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

Dristi Datta, Manoranjan Paul, Manzur Murshed, Shyh Wei Teng, and Leigh M. Schmidtke · 2025

Read ReflectGAN: Modeling Vegetation Effects for Soil Carbon Estimation from Satellite Imagery by Dristi Datta, Manoranjan Paul, Manzur Murshed, Shyh Wei Teng, and Leigh M. Schmidtke (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-18546.txt.

Introduction

Dristi Datta and four co-authors build ReflectGAN, a model that strips the mark of plant cover from satellite reads of soil, so soil carbon can be gauged. Their paper is from two thousand twenty-five.

Soil carbon is a key sign of soil health. From space, plant cover taints the soil's light, which hides its true tone and weakens carbon models. The new model learns how a covered soil's light turns into a bare soil's light. It then rebuilds the bare form from the covered one. The team trains carbon models on both raw and rebuilt reads, from a large land data set. Models fed the rebuilt reads beat those fed raw ones, and beat past plant-strip ways.

Catalog record

CREATORDristi Datta, Manoranjan Paul, Manzur Murshed, Shyh Wei Teng, and Leigh M. Schmidtke
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-18546

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ReflectGAN: Modeling Vegetation Effects for Soil Carbon Estimation from Satellite Imagery Dristi Datta, Manoranjan Paul, Manzur Murshed, Shyh Wei Teng, and Leigh M. Schmidtke 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 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 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. I Introduction Soil organic carbon (SOC) is a fundamental indicator of soil health, influencing agricultural productivity, carbon sequestration, improved soil moisture retention and overall ecosystem sustainability. Accurate estimation of SOC is essential for promoting sustainable agriculture, improving soil management practices, and monitoring environmental changes [ 1 , 2 ] . Traditional methods for estimating SOC rely on laboratory-based soil analyses, which, although precise, are labor-intensive, costly, and limited in spatial coverage [ 3 , 4 ] . These constraints have led to growing adoption of remote sensing technologies as a scalable and cost-effective alternative to conventional field sampling and laboratory analysis. Laboratory-based hyperspectral imaging (HSI) provides a powerful tool for SOC estimation by offering high spatial and spectral resolution, enabling detailed analysis of soil properties without the need for destructive sampling [ 5 , 6 , 7 ] . Numerous studies have validated the effectiveness of HSI in accurately estimating SOC levels [ 8 , 7 ] . However, the widespread deployment of HSI is constrained by the high cost of equipment and limited accessibility, making it impractical for large-scale applications. These limitations have prompted interest in more affordable alternatives, such as visual-band multispectral imaging, which can achieve competitive SOC estimation accuracy at substantially lower costs [ 9 , 10 ] . Still, both HSI and multispectral imaging systems are typically confined to localized areas. To address these spatial constraints, satellite-based remote sensing has emerged as a promising approach for large-scale SOC estimation, offering broader spatial coverage and cost-effective monitoring across diverse landscapes. Multispectral satellite platforms such as Sentinel-2 (S2) and Landsat 8 (L8) offer broad spatial coverage and public accessibility, making them attractive for large-scale SOC estimation. However, their estimation accuracy often lags behind hyperspectral approaches due to environmental noise, limited spectral resolution, and the need for extensive preprocessing [ 11 , 12 , 13 , 14 ] . To mitigate these challenges, recent studies have integrated supplementary features—such as vegetation indices (VIs), soil indices (SIs), and Tasseled Cap Transformation (TCT) components—alongside raw spectral bands. This feature augmentation has been shown to improve the performance of learning-based models by enhancing their ability to capture the complex variability associated with SOC [ 13 ] . Nevertheless, most existing studies have primarily focused on bare soil conditions, where minimal or no vegetation is present. This emphasis stems from the fact that vegetation substantially alters surface reflectance, masking the underlying soil characteristics and complicating SOC estimation. Consequently, models are often trained exclusively on bare soil samples to minimize spectral interference. However, in real-world agricultural and natural landscapes, bare soil exposures are seasonal and limited, underscoring the importance of developing methods that can reliably estimate SOC under vegetated conditions—an area that remains insufficiently explored. Traditional approaches to vegetation correction in satellite-based SOC estimation include vegetation index-based techniques, spectral unmixing models, and radiative transfer frameworks. VIs, such as NDVI and SAVI, are commonly used to estimate vegetation cover and partially reduce its impact on reflectance measurements [ 15 ] . While computationally efficient, these methods do not recover the true underlying soil reflectance and often yield poor SOC estimation performance in densely vegetated areas [ 16 ] . More physically grounded models, such as PROSAIL [ 17 ] , simulate canopy–soil reflectance interactions using radiative transfer theory. However, their application is limited by the need for numerous biophysical parameters (e.g., leaf area index, leaf angle, soil moisture), which are difficult to obtain at scale. Spectral unmixing approaches, such as Spectral Mixture Analysis (SMA) [ 18 , 19 ] and Adaptive SMA (ASMA) [ 20 ] , attempt to separate soil and vegetation components by estimating fractional abundances of endmembers. While conceptually appealing, these methods assume linear mixing models and often rely on idealized endmembers that may not adequately capture the nonlinear reflectance behavior of real-world landscapes. To address these limitations, several nonlinear unmixing models have been proposed, which are better suited for complex environments where reflectance is inherently nonlinear. These include autoencoder-based networks grounded in the Multilinear Mixing Model, which effectively model complex nonlinear interactions in hyperspectral data [ 21 ] , model-based deep autoencoder networks that incorporate prior knowledge of the mixing process to enhance unmixing performance [ 22 ] , and interpretable deep learning (DL) methods that account for nonlinearity and endmember variability through a probabilistic variational framework [ 23 ] . Probabilistic extensions like the Probabilistic Mixture Model-based Spectral Unmixing (PMM-SU) [ 24 , 25 ] account for uncertainty in unmixing but still treat vegetation as a contaminant to be separated rather than learning to reconstruct a soil-equivalent reflectance. An alternative to spectral unmixing is to bypass contaminated signals entirely by identifying and excluding vegetation-affected samples. Recent approaches, such as transformer-guided noise detection, aim to filter out anomalous reflectance patterns that deviate from expected soil signals [ 26 ] . Although effective in improving training quality, these techniques reduce sample availability and limit the model’s applicability in persistently vegetated regions. Reconstruction-based methods offer a promising solution by learning to generate bare soil reflectance directly from vegetation-contaminated inputs. Some attention-based DL models have been proposed to model soil–vegetation interactions [ 27 ] , but they do not explicitly reconstruct bare soil reflectance. As a result, residual vegetation signals may persist in the output, reducing interpretability and potentially degrading SOC estimation accuracy. Explicit reconstruction is essential to ensure that the reflectance conforms to bare soil spectral patterns, thereby enhancing the reliability of the SOC model and facilitating comparisons with bare-soil-based studies. Generative Adversarial Networks (GANs) [ 28 ] offer a data-driven approach for spectral reconstruction. While Pix2Pix [ 29 ] and CycleGAN [ 30 ] have been successful in visual tasks, their objective functions typically prioritize image realism and cyclic consistency over spectral fidelity, making them less suitable for reflectance recovery

Read the complete text