Assessment of a multi-sensor approach for noise removal on Landsat-8 OLI time series using CBERS-4 MUX data to improve crop classification based on phenological features

We investigated a method for noise removal on Landsat-8 OLI timeseries using CBERS-4 MUX data to improve crop classification. An algorithm was built to look to the nearest MUX image for each Landsat image, based on user defined time span. The algorithm checks for cloud contaminated pixels on the Landsat time series using Fmask and replaces them with CBERS-4 MUX to build the integrated time series (Landsat-8 OLI+CBERS-4 MUX). Phenological features were extracted from the time series samples for each method (EVI and NDVI original time series and multi-sensor time series, with and without filtering) and subjected to data mining using Random Forest classification. In general, we observed a slight increase in the classification accuracy when using the proposed method. The best result was observed with the EVI integrated filtered time series (78%), followed by the filtered Landsat EVI time series (76%).