A Meta-Learning Framework for Characterizing and Accessing Training Data for GLOBE Observer Program Land Cover Protocols

Citizen science is increasingly recognized as an important method and source of data for data collection, meeting community needs, and connecting citizens with researchers. An ever-increasing number of geo-tagged field photographs are being collected and shared with the public, creating an exciting opportunity for human and environmental studies at multiple scales. However, labeling and featuring those labels is time-consuming and might be biased for scientific re-use. For GLOBE Observer (GO) -Land Cover, despite users taking single-directional photos, the program has highly standardized protocols using six directional photos. We believe that the intersection of standardized protocol-driven citizen science programs with machine learning models provides important opportunities for GLOBE to lead new types of land surveys and answer key environmental questions that cannot be answered now. In this study, based on GLOBE Observer Land Cover unique sampling protocols, we developed a data fusion and GeoAI framework to take full advantage of GO multi-directional photos to help increase the land cover classification accuracy from the reference data perspective. And preliminary results based on a multilabel variation of GO Land Cover demonstrate the potential of the proposed method, which outperforms the current state of the art by more than 5 percent in terms of accuracy when compared to the use of single-directional photos. This work will begin to answer the question of how citizen science derived data can be used for research purposes and how those photos should be used in order to reestablish a robust citizen science field photo processing pipeline that can be applied to other use cases.