Automated classification of ground deformation processes in Spain: a machine learning approach using a novel national InSAR-based database

InSAR has been widely employed in terrain deformation analysis worldwide. Its significant utility in risk management has led to the development of extensive SAR databases, poised for exploration in land-use planning studies. However, this information still requires specific expertise, hindering its accessibility for non-expert users. In this work, we introduce MOVESAR, an SAR-based database designed for training Machine Learning (ML) classification models capable of providing precise information on the type of deformation process and its cause. MOVESAR is also planned to support the development of deformation time series forecasting models.Each row in MOVESAR is spatiotemporally linked to a deformation time series (DTS) obtained through InSAR processing of SAR images from various satellites (SENTINEL 1, ENVISAT, ERS, COSMO-SkyMed, ALOS and TerraSAR-X), collected from previous studies conducted by the Geological Survey of Spain (IGME) and the Centre Tecnológic Telecomunicacions Catalunya (CTTC). Spatially, our database covers a substantial part of the Spanish territory, represented in 60 deformation polygons (with more than 300,000 measurement points or "MPs"), spanning from 1992 to 2020.Each column in MOVESAR represents a covariate potentially related to the six deformation processes compiled in this initial version of the database: piezometric change-induced deformation, landslide in mining environments, soil landslide, constructive subsidence, subsidence in mining environments, and subsidence in dumps. Covariates include geological, morphometric, hydrological, and geotechnical information, as well as data associated with DTS, land use, land cover, and landslide, subsidence and expansive clays susceptibility/hazard. Dynamic variables, including precipitation and DTS, underwent transformation into static variables by extracting statistical measures such as mean, standard deviation, range, and slope.In this study, we present preliminary results from nine ML models trained using MOVESAR: four single base models (nb, knn, lda, and lr), and five ensemble models (rf, gbc, xgboost, lightgbm, and catboost). We discuss the performance of the models and analyze the importance of covariates. Additionally, we evaluate the impact of applying techniques aimed at reducing noise, bias, and model complexity, such as threshold velocity filtering technique (TVF) for eliminating stables MPs, Recursive Feature Elimination (RFE) for covariate reduction, and Cost Sensitive Learning (CSL) for class balancing.Our future work aims to expand the number of covariates, MPs, and classes using the European Ground Motion Service (EGMS) to enrich MOVESAR, establishing it as a nationally valuable database for forthcoming studies on geohazard management. Additionally, we plan to apply spatiotemporal Deep Learning (DL) models incorporating dynamic variables, providing reliable classifications for decision-making in urban planning and national land-use management.This work has been developed thanks to the pre-doctoral grant for the Training of Research Personnel (PRE2021-100044) funded by MCIN/AEI/10.13039/501100011033 and by "FSE invests in your future" within the framework of the SARAI project "Towards a smart exploitation of land displacement data for the prevention and mitigation of geological-geotechnical risks" PID2020-116540RB-C22 funded by MCIN/AEI/10.13039/501100011033.