths). Each of the geographically weighted evidence combination methods generated spatially distributed measures of belief in hypotheses associated with the presence of individual land cover classes at each location on the grid. These were compared with GlobCover, a global land cover product. The results from the geographically weighted average approach in general had higher correspondence with the reference data and this increased with bandwidth. However, for some classes other evidence combination approaches had higher correspondences possibly because of greater ambiguity over class conceptualisations and / or lower densities of crowdsourced data. The outputs also allowed the beliefs in each class to be mapped. The differences in the soft and the crisp maps are clearly associated with the logics of each evidence combination approach and of course the different questions that they ask of the data. The results show that discordant data can be combined (rather than being removed from analysis) and that data integrated in this way can be parameterised by different measures of belief uncertainty. The discussion highlights a number of critical areas for future research