Improving Classification Error for Mixed Pixels in Satellite Images Using Soft Thresholding Technique

Identification of earth surface features and measurement of their geophysical properties using electromagnetic radiation is referred to as remote sensing. The major characteristics of an image captured by a sensor varies in terms of spectral, spatial, temporal and polarization signature which enables differentiation between targets. As compared to traditional methods of land cover mapping, Synoptic view with its repetitive coverage along with the precisely calibrated sensors to detect changes and observations at different resolutions offers a better solution for earth surface monitoring. Statistical based methods offers solution for land use/land cover mapping which works better in case of normal images but in case of remote sensed images, as their no clear distinction between pixels of two different terrains the problem of mixed pixels occurs. Accurate classification of such pixels becomes extremely difficult. The purpose of this paper is to understand mixed pixel problems and its contribution to overall classification error. Precise classification of such mixed pixels is possible using soft classification techniques compared to its traditional counterpart i.e., hard classification technique. Using similarity based region matching technique and setting a flexible threshold value increases the recognition rate to 79.16% as compared to setting a fixed threshold value which gives recognition rate of 66.1%. Results show considerable improvement in classification accuracy for rough terrains compared to smooth terrains. SAR images are used for testing purpose wherein classification accuracy increases to 76% when single pixel is mapped to multiple classes as compared to classification accuracy when a single pixel is mapped to a dedicated class.