Irrigation-driven land subsidence on the Barind Tract, northwest Bangladesh: two-track Sentinel-1 InSAR, GNSS-tied time series and spatially cross-validated explainable machine learning

The Barind Tract of northwest Bangladesh is the country’s most intensively groundwater-irrigated region, yet its ground motion has been measured only at isolated points. We processedascendinganddescendingSentinel-1small-baseline InSAR networks for 2017–2025, referenced them to a continuous GNSS station and converted line-of-sight rates to vertical velocity. In the ascending track, co herent land subsides at a median of −10.1 mmyr−1 (median formal uncertainty 0.58 mm yr−1, plus a common ±1.7 mmyr−1 GNSS-tie uncertainty), and 51% of it subsides faster than 10 mm yr−1. Rates reach −11.2 ± 1.2 mmyr−1 at Godagari on the Barind margin, against −1.8 ± 1.1 mmyr−1 on the active Ganges flood plain; 1-km block medians from the two tracks correlate at 𝑟 = 0.56. GLDAS-2.2 groundwater storage declined by 7.2±1.2 mmyr−1, and the fitted seasonal ground low lags the April storage minimum by three to four months, consistent with de layed aquitard drainage. An ERA5 tropospheric correction changed velocities by a median of 0.03 mm yr−1 and shifted this timing by at most three weeks at three of four sites. Randomforest andXGBoostmodelsexplain37–47%ofthespatialvariance under 5-km spatial block cross-validation; random cross-validation inflates 𝑅2 by 0.09–0.13, and omitting ramp removal raises it to 0.69. SHAP effects with the same direction in both tracks identify elevation, built-up density and local Boro irrigation intensity as robust controls. Subsidence removes at least 25 million m3 of ground volume per year, of the same order as the groundwater-storage decline, and601km ofmainroadsand101km of railway lieongroundsubsidingfasterthan 10 mm yr−1. Spatial cross-validation and cross-track consistency should precede physical interpretation of machine-learning attributions of subsidence.