The influence of filtration and decomposition window size on the threshold value and accuracy of land-cover classification of polarimetric SAR images

In this study we use ALOS PALSAR satellite data to classify land cover using a decision tree algorithm. We apply polarimetric decomposition methods to coherence and covariance matrices obtained from the data and then use threshold values to classify terrain. We evaluate the influence of speckle filter and decomposition window sizes on the threshold value used in the decision algorithm and on the accuracy of the classification. We also study the sensitivity of the classification to the accuracy of the threshold value.

First, we processed a fully polarimetric Synthetic Aperture Radar (SAR) L-band image using different sizes of speckle filtration and decomposition window (3 × 3 pixels, 5 × 5, 7 × 7, 9 × 9), and the decomposition methods available in PolSARPro software. We evaluated these methods and chose the most efficient. Then we developed a simple hierarchical classification scheme based on threshold values. In the first step we divided the terrain into smooth and rough areas and then separated these into more detailed subclasses (water and agriculture, and forest and urban) which correspond to smooth and rough areas, respectively. A more detailed analysis separated continuous and discontinuous urban fabric and deciduous and coniferous forests. The maximum overall accuracy of the classification was 86.1% for the four main land cover classes, and 80.4% for the six more detailed classes. The accuracy of the classification dropped by about 10% when non-optimal window sizes were used in image filtration or decomposition.