Systematic Evaluation of Machine Learning Models for Regression-Based Error Refinement in SAR-to-Optical Image Translation for Cloud Removal

초록

Generative deep learning-based synthetic aperture radar (SAR)-to-optical image translation (SOIT) has been widely employed for cloud removal. However, since cloud-contaminated regions reconstructed by SOIT inevitably contain prediction errors, an additional error refinement procedure is required to achieve reliable spectral reflectance reconstruction. In this study, three machine learning-based regression models, including Random Forest (RF), eXtreme Gradient Boosting (XGB), and Natural Gradient Boosting (NGB), are comprehensively evaluated for the error refinement of optical imagery initially reconstructed by SOIT. The factors influencing refinement performance are categorized into four components: (1) the sampling strategy of training pixels from cloud-free regions (random vs. quantile-based sampling); (2) the refinement target (actual spectral reflectance vs. residual between actual and initially reconstructed reflectance); (3) SAR features (pixel-level raw SAR features vs. local spatial SAR features); and (4) the cloud fraction in the scene of interest. A systematic sensitivity analysis of their effects on error refinement performance was conducted over cropland using PlanetScope optical imagery and COSMO-SkyMed SAR imagery. The results showed that cloud fraction had the greatest impact on refinement performance. Regarding SAR features for regression, the use of local spatial SAR features improved spectral similarity by up to approximately 4.6%p compared to raw SAR features. In terms of sampling strategy, quantile-based sampling yielded better refinement performance, whereas the effect of the refinement target was less pronounced. These results suggest that local spatial SAR features and quantile-based sampling strategies are the key determinants of regression-based refinement performance in SOIT-based cloud removal.

키워드