Drone imaging and geographic information system (GIS) programs allowing spatial data analysis have been found to be useful in ecology, yet little is known about the methodology and success associated with highly detailed projects. Our investigation focuses on QGIS techniques and their ability to classify flowers based on colors. Part of a larger experiment, we analyzed drone images of reduced input wheat, biologically based wheat, and early successional community plots of the MCSE at the Kellogg Biological Station during summer 2019, and attempted to geospatially identify individual Rudbeckia hirtia (black eyed susan) flowers that we had deployed ourselves, along with natural flowers. Using collected GPS coordinates, drone images were compiled into orthomosaic images before refinement within MetaShapePro. After being uploaded to QGIS, flowers were classified using manipulated image bands / vertices and known regions of interest (ROIs). The semi-automatic classification plugin within QGIS or a script of random forest algorithm then completed the classification. Finally, spectral data of ROIs were analyzed and accuracy assessments comparing classifications with actual image contents were run before completing necessary replications.
Results/Conclusions
From our classification attempts, it was found that increasing the numbers of classes and ROIs within classes increased classification quality and that vertices NDVI, EVI, NIR/Red, NIR/Red Edge, SAVI, Green/Red and Red Edge/Red improved classification accuracy as well. Using the classifications, our study shows that the QGIS semi-automatic classification plugin can accurately classify flower resources in the reduced-input and biologically-based wheat. However, it required more sophisticated algorithms (e.g. random forest) to classify flower resources in the early-successional landscape that was characterized by higher plant diversity. Better understanding techniques used for image classifications with biological applications will provide ecologists with additional insight concerning the effects of spatial data factors such as size, relative distance, and degree of centrality on processes concerning pollination.