Lithological Mapping in High-Vegetation Areas Using Sentinel-2, Sentinel-1, and Digital Elevation Models

Yansi Chen, Genyuan Liu, Zhihong Song, Ming Li, Minhua Wang, Shuang WangSensors 2025 · DOI

TL;DR

This paper tackles the challenge of mapping rock types (lithology) in densely vegetated areas, which are notoriously difficult to analyze with remote sensing. By fusing data from Sentinel-2 (optical), Sentinel-1 (radar), and Digital Elevation Models (DEM), and employing the Random Forest algorithm, they achieved significantly improved lithological classification accuracy, demonstrating the power of multi-source data integration.

What Problem Does This Paper Address?

Accurately identifying rock types (lithology) is fundamental for geological research, mineral exploration, and understanding Earth's surface processes. However, in regions with dense vegetation cover, traditional remote sensing methods struggle because the vegetation obscures the underlying rock formations. This makes it difficult to extract spectral and structural information directly from the surface. While combining different remote sensing data sources has shown promise, a comprehensive approach integrating optical, radar, and topographic data for lithological mapping in these challenging, high-vegetation environments has remained relatively unexplored.

Methods

The core of this study lies in the application of the Random Forest (RF) algorithm for lithological classification. RF is an ensemble learning method that builds multiple decision trees and combines their predictions, offering high accuracy, robustness to overfitting, and the ability to handle high-dimensional data without extensive preprocessing.

The researchers explored various combinations of data sources:

  • Sentinel-2 (S2): Optical multispectral imagery providing spectral information.
  • Sentinel-1 (S1): Radar imagery, capable of penetrating clouds and vegetation to capture surface roughness and dielectric properties.
  • Digital Elevation Models (DEM): Topographic data providing information on elevation, slope, and aspect.

They systematically evaluated the performance of RF using individual data sources and several fused combinations (S2+S1, S2+DEM, S1+DEM, and S2+S1+DEM). A key aspect of their methodology was hyperparameter optimization using the Out-of-Bag (OOB) error to find the best settings for the RF model for each data combination. This involved tuning parameters like the number of trees and the fraction of data sampled per tree.

Datasets

The study utilized the following datasets:

  • Sentinel-2 L2A imagery: Optical multispectral data from June 2023, specifically using bands B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12.
  • Sentinel-1 imagery: Radar data from June 2023, using VV and VH polarization bands.
  • Shuttle Radar Topography Mission (SRTM) DEM: Providing 30m resolution topographic data.
  • Geologically modified map: A refined lithological classification map created by professional geologists, serving as the ground truth for validation. This map categorized 11 lithological types, plus a 'Water Body' class.

The study area was Taiwan Province, China, chosen for its complex geology and high vegetation coverage.

Key Insights and Ideas

  • Data Fusion is Crucial for High-Vegetation Areas: The most significant takeaway is that combining multiple data sources (optical, radar, and topographic) is far more effective than using any single source for lithological mapping in densely vegetated regions.
  • Sentinel-2 and DEM are a Powerful Pair: The combination of Sentinel-2 (spectral information) and DEM (topographic information) emerged as the optimal fusion for this study, significantly outperforming other combinations. This highlights how spectral signatures, even when partially obscured by vegetation, can be effectively complemented by terrain features.
  • Topography Matters: Elevation, slope, and other topographic features derived from DEM play a substantial role in distinguishing lithologies, especially when vegetation interference is high. This is because different rock types often have distinct topographic expressions.
  • Feature Importance Varies: Not all features contribute equally. While spectral bands (like B2) and vegetation indices (like GCVI) were important, radar data (S1) and some topographic features (like SR) showed less significance in this specific study, suggesting that their utility can be context-dependent.
  • Random Forest is Well-Suited: The RF algorithm proved to be a robust and effective choice, handling the high dimensionality of the fused data and providing reliable classification without requiring extensive feature selection or complex preprocessing.

Key Results

  • The optimal data combination of Sentinel-2 and DEM achieved the highest validation Overall Accuracy (OA) of 84.30% and a Kappa coefficient of 0.83.
  • When using Sentinel-2 alone, the model achieved a validation OA of 81.90% and a Kappa of 0.80.
  • Using DEM alone resulted in a significantly lower validation OA of 57.22% and a Kappa of 0.54.
  • The combination of Sentinel-1, Sentinel-2, and DEM also performed well, with a validation OA of 82.78% and a Kappa of 0.81.
  • Feature importance analysis showed that spectral features like B2 and NDBI, along with elevation, were among the most influential variables for classification.

Limitations and Future Work

  • Limited Environmental Factors: The current model primarily incorporates spectral and terrain variables. It does not account for other environmental factors like hydrology, land surface temperature, or detailed geological structures, which could further influence lithological differentiation.
  • RF vs. Deep Learning: While RF is effective, its ability to capture complex and nonlinear spatial patterns might be limited compared to more advanced deep learning methods like Convolutional Neural Networks (CNNs) or transformer-based models.
  • Future Directions: The authors suggest future research should investigate the impact of different vegetation coverage types and degrees on lithology classification. They also propose integrating hydrological, climatic, and structural geological information. Furthermore, exploring more sophisticated algorithms like CNNs or transformer models is recommended to enhance classification performance in geologically complex regions.

Summary updated

Abstract

The extraction of lithological information in areas with high vegetation coverage presents numerous challenges, particularly in identifying concealed lithological features. This study focuses on a typical high-vegetation coverage area in Taiwan Province, China, utilizing multi-source data from Sentinel-2, Sentinel-1, and DEM, and using the Random Forest algorithm for lithological mapping. The results demonstrate that the optimal combination of Sentinel-2 and DEM significantly enhances the classification performance, achieving an overall accuracy (OA) of 84.30% and a Kappa coefficient of 0.83 in the validation set. Geological conditions have specific limiting effects on ecosystems, as spectral features (such as B2 and NDBI) and topographic features (such as elevation) contribute significantly to the classification results. This study provides valuable reference information for lithological information extraction in areas with high vegetation coverage.
BibTeX
@inproceedings{chen2025lithological,
  author = {Yansi Chen and Genyuan Liu and Zhihong Song and Ming Li and Minhua Wang and Shuang Wang},
  title = {{Lithological Mapping in High-Vegetation Areas Using Sentinel-2, Sentinel-1, and Digital Elevation Models}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25072136},
}
Figure