The statistical validation meets the CEOS Land Product Validation Stage 4 requirements. The version 3.0 annual 100m land cover classes were mapped with high temporal stability across years and an overall mapping accuracy just over 80% (80.6% in 2015, 80.3% in 2019), when compared to the 28K independent validation points. Improved time series break detection and temporal post-processing, combining a BFAST algorithm on a time series of MODIS Near-Infrared Reflectance on Vegetation (NIRv) data and a Hidden Markov Model.Random Forest model of base year (2015) applied to metrics for later years.near-real time: for the most recent complete year, with full year prior and three months posterior data.consolidated: for epochs with full year prior and posterior data available.Classifier and regression models are re-used in later years. base: reference year, calculated using full year prior and posterior data.Metrics calculated per epoch of three years, with annual updates and three different processing modes:.Addition of near-real time processing mode for year 2019, with one year prior but only three months posterior satellite observations as input.Global coverage, following the demonstration of annual change maps over Africa (for period 2015-2018).Temporal post-processing, including time series break detection, to ensure consistency of the classification and cover fractions across epoch years 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.168 000 points) from Geo-WIKI's crowd-sourcing Extraction of metrics: base reflectances, time series harmonics, vegetation indicators, descriptive over the 3-year epoch period.Data fusion using Kalman filtering between 5-daily 100m and daily 300m input data.Calculation of input data density indicator.
Data cleaning by sensor specific status masks and (temporal) outlier detection techniques.Dedicated satellite pre-processing on Sentinel-2 tiling grid (110x110km), with geometric and atmospheric corrections.Classification and cover fraction generation, per 3-year epoch, including:
0 kommentar(er)
