Global Land cover

For these global 100m resolution maps, the main inputs are PROBA-V satellite observations, organized into millions of Sentinel-2 equivalent tiles of 110x110km. The processing in this tiling grid, and with UTM projection, ensures high quality and facilitates the continuity with Sentinel-2 observations.

These satellite observations are put through an innovative algorithm, that includes:

its own satellite data pre-processing, with geometric and atmospheric corrections;
data cleaning by sensor specific status masks and (temporal) outlier detection techniques;
calculation of the input data density indicator
data fusion between 5-daily 100m resolution and daily 300m resolution data;
extraction of 183 metrics, including the base reflectances, vegetation indicators, time series harmonics and descriptive statistics.
the use of 168K training points, collected at 10m resolution, from GeoWIKI’s crowd-sourcing, for year 2015
the use of well-established, external datasets for the shoreline masking, ecological regionalisation, built-up (urban) cover, permanent and seasonal water cover, arctic vegetation, weather and topography;
supervised classification and regression.

The classified metrics are calculated over a three year period (epoch), in three processing modes:

base maps for epoch 2015, that serves as reference for the classifier and regression models,
consolidated maps (epochs 2016, 2017 and 2018) with full years of prior and pastor data and
the near-real time (nrt, 2019) maps with one year prior and only three months pastor data.

Time series break maps are computed using firstly a BFAST break detection algorithm on a time series of MODIS Near-Infrared Reflectance of Vegetation(NIRv) input data and secondly a Hidden Markov Model. These break maps show the areas where changes occurred between years and are used in temporal post-processing rules to improve the cover fraction time series. The changes in land cover maps are aligned with these break predictions as well.

The final maps are validated using an independent set of 21.7K validation points, also sourced from Geo-WIKI.

The processing continued to use the innovative Big Data techniques on the PROBA-V Mission Exploitation Platform.