Seasonal fractional cover - Landsat, JRSRP algorithm Version 3.0, Australia coverage

Summary of processing: Landsat surface reflectance data > multiple single-date fractional cover datasets > medoid calculation for seasonal composite of fractional cover
Further details are provided in the Methods section.

Data Creation
Image Pre-Processing: All input Landsat TM/ETM+/OLI imagery was downloaded from the USGS EarthExplorer website as level L1T imagery. Images which the EarthExplorer site rated as having greater than 80% cloud cover were not downloaded. The imagery has been corrected for atmospheric effects, and bi-directional reflectance and topographic effects, using the methods detailed by Flood et al (2013). The result is surface reflectance standardised to a fixed viewing and illumination geometry. Cloud, cloud shadow and snow have been masked out using the Fmask automatic cloud mask algorithm. Topographic shadowing has been masked using the Shuttle Radar Topographic Mission DEM at 30 m resolution. Water has been masked out using the methods outlines in Danaher & Collett (2006).
Fractional Cover Model (Version 3.0): A multilayer perceptron (MLP) model is used to estimate percentage cover in three fractions – bare ground, photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) from surface reflectance, for every image captured within the season. The MLP model was trained with Tensorflow using Landsat TM, ETM+ and OLI surface reflectance and a collection of approximately 4000 field observations of overstorey and ground cover. The field observations covered a wide variety of vegetation, soil and climate types across Australia, and were collected between 1997 and 2018 following the procedure outlined in Muir et al (2011). The model was assessed to predict the vegetation cover fractions with MAE/wMAPE/RMSE of: bare - 6.9%/34.9%/14.5% PV - 4.6%/37.9%/10.6% NPV - 9.8%/25.2%/16.9%.
Seasonal Compositing: The method of compositing used selection of representative pixels through the determination of the medoid (multi-dimensional equivalent of the median) of three months (a season) of fractional cover imagery. The medoid is the point which minimises the total distance between the selected point and all other points. Thus the selected point is “in the middle” of the set of points. The value selected is a specific data point and not an averaged or blended value. It is robust against extreme values, inherently avoiding the selection of outliers, such as occurs when cloud or cloud shadow goes undetected. At least three pixels from the time-series of imagery for the season must be available. Unfortunately, due to the high level of cloud cover in some areas, often three cloud free pixels are not available, resulting in data gaps in the seasonal fractional cover image. For further details on this method see Flood (2013).