Unraveling the Spatial Heterogeneity of Land Subsidence in the Yellow River Delta: A Spatially Adaptive Ensemble Learning Approach

The primary subsidence measurements were derived from 83 scenes of Sentinel-1A Single Look Complex (SLC) SAR images acquired between March 2017 and June 2023, processed using the Small BAseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technique. This processing yielded high-precision annual subsidence rates at 30 m spatial resolution, with accuracy validated against leveling survey data (R² = 0.877, RMSE = 17.24 mm/yr).To characterize the driving factors of subsidence, nine predictive variables were compiled from multiple sources: (1) topographic factors (elevation, slope, and sine/cosine-transformed aspect) derived from the ALOS Digital Elevation Model; (2) lithological classes obtained from regional geological maps, encoded via one-hot transformation; (3) anthropogenic activity indicators, including Euclidean distance to salt pans and binary salt pan mask (vectorized from 1:310,000 scale maps) and cropland proportion extracted from 2022 land cover data; and (4) hydrometeorological variables comprising current-month and one-month lagged precipitation, serving as proxies for groundwater recharge due to limited in-situ well observations.All predictor layers were resampled to 30 m resolution to maintain consistency with InSAR subsidence data.

Spatial features were enhanced through nonlinear transformations: inverse transformation of salt pan distance to model decay effects, and square root transformation of local topographic variance (computed within 3×3 moving windows) to characterize micro-topographic influences. The complete dataset comprises approximately 1.7 million sample points covering the eastern Yellow River Delta, providing comprehensive spatial coverage of both subsidence measurements and potential controlling factors.