Potential and limitations of crowdsourced data for high-resolution rice mapping in Madagascar: The importance of representation

Abstract

Given the projected population growth and rice consumption in Madagascar, assessing its rice production potential is important. Existing spatial representations of rice are characterized by coarse resolutions and diverge widely. Mapping rice production based on remote sensing offers a potential solution to this problem but its application in Sub-Saharan Africa is hindered by scarce ground truth data and a complex array of cropping systems. While crowdsourcing initiatives provide data that could in principle be used to overcome this hurdle, their usefulness in rice mapping is hitherto underexplored. Here, we investigate the usefulness of crowdsourced data to enhance the mapping of rice production areas in the data-poor context of Madagascar. At present, dedicated crowdsourcing initiatives only limitedly cover rice but do provide sufficient information to build a dataset of data points characterizing a No-Rice class. On the other hand, geotagged pictures from social media, which do document rice production areas, predominantly represent 'scenic' fields appealing to social media contributors, i.e. irrigated, intensive rice production systems. The archetypal nature of these data leads to distinct feature signatures. Consequently, these data could be successfully used for rice mapping in the Lake Alaotra basin, which is characterized by the type of intensive rice cropping that is well-represented by the crowdsourced data. However, it was not possible to construct an accurate map of rice production in the greater Antananarivo region or at the national scale using crowdsourced data as they do not sufficiently represent the present variety in rice production systems. The value of crowdsourced data thus depends on the representativeness of the data, which in turn depends on the application and can be improved through applying a probabilistic sampling design. Our study also highlights the need for data representation checks prior to data augmentation or classification for future applications wanting to construct credible maps using crowdsourcing within their sampling design for which we present a multi-faceted evaluation framework.