One of the major limitations of remote sensing flood detection is the presence of vegetation. Our study focuses on a flood classification using Radarsat-2 Quad-Pol data in a natural floodplain during leafless, dry vegetation (early spring) state. We conducted a supervised classification of a data set composed of nine polarimetric decompositions and Shannon entropy followed by the predictors' importance estimation to reveal which decomposed component had the strongest effect on classification models. Also, we tested two variants of polarimetric speckle filtering to see if this step influences the results. The classification accuracy was 0.78 and 0.83 for the boxcar and improved Lee sigma filter respectively. The Cloude - Pottier decomposition produced the highest accuracy (0.67) in a single-decomposition scenario, but the volume component of Pauli decomposition was the most important for classification in a multi-decomposition scenario.