Image Upscaling Assesment From UAV To Sentinel-2 In Coastal Wetlands

Coastal wetlands provide a range of ecosystem services and can support quite high biodiversity as a result of their high productivity. There are a range of techniques applied to monitoring and assessing ecological status and ecosystem service provision, however, traditional techniques can be quite time consuming and costly. In recent years, there has been a strong push to use remotely sensed data to evaluate ecological condition as well as estimate a range of ecosystem services within coastal wetlands.  Unmanned Aerial Vehicles (UAV) platforms have increasingly been used in the field of remote sensing of coastal wetlands because they provide detailed radiometric data to carry out the classification of the high-resolution images. Classifications using supervised Machine Learning algorithms can be performed on those images, providing robust datasets for a range of variables.

However, in spite of the flexibility of performing flight plans to monitor coastal wetlands with high accuracy, it is often not feasible to capture large areas using UAV systems. Satellite imagery can be used to undertake evaluations of a wide range of environmental variables in coastal wetlands over much larger areas. Finding synergies between images taken from UAVs and satellite could provide the possibility to extend local observations of plant functional diversity or ecosystem service provision in coastal wetlands to larger areas or to regions. Using validation techniques based on ground-truth data, high-resolution UAV derived images can be used to characterize terrain and ecological features, such as plant communities and then upscale them to satellite resolutions.

The present study presents a methodology to compare images taken from a UAV multispectral camera and the freely available Multispectral Instrument (MSI) sensor images from the Sentinel-2 satellite because their spectral bands overlap with those commonly used for plant community assessments in coastal wetlands using drones. First, each pixel of Sentinel-2 image is characterized by the most frequent category of plant communities obtained from a ML supervised classification of high-resolution UAV image. Then, the results of classifying the study areas with the Sentinel-2 image are compared with the previous process by analyzing the differences and similarities of categories in each pixel. By this way, synergies between the UAV and Sentinel-2 images can be found in order to have a reliable upscaling of UAV-based data. 
Keywords: Remote Sensing, UAV, Machine Learning, Upscaling