Observation-Driven Versus Machine-Learning Approaches for Land Subsidence Assessment in Arid Regions

Water scarcity, land subsidence, and desertification constitute major environmental challenges in arid and semi-arid regions worldwide, with profound impacts on ecosystems, agricultural productivity, infrastructure, and long-term sustainable development. In many of these regions, intensive groundwater extraction has become a dominant driver of land subsidence, exacerbating water insecurity and environmental degradation.
Over the past decades, multi-decadal satellite observations from remote sensing and gravity missions have played a crucial role in estimating groundwater storage changes and quantifying the extent and rates of land subsidence at both local and regional scales. More recently, machine-learning (ML) approaches have been increasingly applied to map and assess land-subsidence hazards using diverse geospatial, hydrological, and satellite-derived datasets. While these models offer promising new capabilities, their results can vary substantially depending on model design, input data, and training strategies, sometimes leading to conflicting or uncertain outcomes.
In this contribution, we first focus on Iran, where land subsidence and water scarcity have emerged as widespread and critical issues, currently affecting more than 260 of the country’s 429 counties. We present results from a multi-decadal satellite-based analysis of land subsidence and groundwater dynamics and systematically compare these observations with outputs from several published machine-learning models. This comparison highlights both consistencies and discrepancies between observation-driven assessments and data-driven predictive approaches.
We then extend the analysis to selected regions in Central Asia, including Uzbekistan and Afghanistan, where similar hydrogeological and socio-environmental pressures are present but data availability and monitoring capacities are more limited. Finally, we discuss the key challenges and opportunities associated with integrating remote-sensing observations and machine-learning models for land-subsidence assessment, with particular emphasis on data quality, model transferability, uncertainty quantification, and implications for regional-scale hazard monitoring and water-resources management.