Evaluation of Polarimetric SAR Despeckling Methods for Crop Classification from RCM Compact Polarimetry Data

Abstract. The presence of speckle in RADARSAT Constellation Mission (RCM) Compact Polarimetry (CP) Synthetic Aperture Radar (SAR) images can impair the performance of information extraction applications such as classification. Therefore, a critical preprocessing step known as despeckling is necessary to mitigate this granular, noise-like phenomenon in these images. This paper compared several PolSAR speckle reduction methods, including Box Car, IDAN, Lee Refined, Lee Sigma, Improved Lee Sigma, and Lopez filters. A CP SAR dataset collected over agricultural land in southern Quebec, QC, Canada, was utilized for the study. The assessment of despeckling was based on various no-reference quantitative indicators. Each despeckling method was evaluated for its effectiveness in reducing speckle in homogeneous areas, preserving details, and avoiding radiometric distortion. Additionally, the impact of despeckling on the classification of this agricultural land was assessed using the Random Forest classifier. The Stokes parameters, m-chi decomposition, and intensity images were utilized for this purpose. Experimental results indicated that the Box Car method excelled in speckle suppression at the expense of edge over-smoothing. Furthermore, the Lee Sigma and Improved Lee Sigma methods were the most effective in speckle reduction from homogeneous areas while preserving edges and preventing radiometric distortion. Moreover, the classification results demonstrated that appropriate despeckling could significantly enhance classification accuracy.