Towards an Automated Mapping of Trees Outside Forest using Google Earth Engine

The Google Earth Engine cloud-computing platform (GEE) allows users to perform geospatial analysis from local to planetary scale based on Google's cloud infrastructure in a short time. Users can access and analyse data from a large repository of publicly available geospatial dataset, including more than forty years of historical imagery, such as the entire Landsat archive as well as the complete Copernicus Sentinel archive, and a variety of earth science-related datasets.Applications of GEE are growing in many fields of earth science, such as global forest changes, effects on the global water cycle, land cover/land use changes, flood mapping, urban mapping and so on. In this study we used GEE for a just apparently trivial problem: mapping Tree Outside Forest (TOF). Since small woods, tree hedgerows, scattered and isolated trees are key TOF features of rural and urban landscapes all around the world and provide a variety of products and ecosystem services essential for human well-being, the interest in estimating and mapping TOF coverage is increasing. For these reasons, filling the lack of information on TOF extension is crucial to develop effective agro-environmental measures and rural development policies.We followed the approach proposed in our recent studies, based on Sentinel-2 imagery, consisting in: 1) an automatic identification of tree covered surface, by applying a statistical threshold on several vegetation related optical indices (NDVI, EVI, Negative Luminance etc.), 2) an object-based image analysis to classify TOF elements, 3) a ground truth validation process.We developed an openly available GEE script performing the following steps: 1) retrieval of a collection of Sentinel-2 images between two dates in a dynamically chosen Area of Interest (AoI), with a user-defined maximum cloud coverage. 2) Extraction of the minimum for each band, obtaining a single image, from which green, blue, red and near-infrared (both 8 and 8A bands) are extracted and clipped to the AoI. 3) Evaluation of an optical index and relative histogram for manually choosing the threshold value for tree cover identification. 4) Creation of a polygon coverage for trees, taking the optical index values that exceed the given threshold. The polygons are then equipped with a shape factor and a pixel count field. 5) Extraction of TOF polygons from trees, based on shape and size. 6) Graphical GIS-like presentation of trees and TOF polygons over the optical indices and Sentinel-2 layers. All the layers can be exported as georeferenced shapefiles or geoTiffs, for ground truth validation and further off-cloud spatial analysis.In this study we exploited many capabilities of GEE. Cloud-computing speeds-up the conventional pre-processing phase to an unprecedented level (images retrieval, fusion and index computation is almost instantaneous and visually driven) as well as the processing phase (vegetation index extraction, histogram evaluation, thresholding and coverage vectorisation), including some geospatial raster (filtering, thresholding, comparison) as well as vector analysis. The only operator-based step consists in the vegetation index choice and threshold selection for the trees identification. The identification rationale consists in the presence of a distinguished peak in the vegetation index of the AoI, which is to be selected by the operator within the GEE script execution, mouse-clicking on the histogram graph.Our script has been developed and tested on an agroforestry landscape in Umbria, central Italy, but it could be run potentially all over the world, choosing the best-performing optical index or linear bands combination.Further improvements could include a fully automated vegetation index choice and threshold selection, limiting the operator's role to the choice of the AoI. The ground truth validation is meant to be performed manually, but it could be easily implemented within the GEE platform itself, uploading ones' vector validation layers as so-called GEE assets, from ordinary classified shapefiles.