Extended Abstract

Introduction

Land subsidence in alluvial plains cannot be interpreted for risk management solely from deformation rates because spatial priority also depends on exposed people and assets and on physical–hydrogeological susceptibility. The Marvdasht Plain in southwestern Iran is an agriculturally important alluvial plain that depends heavily on groundwater and contains rural settlements, transportation and energy infrastructure, and major archaeological sites. Previous studies have documented subsidence and hydrogeological stress in the region, but an integrated framework combining deformation hazard, exposure, susceptibility, component discrimination, and uncertainty has remained limited. This study developed a reproducible relative land-subsidence risk index (RI) by integrating ascending and descending Sentinel-1 LiCSBAS time series with spatial exposure and physical–hydrogeological susceptibility. The RI is intended for screening and monitoring prioritization, not for estimating structural-failure probability, economic loss, or absolute disaster risk.

Methodology

Sentinel-1 ascending and descending LiCSAR time-series products were processed using LiCSBAS. Annual deformation rates were estimated for valid pixels, and the series were divided at 2019.5 into P1 and P2. Vertical-equivalent rates were calculated under the assumption of predominantly vertical motion, and inter-period change was defined as Δv = vP2 − vP1. Agreement between ascending and descending products was assessed using Pearson correlation, bias, and RMSE as an inter-geometry consistency test rather than independent validation.

The model comprised subsidence hazard (H), exposure (E), and physical–hydrogeological susceptibility (V). Hazard was derived from aligned deformation-rate products and normalized to 0–1. Exposure was constructed from 19 layers grouped into settlements, transportation, lifeline infrastructure, and productive land uses, using distance-decay functions and robust scaling. Susceptibility combined lithology, a groundwater-abstraction pressure proxy based on smoothed well density, and fault proximity. Because complete engineering and social vulnerability data were unavailable, V was treated as environmental susceptibility.

Upper-level weights were obtained using a role-based virtual AHP framework. Consensus weights were 0.548551 for H, 0.293941 for E, and 0.157508 for V, giving RI = 0.548551H + 0.293941E + 0.157508V. The index was classified into five equal-width classes. Robustness was evaluated through deterministic weighting scenarios, leave-one-component-out tests, and 10,000 Monte Carlo simulations of AHP weight uncertainty. Villages, population, households, infrastructure, productive land uses, and selected heritage sites were screened against the final classes.

Results and Discussion

The final model covered 897.131 km², or 99.223% of the study boundary. Very-low, low, moderate, high, and very-high classes occupied 269.105, 223.920, 273.288, 128.352, and 2.465 km², respectively. Thus, 130.817 km², or 14.582% of the valid area, was classified as high or very high. These areas were concentrated mainly in the central and central-eastern plain, showing that high RI values resulted from the coincidence of subsidence hazard, exposed assets, and susceptible ground conditions rather than deformation magnitude alone. The moderate class covered 30.463% of the valid area.

Hazard contributed 86.60% of the calculated RI variance, compared with 6.99% for exposure and 6.41% for susceptibility. Although hazard dominated spatial variability, E and V still modified the location and ranking of priority areas.

Fourteen of the 73 villages with valid coverage were located in the high class. They contained 30,725 people and 9,308 households, representing 47.60% of the assessed rural population and 48.27% of rural households. High and very-high classes also coincided with 43.92% of valid railway length, 36.90% of power-transmission lines, 36.75% of freeways and highways, 36.40% of valid bridges, and 43.90% of industrial sites. These values indicate spatial co-location only and do not demonstrate observed damage or functional disruption.

The high-plus-very-high share remained close under the AHP-consensus, equal-weight, and initial 0.60–0.30–0.10 scenarios, at 14.582%, 14.879%, and 12.954%, respectively. In contrast, removing individual components produced a wider range of 5.604–26.097%, indicating that model structure influenced the result more strongly than limited changes among the principal weighting schemes. Monte Carlo analysis yielded a mean high-plus-very-high share of 15.387%, a median of 14.708%, and a 95% simulation interval of 11.182–22.372%. The mean dominant-class probability was 0.963, and 92.20% of valid pixels had a dominant-class probability of at least 0.80; 115.236 km² had at least an 0.80 probability of remaining in the high or very-high classes. These probabilities represent classification stability under weight uncertainty, not damage probability.

Temporal comparison showed high agreement between ascending and descending rates for P1 and P2, with r = 0.952 and 0.978, but lower agreement for Δv, with r = 0.462. The ascending track showed 19.83% intensified subsidence, 38.07% relative stability, and 42.10% reduced intensity; the descending track showed 0.13%, 90.53%, and 9.34%, respectively. Thus, the broad spatial pattern was more robust than the inferred inter-period change, and Δv should be treated as an exploratory temporal indicator rather than definitive vertical acceleration or independent validation. Heritage screening classified Tall-e Bakun as high, Toll-e Shoqa as moderate, and Persepolis as very low at representative points, while Naqsh-e Rostam, Naqsh-e Rajab, and Estakhr were outside valid coverage. NoData does not indicate low risk.

Conclusion

The framework provides a transparent basis for prioritizing subsidence monitoring in the Marvdasht Plain. High-priority areas emerge from the combined spatial influence of hazard, exposure, and physical–hydrogeological susceptibility rather than subsidence intensity alone. The concentration of nearly half of the assessed rural population and households in only 14 high-class villages also shows that management significance cannot be inferred from area percentage alone. Sensitivity and Monte Carlo analyses indicate that the main spatial pattern is relatively stable under reasonable changes in upper-level weights, while component removal produces larger changes, highlighting the importance of model structure. Nevertheless, RI remains a relative screening index. The absence of independent damage and control data prevented external ROC/AUC-type validation, and the model does not estimate structural failure, economic loss, or complete engineering and social vulnerability. Future work should incorporate independent damage inventories, groundwater-level and abstraction time series, GNSS and leveling observations, hydrogeomechanical modelling, infrastructure fragility information, and polygon-based heritage assessment.