SAR-Based Soil Moisture Estimation with Machine Learning

An information-theoretic account of when soil-texture-specific calibration of SAR soil moisture retrieval helps and when it cannot, tested on 677 paired observations across five soil textures. Published in Remote Sensing Letters.

Period: 2024–2025
Where: Indian Institute of Science
Role: Research Associate
Methods: Random Forest, SAR Remote Sensing, Python/Scikit-learn

Result: Developed soil-specific Random Forest calibration achieving 34% accuracy improvement for sandy textures in SAR-based soil moisture retrieval.

Findings:

Analyzed 677 paired SAR-soil moisture observations across 5 soil textures
Discovered counterintuitive vegetation enhancement effect (r=0.743 vegetated vs r=0.380 bare soil)
Developed information proxy framework predicting calibration success
Published: Remote Sensing Letters 17(4), 416-428 (2026). doi:10.1080/2150704X.2026.2647288

Approach: Combined physics-based understanding with machine learning to solve operational remote sensing challenges.