Mapping of invasive plant species with Sentinel-1 and -2 data calibrated with UAV-based training data

Up-to-date maps on depicting the current status of plant invasions are highly valuable for efficient mitigation measures. In this context, remote sensing has been reported to be a useful tool for continuous mapping of invasive species over larger spatial extents. So far an important restriction for the development of flexible, operational approaches was the limited availability of cost-free and spatially highly-resolved datasets in many parts of the world. Since 2014, ESA´s Sentinel-1 and 2 satellites provide cost-free EO data with global coverage, relatively high temporal and spatial resolution. This data is ascribed a high potential for differentiating plant species. One important perquisite to operationally use such data for an operational mapping of invasive species is an efficient collection of reference data to train and validate mapping procedures applied to Sentinel data. We hypothesize that data collected from unmannead aerial vehicles (UAV) can be an efficient alternative of traditional field surveys. Applying UAV data instead of GNSS-coded (global navigation satellite system) field data comprises several advantages: (1) more area can be mapped in a given time frame, (2) increased area accessibility, (3) the UAV data share the bird’s eye perspective of the satellite data and are hence directly compatible, and (4) the extraction of target species may be possible with automatized classification algorithms. We hence developed and tested a workflow for three invasive woody plant species in southern Chile which firstly includes an automatic extraction of spatially continuous species cover from UAV imagery by combining a sparse set of photo-interpreted presence points of the target species with a Maximum Entropy (MaxEnt) classifier. Secondly, we use this species cover data as reference to train a random forest model with multitemporal Sentinel observations to predict the canopy cover of the three target species over large areas. Our results show that the three invasive species considered can be mapped with very high accuracy as there is a very high agreement between UAV and Sentinel predictions (R2 > 0.9, NRMSE < 10%). We thus conclude that the proposed methodology can be used as a blueprint for operational monitoring of invasive plant species.