Exploring the use of crowd generated geospatial content in improving the quality of ecological feature mapping

Generalisation and uncertainty in ecological data are hugely problematic for environmental monitoring and modelling analyses and as such, land use change calculations, global climate change monitoring and food security analyses suffer. Low resolution, incomplete and often inconsistent land cover data sets signify that the digital representation of the Earth’s biophysical surface is very poor. User generated content is gaining increasing traction within the scientific community; researchers are increasingly looking to Volunteered Geographic Information as a solution to fulfil the inadequacies of authoritative information as it can often be obtained more frequently and with less expense. Throughout this thesis, the use of VGI for land cover classification and Phase 1 Habitat classification is trialled with the view that crowd generated content could make appropriate training data for remote sensing classifications. In the field of ecology, VGI has mostly been used in the context of validating existing data and a great deal of research is required into how ecological data can be collected most effectively and the extent to which it can be filtered to achieve usable datasets. This thesis examines the utility of both open web data (passively submitted to sources such as Geograph) and data obtained via directed crowdsourcing as ancillary data for remote sensing classification. The open web land cover data sources analysed (OpenStreetMap and Geograph) have some very detailed and potentially minable information, though require substantial edits and filtering prior to use. Directed crowdsourcing approaches are shown to have a stronger match with their authoritative counterpart and several factors (characteristic profiling, classification time and user confidence in classification) are found to be useful ways of filtering the crowd contributions to achieve the best quality data set. The investigations show how datasets can be synthesised that, whilst not reaching the accuracies of authoritative data, could be seen as an asset to ecological monitoring and inventory. Extending the directed crowdsourcing method to use data collected iteratively in situ via a mobile application forms the focus of future research recommendations. Active learning is also suggested as a complementary method which could reduce the number of training samples required by focusing sampling efforts on where existing classifications are most uncertain and where the contributor’s knowledge profile is most appropriate