Auditing observability bias in InSAR-derived subsidence exposure estimates with a reproducible geospatial workflow

Technical Summary: Auditing Observability Bias in InSAR-Derived Subsidence Exposure Estimates

Problem Statement
Public Interferometric Synthetic Aperture Radar (InSAR) products are increasingly utilized as primary inputs for land subsidence screening. However, a critical informatics gap exists in how these products are translated into exposure estimates (e.g., population or built-up land at risk). Standard workflows often treat missing or weakly observable radar support as "decision-neutral," effectively removing cells with low observability from exposure calculations before the accounting step. This practice silently alters the exposure denominator. When low-observability areas overlap with significant population or infrastructure, "visible-only" summaries underestimate the total exposure within the deformation-relevant domain, potentially leading to incomplete risk assessments. The paper identifies this not merely as missing data, but as a systematic bias where the valid-support mask of the radar product is applied prematurely, converting "unobserved" areas into "absent" exposure.

Methodology
The authors propose a reproducible geospatial workflow designed to audit this bias by decoupling radar observability, deformation relevance, and exposure accounting until the final step. The workflow operates on four inputs: a spatial analysis grid, a public InSAR observability layer, a deformation-screening layer, and exposure layers (population/built-up).

Key methodological components include:

1. Separation of Layers: The workflow maintains radar observability, deformation status, and exposure as distinct layers. It avoids clipping the deformation domain to valid radar support before exposure is calculated.
2. Frozen Deformation Screen: A "strong-deformation" screen is applied first to define the deformation-relevant domain. This screen is fixed before observability classes are assigned to prevent tuning the deformation definition based on exposure results.
3. Observability Masking: Valid-support information from the public InSAR product is converted into an observability fraction for each analysis cell. Cells are partitioned into "majority observable" and "not majority observable" (hidden) classes.
4. Hidden Exposure Estimation: The core audit calculates exposure within cells that pass the deformation screen but lack majority observability. Two estimation methods are used:
   • Area-weighted allocation: Distributes exposure based on the fraction of the cell classified as hidden.
   • Native-pixel allocation: Repeats the calculation at the native resolution of the exposure product (e.g., GHSL pixels) to test for grid-allocation artifacts.
5. Robustness and Bounded Evidence Protocol: The study employs a hierarchy of checks to ensure the results are not artifacts of specific thresholds or spatial partitions:
   • Threshold-surface analysis: Tests 120 combinations of deformation and observability cutoffs.
   • MAUP (Modifiable Areal Unit Problem) sensitivity: Varies reporting partitions and origins.
   • Monte Carlo propagation: Perturbs exposure-scale and threshold assumptions over 5,000 iterations.
   • Placebo randomization: Tests if hidden masks placed randomly would produce similar associations.
   • Decision-ranking analysis: Evaluates how the audit affects operational priority lists compared to visible-only lists.
   • External Evidence Ladder: Uses external datasets (e.g., DWR/TRE, Cyprus EGMS, Po River) as bounded controls to verify code paths and transferability without claiming dense site-level validation.

Key Results
The workflow was applied to the Chao Phraya urban delta as a lead case. The results demonstrate that a substantial portion of exposure lies within the deformation-relevant domain but is weakly supported by the public InSAR product:

• Hidden Exposure: The area-weighted estimator identified 3.77 million people and 404.77 km² of built-up land in the "not-majority-observable" portions of strong-screen cells. The native-pixel sensitivity yielded similar results (3.63 million people and 404.77 km²), confirming the finding is not an artifact of coarse grid allocation.
• Robustness: The hidden exposure remained material across the threshold surface (ranging from 1.33 to 5.36 million people) and Monte Carlo simulations (median 3.65 million, 95% CI 1.64–5.48 million).
• Operational Impact: Decision-ranking experiments showed that while the top 100–250 priority cells remained similar between visible-only and observability-aware lists, the lists diverged significantly as the portfolio expanded. At 2,000 cells, the observability-aware ranking added over 572,000 hidden people and 10.83 km² of built-up land, indicating the bias is most relevant for broader screening portfolios.
• Transferability: Supporting checks in the Po, Brantas, Rhone, and Rhine deltas, along with product-lineage tests using Japanese and Iranian InSAR products, confirmed the workflow's logic is portable, though regional evidence roles were strictly bounded (e.g., Po served as a transfer case with sparse geodetic context, not a dense validation).

Significance and Claims
The paper positions itself as a methods and workflow contribution to Earth-science informatics rather than a regional geotechnical diagnosis or a dense validation study. Its primary claims are:

1. Informatics Correction: The workflow exposes a semantic error in current exposure translation: treating product support masks as neutral filters rather than explicit variables. By retaining "hidden" exposure as a separate accounting class, the audit makes the incompleteness of public products visible.
2. Screening Utility: The method improves the honesty of public-product exposure translation. It does not claim that every hidden cell is actively subsiding, but rather that these areas represent exposure that requires monitoring or independent checking because they fall within the deformation-relevant domain yet lack strong radar support.
3. Reproducibility: The study provides a reproducible audit layer that records the exact computational semantics of layer integration. It allows downstream users to distinguish between "zero exposure" and "unobserved exposure."
4. Limitations: The authors explicitly state that the workflow does not replace local geotechnical diagnosis or dense independent validation. The hidden exposure class is an audit artifact indicating where the public product is insufficient, not a confirmed deformation map. The absence of dense independent validation for the Chao Phraya hidden-exposure estimate is acknowledged as a limitation, framing the work as an "observability-exposure audit" rather than a physical validation study.

In conclusion, the paper argues that for public InSAR products used in screening, valid support information must travel with the exposure estimate. Removing support before accounting changes the question from "how much exposure is in the deformation-relevant domain?" to "how much exposure is visible to this specific product?", a distinction the proposed workflow seeks to restore.