Abstract Soil degradation poses a significant threat to both food security and climate change. Remote Sensing offers valuable insights for soil monitoring, enabling cost-effective observation across extensive regions and extended periods of time. This study evaluates bare soil reflectance mapping at medium spatial resolution by benchmarking the performance of various compositing approaches, providing an assessment of the contribution of different techniques—including the simultaneous use of multiple spectral indices, different compositing, masking or thresholding techniques, and time series parameters like cloud cover and time range. Focusing on Greece, a Mediterranean country with diverse microclimates and soil types, the study leverages Landsat 8 images spanning from 2015 to 2020 and the LUCAS 2015 database to evaluate the results. A wide range of experiments were conducted to determine the best approach for creating a bare soil reflectance composite (BSC), evaluated based on its correlation per spectral band with spectral reference data and its performance in soil organic carbon prediction. The study demonstrated that estimating bare soil reflectance from multispectral satellite image time series can be significantly improved through careful selection and optimization of parameters, offering a strong basis for refining methodologies in future soil monitoring efforts.
1. Introduction
To monitor soil health and combat degradation, scientists require accurate maps of soil reflectance. However, capturing "pure" bare soil from space is inherently difficult; satellite sensors are often obscured by green vegetation, dry crop residues (stubble), and varying soil moisture, all of which alter the spectral signature. While spectral indices are commonly used to mask vegetation, there is no consensus on the optimal way to combine multiple images over time (compositing) or how to set thresholds for these indices, especially across large, heterogeneous regions. This paper systematically evaluates these variables to establish a robust workflow for national-scale soil property estimation.
2. Methodology
The research team developed a workflow using Google Earth Engine to process Landsat 8 data (2015–2020) over Greece. The experimental design focused on four primary dimensions:
- Spectral Indices: Four indices were tested to identify bare soil: NDVI (green vegetation), NBR2 (crop residues), BSI (bare soil), and S2WI (soil moisture).
- Thresholding Strategies: The study compared Fixed Thresholding (applying a single value nationwide) against Dynamic Thresholding (calculating pixel-specific thresholds based on the 40th and 60th percentiles of the time series).
- Compositing Methods: Once bare soil pixels were identified, the time series was collapsed into a single map using Mean, Median, or selection based on index extrema (e.g., Min NDVI).
- Validation: Results were validated against the LUCAS 2015 database, which provided 641 ground-truth samples in Greece. Performance was measured by spectral correlation and the accuracy of Soil Organic Carbon (OC) predictions using Machine Learning algorithms like Kernel Ridge (KR) and Random Forest (RF).
3. Results and Discussion
The study revealed several critical insights into bare soil mapping. A "frequency filter" requiring a pixel to be identified as bare soil at least twice in the time series proved highly effective at removing transient noise. Furthermore, while NDVI is standard for vegetation removal, NBR2 was found to be essential for distinguishing soil from dry crop residues.
In terms of predictive power, dynamic thresholding strategies significantly outperformed fixed baselines. Specifically, the Kernel Ridge (KR) regressor using a composite based on the Dynamic Bare Soil Index (DyBSI) achieved an $R^2$ of 0.69. This represents a massive improvement over fixed-threshold configurations, which often yielded $R^2$ values near zero.
Visually, the Mean compositing method was preferred over the Median as it produced smoother maps with less "salt and pepper" noise. Most bare soil detections occurred in late spring during field preparation; winter detections were frequently skewed by high soil moisture, which lowers reflectance and degrades composite quality.
4. Limitations and Future Work
A primary limitation of the study is the inherent data imbalance within the LUCAS dataset for Greece. As shown in the frequency distributions, the samples are heavily skewed toward low Organic Carbon values (0–25 g/kg), making it difficult for models to accurately predict carbon-rich soils.
Additionally, the soil moisture index (S2WI) did not provide the expected performance boost, suggesting a need for more sensitive moisture detection. Future research should focus on incorporating higher-resolution data, such as Sentinel-2, and utilizing more balanced training datasets to transition from experimental composites to operational national monitoring systems.
5. Conclusion
This research demonstrates that optimizing spectral indices and employing dynamic thresholding can significantly enhance the accuracy of bare soil reflectance estimation. By refining these methodologies, it is possible to create high-quality soil maps that serve as a vital tool for monitoring soil health and addressing the challenges of soil degradation at a national scale.
Summary updated