Lee et al. (2026) Class-Specific Interpretation and Validation of Optical and SAR Sensor Contributions in Multi-Sensor Land-Cover Classification Using SHAP and Leave-One-Sensor-Out Ablation

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-09-16
• Authors: Jeonghee Lee, Kwangseob Kim, Kiwon Lee
• DOI: 10.3390/rs18183186

Research Groups

• Department of Computer Science, University of California, Los Angeles (UCLA)
• Google Earth Engine team

Short Summary

This study compares the performance of four machine learning classifiers on a harmonized satellite image dataset and highlights conditions under which feature-level attribution and sensor-level necessity diverge. The results show that strong within-sensor collinearity can lead to attribution dilution across certain bands.

Objective

• To evaluate and compare the performance of Random Forest, Classification and Regression Trees, Support Vector Machine, and Gradient Tree Boosting classifiers on a harmonized satellite image dataset.

Study Configuration

• Spatial Scale: Global (using Google Earth Engine)
• Temporal Scale: Not specified

Methodology and Data

• Models used: Random Forest, Classification and Regression Trees, Support Vector Machine, and Gradient Tree Boosting
• Data sources: KOMPSAT-3/5 and Sentinel-1/2 satellite images harmonized to a common 2.8 m grid within Google Earth Engine

Main Results

• The study found that Random Forest and Gradient Tree Boosting reached overall accuracies of 92.4% and 92.8%, respectively.
• Class-specific SHAP was cross-interpreted against permutation importance, feature-correlation analysis, and a leave-one-sensor-out ablation, with SHAP and permutation-importance rankings in broad agreement (Spearman ρ = 0.72–0.92).
• The results highlighted two conditions under which feature-level attribution and sensor-level necessity diverge: strong within-sensor collinearity and class-conditional information concentration.

Contributions

• This study characterizes the conditions under which feature-level attribution and sensor-level necessity diverge, providing original insights into the behavior of machine learning classifiers on satellite image datasets.

Funding

• Not specified

Citation

@article{Lee2026ClassSpecific,
  author = {Lee, Jeonghee and Kim, Kwangseob and Lee, Kiwon},
  title = {Class-Specific Interpretation and Validation of Optical and SAR Sensor Contributions in Multi-Sensor Land-Cover Classification Using SHAP and Leave-One-Sensor-Out Ablation},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18183186},
  url = {https://doi.org/10.3390/rs18183186}
}