AI model predicts terrain clutter for satellite ground stations with 60% less error

Siting satellite ground stations, especially for low Earth orbit (LEO) networks, depends on knowing local terrain obstructions that affect radio signals. The current standard, ITU-R P.452-18, uses fixed clutter heights per land-use class—a coarse approach that misses within-class variation, leading to conservative exclusion zones and poor site ranking. A team of eight researchers from academia and industry (Shohini Sarkar et al.) addresses this with an explainable geospatial AI framework that predicts representative clutter height (RCH) using open data.

The model, based on LightGBM, was trained on LiDAR-derived RCH labels (75th percentile clutter height) from the U.S. Geological Survey’s 3D Elevation Program and features from global land-cover, terrain, demographic, thermal, and optical remote sensing datasets. It achieves a mean absolute error of 1.79 meters and R²=0.765, a 60% reduction in absolute error over the ITU baseline. SHAP feature attribution shows tree canopy cover, land-cover semantics, and spectral reflectance as the strongest predictors. The framework was validated internationally on land-cover matched data, and ablation studies confirm robustness outside forested areas. Accepted at IEEE CASE 2026, this work opens the door to scalable, interpretable clutter modeling for better ground station placement and spectrum coordination.