ReflectGAN: Modeling Vegetation Effects for Soil Carbon Estimation from Satellite Imagery

TL;DR

SOC estimation from satellite imagery is hampered by vegetation, which distorts soil reflectance. ReflectGAN introduces a paired-GAN that learns a conditional, residual spectral translation from vegetated to bare-soil reflectance, enabling more accurate SOC estimation across mixed land covers. The method is validated on Landsat 8 (and cross-validated on Sentinel-2), showing substantial improvements in $R^2$, RMSE, and $RPD$ over traditional vegetation corrections and unpaired GANs, with best results around $R^2\approx 0.53$–0.55 and $RMSE\approx 3.9$–$5.3$. This approach expands reliable soil monitoring to vegetation-rich landscapes and can be integrated into operational remote sensing workflows, offering scalable, physically meaningful spectrally corrected inputs for SOC models.

Abstract

Soil organic carbon (SOC) is a critical indicator of soil health, but its accurate estimation from satellite imagery is hindered in vegetated regions due to spectral contamination from plant cover, which obscures soil reflectance and reduces model reliability. This study proposes the Reflectance Transformation Generative Adversarial Network (ReflectGAN), a novel paired GAN-based framework designed to reconstruct accurate bare soil reflectance from vegetated soil satellite observations. By learning the spectral transformation between vegetated and bare soil reflectance, ReflectGAN facilitates more precise SOC estimation under mixed land cover conditions. Using the LUCAS 2018 dataset and corresponding Landsat 8 imagery, we trained multiple learning-based models on both original and ReflectGAN-reconstructed reflectance inputs. Models trained on ReflectGAN outputs consistently outperformed those using existing vegetation correction methods. For example, the best-performing model (RF) achieved an $R^2$ of 0.54, RMSE of 3.95, and RPD of 2.07 when applied to the ReflectGAN-generated signals, representing a 35\% increase in $R^2$, a 43\% reduction in RMSE, and a 43\% improvement in RPD compared to the best existing method (PMM-SU). The performance of the models with ReflectGAN is also better compared to their counterparts when applied to another dataset, i.e., Sentinel-2 imagery. These findings demonstrate the potential of ReflectGAN to improve SOC estimation accuracy in vegetated landscapes, supporting more reliable soil monitoring.

Figures (5)

• Figure 1: The proposed ReflectGAN architecture to illustrate the interaction between the Generator and Discriminator, where the Generator transforms vegetated soil reflectance into bare soil-like spectra, and the Discriminator evaluates their authenticity.
• Figure 2: Schematic Diagram of the Proposed ReflectGAN Generator, illustrating the flow from input vegetated soil spectral signatures through the series of residual blocks to the output of generated bare soil spectral signatures.
• Figure 3: Schematic Diagram of the Proposed ReflectGAN Discriminator, showcasing the sequence of processing layers from concatenated vegetated and bare soil spectral signature inputs to the binary classification output.
• Figure 4: Illustration of the proposed framework for soil organic carbon estimation, encompassing data preprocessing, ReflectGAN-based reflectance correction, model training, and evaluation.
• Figure 5: Comparison of different soil reflectance types against ReflectGAN-reconstructed reflectance, illustrating how the proposed method transforms vegetated reflectance into bare soil reflectance under varying vegetation densities.