Estimating the distribution of reed Phragmites australis in Britain demonstrates challenges of remotely sensing rare land cover types at large spatial scales

Abstract

Reed Phragmites australis is important for biodiversity, for ecosystem services, and as a resource for humans. Already one of the mostly widely distributed wetland plants globally, reed has recently expanded outside of its native range, modifying ecosystems. However, like most wetland plant communities, reedbed has rarely been mapped at large geographical scales, restricting the information available to ecologists and resource managers. Using Sentinel-2 data and machine learning in open-source software, we produce the first remotely-sensed reedbed map of Britain. A random forest was trained on 79.2 ha of reedbed and 2,719.2 ha of non-reedbed land cover, using free online imagery. Accuracy was high within the training area (AUC > 0.998); however, field validation accuracy was much lower (AUC = 0.671), with many false positives (commission error of 93.4%). A similar workflow carried out in Google Earth Engine, using nearly an order of magnitude more images, gave a lower commission error but a disproportionately higher omission error. Due to the classification error, our map is more useful for a non-spatial estimate of the overall reedbed extent in Britain, rather than for the spatial location of reedbeds in Britain. Using the known commission and omission error, we estimate that in 2015 - 2017 c. 7800 ha of Britain was reedbed. Our study highlights the issues that present enduring barriers to accurate land cover classification at large spatial scales, perhaps suggesting fruitful areas for technological innovation. Even with a big data approach and even if technological issues are resolved, ecological factors such as confusion land cover types and geographical variation in temporal reflectance function will probably continue to impose upper limits on the size of area for which land cover can be classified accurately, and therefore on the utility of remote sensing to resource managers.