Learning-based Soil Property Estimation using Remote Sensing Data Focusing on Spectral Complexity, Noise and Vegetation Interference

Healthy soils are essential to agricultural productivity, ecosystem resilience, and sustainable land management. Estimating key soil properties, such as soil organic carbon (SOC), moisture, and nitrogen content, has traditionally relied on laboratory analyses, manual field surveys, and expert visual assessments. While these approaches offer accurate local measurements, they are time-consuming, costly, and unsuitable for monitoring large or diverse regions. Satellite remote sensing offers a more scalable and timely alternative by enabling repeated observation across vast areas. However, existing approaches using hyperspectral and multispectral satellite images often face limitations due to spectral noise, vegetation interference, and complex interactions between soil and vegetation that distort reflectance signals. As a result, conventional remote sensing models frequently fall short in delivering reliable and consistent soil property estimates under varying environmental conditions.

To address these challenges, this thesis presents a series of innovative methodologies that improve the estimation of soil properties from satellite imagery using data-driven approaches. It begins by developing machine learning frameworks tailored to hyperspectral and multispectral satellite data, with a focus on effective feature selection and dimensionality reduction to improve estimation performance. The research then examines the potential of using limited spectral information, such as visual bands, to support rapid and accessible assessments where full-spectrum data may not be available. Together, these developments lay the foundation for building more accurate and scalable soil monitoring systems.

To mitigate the confounding effects of moisture and vegetation, new spectral indices are introduced that specifically account for dry soil and dry vegetation conditions. Beyond feature engineering, advanced modeling approaches are developed, including a Hybrid Bayesian Attention Model that captures both key statistical relationships and uncertainty, and a Transformer-guided framework for detecting and correcting noise in satellite reflections. Additionally, a generative adversarial network, named as ReflectGAN, is proposed to reconstruct bare soil reflectance from vegetation-contaminated observations in order to improve soil property estimation in mixed or densely vegetated landscapes, where direct interpretation of satellite signals is often unreliable.
Validated across multiple benchmark datasets and diverse landscape conditions, the proposed methodologies demonstrate clear improvements over existing approaches in handling noise, vegetation interference, and spectral complexity. These innovations contribute to more reliable, interpretable, and scalable estimation of soil properties from satellite imagery. By addressing long-standing challenges in remote sensing-based soil monitoring, this research advances data-driven strategies for sustainable agriculture, environmental observation, and informed land management.