Recent Advances in Remote Sensing of Soil Science

Soil is the foundation of terrestrial life, underpinning ecosystem services, food production, and climate regulation. The accurate characterization of soil properties at regional to global scales is both a scientific and societal priority. Remote sensing (RS) provides the observational backbone for such characterization. When combined with artificial intelligence (AI) and machine learning (ML) algorithms, it enables the large-scale, temporally consistent mapping and monitoring of diverse soil attributes that would otherwise remain inaccessible. Our Special Issue, entitled “Remote Sensing Advances for Soil Properties”, aimed to highlight novel methodologies and workflows for estimating and monitoring soil properties utilizing Earth observation data. It attracted nine contributions spanning a wide thematic and methodological spectrum, from AI-driven digital soil mapping (DSM) and uncertainty quantification to erosion modeling and ecological vulnerability assessment.

In an era defined by the urgency of climate change mitigation, the twin imperatives of carbon farming and climate-smart agriculture have brought soil organic carbon (SOC) to the center of policy agendas. Hence, two contributions address large-scale SOC estimation. Kalopesa et al. [1] present a multimodal data fusion framework that combines near-infrared (NIR) spectral recordings from a low-cost handheld sensor with open geospatial covariates derived from RS, including landform descriptors, climate layers, and vegetation indices. Their results demonstrate that a hybrid convolutional neural network–extreme gradient boosting (CNN-XGBoost) architecture substantially outperforms traditional ML methods, reaching an R² of 0.72. The CNN-derived spectral features have been recognized as the dominant predictors, and rainfall data provided the most informative climatic contribution. Kakhani et al. [2] approach SOC estimation from a different perspective, addressing the insufficiently addressed challenge of uncertainty. By integrating conformal prediction with a random forest model trained on the European Land Use and Coverage Area frame Survey (LUCAS) soil dataset, they generate statistically rigorous uncertainty intervals for SOC estimates across seven different ecosystems. Their framework outperforms conventional uncertainty quantification approaches, such as quantile versions of gradient boosting and linear regression. In that regard, their approach provides end users with informative prediction sets that identify samples associated with higher modeling uncertainty, offering clear value for operational soil monitoring services. However, a critical need still exists for SOC monitoring approaches that can reliably track changes in soil health while supporting sustainable agriculture and emerging carbon market opportunities [3]. Contributions to this Special Issue highlight the potential of combining proximal sensing, remote sensing, and machine learning to improve SOC estimation and support more scalable soil monitoring frameworks, while also emphasizing the need for broader validation before operational deployment.

Three-dimensional DSM at the national scale is addressed by Stumpf et al. [4], who present a comprehensive framework for Switzerland, targeting clay content, SOC, pH, and potential cation exchange capacity. Applying Quantile Regression Forest to link a national soil database with covariates derived from light detection and ranging-based elevation models, multitemporal satellite imagery, and climate rasters, they achieve overall model accuracy (R² ranging from 0.64 to 0.76) for the four properties. Their analysis reveals that multiscale terrain covariates are universally influential, while bare soil reflectance is a primary driver for SOC, confirming the supporting role of spectral and topographic information in DSM workflows. The challenge of mapping soil properties beyond bare soil pixels is addressed by Baby George et al. [5], who propose a composite mapping strategy applied to AVIRIS-NG airborne hyperspectral data over agricultural fields in India. By classifying pixels according to their bare soil fractional cover using spectral unmixing and fitting class-specific regression models for clay content, they extend the mapped area by 42.4% relative to strict bare-soil-only approaches. The study further demonstrates the feasibility of transferring this framework to spaceborne hyperspectral missions, such as the DLR Earth Sensing Imaging Spectrometer and PRISMA successors, which will deliver similar spectral datasets at the global scale.

Moving from soil properties to the topic of degradation, we received two relevant studies. Polovina et al. [6] present an innovative application of RS for soil erosion mapping over Bosnia and Herzegovina. They introduce a novel approach for deriving the coefficient of erosion types and extent (φ) in the erosion potential method by applying the bare soil index to a ten-year Landsat time series processed via the Google Earth Engine [7]. The resulting fractional bare soil cover maps achieve an overall accuracy of 85.79% and a Kappa statistic of 0.82, demonstrating that freely available multispectral satellite archives, when combined with temporal compositing strategies, can provide valuable results. These products can reliably support regional erosion monitoring and decision-making for torrent flood defense. Similarly, Han et al. [8] assessed ecological vulnerability at a national scale. They proposed a holistic indicator integrating multiple soil and land degradation signals in Mongolia. Using Moderate-Resolution Imaging Spectroradiometer (MODIS)-derived data spanning from 2000 to 2022, they introduce a new RS-based ecological vulnerability index with an accuracy of 89.39%. Their analysis reveals that Mongolia is classified as moderately vulnerable (mean index 1.57) with a center of vulnerability shifting southwestward, driven predominantly by maximum temperature at high altitudes and in arid regions, and by precipitation in eastern and central zones. The application of the geographical detector and gravity center model provides a methodological blueprint for tracking the temporal dynamics of ecological vulnerability under climate change scenarios.

Collectively, the nine contributions to this Special Issue document the breadth and depth of current research at the intersection of RS and soil science. They reflect a discipline increasingly defined by data fusion across spectral, spatial, and temporal scales, where the application of AI architectures seems capable of exploiting this data richness. Moreover, three works highlighted a growing emphasis on uncertainty quantification and operational transferability. Challenges remain, however, including the need for harmonized, open, and well-curated ground truth soil datasets [9], as well as the reconciliation of satellite-derived products with the inherently three-dimensional nature of soil systems. Future advances in spaceborne hyperspectral missions, multimodal data fusion [10,11], and soil physics-informed ML models [12] are expected to progressively address these limitations, further closing the gap between laboratory-scale soil science and global Earth observation.