Conditional GAN-Based NDVI Estimation From Multifrequency SAR Data in Forest Ecosystems

Abstract

The combined use of remote sensing satellite imagery, particularly the extraction of optical data from synthetic aperture radar (SAR) data to fill gaps, is used to overcome the disadvantages of optical data in temporal analysis. This study investigates the contribution of multifrequency SAR data to the estimation of optical data within this context. This letter presents an SAR-based framework for estimating Sentinel-2 normalized difference vegetation index (NDVI) by fusing C-band Sentinel-1 and L-band ALOS-2 PALSAR-2 data over a heterogeneous forest ecosystem. Three machine learning approaches: random forest (RF), extreme gradient boosting (XGBoost), and a conditional generative adversarial network (cGAN) were comparatively evaluated. A compact feature set derived from multipolarization and multifrequency SAR observations was selected to represent canopy structure. Results demonstrate that the cGAN significantly outperforms ensemble tree-based models, achieving an R-2 of 0.95 and a mean absolute error (MAE) of 0.0031, while RF and XGBoost yielded R-2 values below 0.55. The findings indicate that adversarial learning effectively captures the complex, nonlinear relationship between SAR backscatter and NDVI, enabling reliable all-weather vegetation monitoring in structurally complex forest environments.