Quantifying the sensitivity of SAR and optical images three-level fusions in land cover classification to registration errors

Abstract

Synthetic aperture radar (SAR) and optical (SAR-optical) data are two important remote sensing data sources. Fusing them is expected to lead to complementary information that benefits land cover classification (LCC). The fusion of SAR-optical images occurs in three levels, i.e., pixel-level, feature-level, and decision-level. However, accurately registering SAR-optical images is still a challenge, and geometric registration errors will bring great uncertainty to LCC based on SAR-optical images fusion. Therefore, this study quantitatively evaluates the sensitivity of SAR-optical images three-level fusions to registration errors through simulation experiments. The results show that: (1) geometric registration errors affect LCC based on SAR-optical images fusion at three levels in a significant manner. Among them, feature-level fusion is the least sensitive to registration errors. (2) the fusion taken optical image as reference image presents better tolerance to registration errors than that based on SAR image. (3) the response of SAR-optical images fusion to registration errors in LCC of heterogeneous regions is greater than that in homogeneous regions. (4) during the LCC, the fusion of SAR-optical images with comparable spatial resolution has a higher tolerance of registration errors. (5) fusing SAR-optical data does not always guarantee the improvement of LCC compared to using optical data alone, depending on fusion levels, fusion methods and classifiers. We believe these findings can greatly benefit the research community in developing new SAR-optical images fusion methods to improve LCC in the future.

Keywords

• SAR and optical (SAR-optical) images fusion
• Registration errors
• Sensitivity quantification
• Land cover classification (LCC)