The optimal processing chain for flood mapping using polarimetric SAR in a temperate zone wetland

In this study, we investigated the influence of speckle filters and decomposition methods with different combinations of filter and decomposition windows sizes on classification accuracy. The study area was a part of Biebrza National Park, located in Northeast Poland. The C-Band SAR image from Radarsat 2 sensor was processed using various speckle filters (boxcar, IDAN, improved Lee sigma, refined Lee) in 5x5, 7x7, 9x9, and 11x11 pixel window sizes. We processed the filtered data using nine polarimetric decompositions also in 5x5, 7x7, 9x9, and 11x11 pixel window sizes. We used the calculated polarimetric features to conduct a supervised classification with random forest machine learning algorithms for each combination of processing parameters in three different scenarios: (1) each decomposition product was used separately as a model input; (2) all decomposition products with the same speckle filtering method were used as a model input; (3) all decomposition products with all speckle filtering methods were used together as the model input. Overall, the most accurate classification model (87%) was produced in scenario 3 with an 11x11 filter and decomposition windows. In scenario 1, the highest overall accuracy achieved the Cloude-Pottier decomposition (72%) and the lowest produced the Touzi decomposition (38%). In scenario 2, the IDAN filter provided the highest accuracy (84%) with an 11x11 filter window and a 9x9 decomposition window. The obtained results show that the selection of appropriate processing parameters is an important step in the SAR data classification workflow. Our study also indicates the most suitable combination of radar image processing parameters for wetland classification.