Revisiting the effect of spatial resolution on information content based on classification results

TL;DR

This study investigates how spatial resolution and filtering techniques affect the classification accuracy of PolSAR images, revealing that filtering generally improves classification performance up to a certain filter size.

Contribution

It provides an empirical analysis of the impact of different filtering methods and parameters on classification accuracy, Kappa, and overall map quality in PolSAR images.

Findings

Filtering improves classification accuracy up to 7x7 filter size.

Support Vector Machine outperforms Maximum Likelihood in classification.

Filtering enhances map quality metrics like Kappa and Overall Accuracy.

Abstract

Polarimetric Synthetic Aperture Radar (PolSAR) images are an important source of information. Speckle noise gives SAR images a granular appearance that makes interpretation and analysis hard tasks. A major issue is the assessment of information content in these kind of images, and how it is affected by usual processing techniques. Previous works have resulted in various approaches for quantifying image information content. As Narayanan, Desetty, and Reichenbach(2002) we study this problem from the classification accuracy viewpoint, focusing in the filtering and the classification stages. Thus, through classified images we verify how changing properties of the input data affects their quality. Our input is an actual PolSAR image, the control parameters are the filter (Local Mean or Model Based PolSAR, MBPolSAR), the size of them and the classification method (Maximum Likelihood, ML, or Support Vector Machine, SVM), and the output are the classification precision obtained applying the classification algorithm to the filtered data. To expand the conclusions, this study deals not only with Classification Accuracy, but also with Kappa and Overall Accuracy as measures of map precision. Experiments were conducted on two airborne PolSAR images. Unless Narayanan, Desetty, and Reichenbach(2002) almost all measure values are good and increase with degradation, i.e. the filtering algorithm that we used always improves the classification results at least up to 7x7.