ECDF-Stabilized Sentinel-1 InSAR Coherence for Verification-Oriented Heritage Monitoring: A TabICLv2-SHAP and Three-Horizon DEMATEL Architecture

Highlights

What are the main findings?

• Reliability-filtered Sentinel-1 InSAR coherence is transformed into a fixed-support ECDF-stabilized ordinal evidence register, with a four-state construction retained upstream and three observed labels represented in the retained TabICLv2 holdout evaluation.
• Leakage-controlled TabICLv2 learning, post hoc SHAP attribution, and three-horizon DEMATEL prioritization are combined with confusion-matrix support to preserve verification-oriented interpretation.

What are the implications of the main findings?

• Multi-epoch coherence evidence can support transparent heritage-monitoring priorities without being treated as direct field diagnosis.
• The proposed framework supports persistence-aware screening, narrow ROI selection, and subsequent verification in heritage-sensitive landscapes.

Abstract

Cultural-heritage monitoring increasingly requires evidence-to-decision architecture capable of moving beyond descriptive coherence maps toward analytically transparent and verification-oriented decision support. Within this evidential setting, the analysis presents a case-study-demonstrated architecture for multi-epoch Sentinel-1 interferometric synthetic aperture radar (InSAR) coherence analysis in a heritage-sensitive landscape, designed to transform heterogeneous pairwise outputs into an interpretable, cross-epoch comparable, sector-level evidence register. Quantitatively, the architecture is demonstrated on 12 short-baseline intervals and 1193 valid fixed-support points; it is organized around four linked operations: (i) reliability-filtered coherence stabilization through fixed-support sampling, temporal point-history construction, ECDF-based rank alignment, and ordinal recoding; (ii) separation of the continuous radar response from the upstream four-state ordinal monitoring-support register before model fitting; (iii) leakage-controlled TabICLv2 learning on RAW/stabilized radar-meteorological descriptors, excluding SHAP-derived, prediction-derived, probability-derived, feature-importance-derived, and global mean-coherence proxy variables; and (iv) post hoc SHAP attribution transferred to a three-horizon DEMATEL prioritization register. Under the retained TabICLv2 configuration, holdout evaluation showed strong ordinal separability: accuracy = 0.9799, balanced accuracy = 0.9798, macro-F1 = 0.9798, and Cohen’s kappa = 0.9699. To delimit the performance claim, the corresponding holdout confusion matrix contained three observed ordinal labels after filtering and split construction; support was balanced across State 0, State 1, and State 2. At the decision-transfer stage, the SHAP-informed DEMATEL synthesis distinguished three prioritization regimes: an immediate coherence-led H1 structure, a seasonal hydro-meteorological amplification regime in H2, and a strategic coherence-selective prioritization structure in H3. The study is framed as a methodological demonstration for persistence-aware, verification-oriented heritage monitoring, not as field validation, archeological confirmation, structural diagnosis, or causal physical inference.

1. Introduction

Cultural-heritage monitoring [1,2,3,4] increasingly requires remote-sensing processing architectures capable of moving beyond descriptive mapping toward transparent, verification-oriented decision support. In archeological and heritage-sensitive landscapes, the relevant problem is rarely limited to the detection of a single spatial anomaly; rather, it concerns the repeated observation of territorial sectors whose surface response, accessibility, and monitoring priority may vary across time [1,2,3,4]. Satellite remote sensing and geographic information systems therefore provide a scalable basis for heritage documentation, risk screening, and landscape-scale monitoring, especially where continuous field inspection is impracticable or where rural settings require recurrent evidence updates.

Within this field, SAR/InSAR heritage applications, time-series coherence analysis, and archeological screening provide the technical basis for repeated observation under constrained field-access conditions [5,6,7,8,9,10,11,12,13,14,15,16,17,18] because they provide repeated, weather-independent observations and can support temporal appraisal when optical interpretation is constrained by cloud cover, vegetation, illumination differences, or acquisition gaps. Sentinel-1 coherence is retained here because it expresses phase consistency between two acquisitions and can support temporal-stability and recurrent-decorrelation screening. At the same time, coherence is not a direct heritage-condition measurement: it is influenced by acquisition geometry, temporal baseline, surface scattering, vegetation dynamics, soil moisture, rainfall history, and terrain-related observability conditions. Consequently, a multi-epoch coherence archive requires explicit stabilization, quality control, and aggregation before it can be used as a decision-support register.

More narrowly, the unresolved methodological problem concerns the SAR-to-decision step: reliability-filtered pairwise coherence rasters must be converted into a reproducible monitoring-support structure before learning, attribution, or prioritization are introduced. Many remote-sensing workflows generate coherence maps, pairwise change indications, or visually interpretable raster products; fewer define a compact pathway that samples reliability-filtered coherence on a common admissible support, aligns heterogeneous pairwise distributions, and aggregates the stabilized evidence into sector-level descriptors suitable for learning and prioritization. In the absence of this intermediate evidence layer, later analytical stages risk appearing either as decorative additions to raster mapping or as decision claims insufficiently grounded in the radar evidence. A verification-oriented workflow must therefore preserve a strict analytical order: the radar evidence is first filtered, sampled, and stabilized; the resulting descriptors then define a transparent learning task; model-internal attribution is summarized without causal overclaim; and decision-analytic prioritization is applied only to already screened driver families.

Set against this SAR-to-decision gap, the present study demonstrates a compact evidence-to-decision architecture for multi-epoch Sentinel-1 InSAR coherence monitoring in a heritage-sensitive landscape. The purpose is to address the unresolved SAR-to-decision step between pairwise coherence rasters and verification-oriented sector prioritization, by making explicit the sequence by which radar evidence is filtered, rendered comparable, interpreted through a leakage-controlled learning task, and transferred into verification-oriented monitoring priorities. Methodologically, this distinction is important because the decision-support value of the workflow depends less on the isolated novelty of any single component than on the controlled order in which the components are connected: Figure 1 provides a compact workflow overview by summarizing the four-operation structure connecting fixed-support coherence sampling, ECDF-based ordinal stabilization, leakage-controlled TabICLv2 learning, and SHAP-informed DEMATEL prioritization.

In methodological terms, the contribution is organized around four linked operations. First, reliability-filtered Sentinel-1 coherence is sampled on a fixed admissible support to construct temporal point histories. Second, ECDF-based rank alignment and ordinal recoding separate the continuous radar response from the upstream four-state monitoring-support register before model fitting. Third, leakage-controlled TabICLv2 learning is performed on RAW/stabilized radar-meteorological descriptors, while SHAP-derived, prediction-derived, probability-derived, feature-importance-derived, and global mean-coherence proxy variables are excluded from the learner input space. Fourth, post hoc SHAP attribution is transferred into a three-horizon DEMATEL register for immediate, seasonal, and strategic monitoring prioritization. Retained learning performance is therefore reported separately through the observed holdout labels, confusion matrix, and per-class support counts, rather than being conflated with the upstream four-state evidence register.

2. Materials and Methods

For interpretive control, the analysis follows a fixed sequence: fixed-support Sentinel-1 coherence sampling, ECDF-based ordinal evidence-register construction, leakage-controlled TabICLv2 learning with post hoc SHAP interpretation, and SHAP-informed DEMATEL transfer. This order is retained because interpretation and prioritization are meaningful only after the radar-side evidence has been filtered, stabilized, and aggregated into a comparable descriptor space.

2.1. Study-Area Context, Sentinel-1 Archive, and Admissible Support

Within the present case-study frame, Figure 2 is introduced to define the geographical, topographic, heritage-administrative, and operational boundary used for fixed-support extraction [19,20].

Figure 2 fixes the spatial frame used for fixed-support extraction by locating Masseria Antalbo in southwestern Sicily, and relating the site to its topographic and altimetric context, and delineates the operational boundary used for extraction: this boundary defines the spatial domain within which coherence observations are sampled and later aggregated; it is a sampling frame, not independent archeological or structural validation.

Within this boundary, the fixed-support scheme retained 1193 valid points. For each valid point p and interferometric interval i, the elementary point-epoch evidence unit is expressed in Equation (1), and the corresponding 12-interval temporal point history is expressed in Equation (2).

e_p,i = (c_p,i, v_p,i, o_p,i),

(1)

H_p = {e_p,1, e_p,2, …, e_p,12},

(2)

In Equation (1), c_p,i; is the reliability-filtered coherence value, v_p,i marks valid-support membership, and o_p,i records the ordinal occurrence state used later in monitoring-support construction. Equation (2) shifts the analytical unit from a single coherence pixel to a repeated evidence record across the 12-interval archive. At this stage, the methodological aim is not condition inference, but stable cross-epoch observation support. At the pairwise-evidence level, Figure 3 summarizes the retained coherence archive, whose interpretation depends on Sentinel-1 mission design, TOPS acquisition geometry, SNAP-StaMPS processing, SAR interferometric processing, and coherence-decorrelation constraints [21,22,23,24,25].

In contrast to a diagnostic map, the panel acts as an archive-structuring device: it reports the P01–P12 short-baseline coherence intervals and shows how repeated raster observations are constrained within a common support before being translated into point histories and sector descriptors: this archive-level organization provides the fixed input support for ECDF-based four-state target construction in Section 2.2.

2.2. ECDF-Based Four-State Ordinal Target Construction

Once the common support had been established, comparability across interferometric pairs became the main constraint. Raw coherence values are not directly exchangeable across intervals because acquisition geometry, temporal baseline, surface scattering, vegetation state, soil moisture, rainfall history, and observability conditions may differ from pair to pair. For that reason, the target was not assigned from raw coherence alone. Instead, each interval was first stabilized through an empirical cumulative distribution function (ECDF) transform, so that the observation could be read relative to its own pairwise distribution. For mathematical transparency, Equations (3) and (4) define the ECDF transform and the corresponding aligned rank variable, following established quantile and empirical distribution-function conventions [16,26,27].

$${\widehat{F}}_i\left(c\right)={\displaystyle \frac{1}{n_i}}{\sum }_{p\in {\mathcal{P}}_i}\mathbf{1}\left(c_{p,i}\le c\right),$$

(3)

$$r_{p,i}={\widehat{F}}_i\left(c_{p,i}\right), r_{p,i}\in \left[0,1\right],$$

(4)

Within this rank-based frame, the aligned value defines a continuous response only after the direction of interpretation has been fixed: in the present workflow, larger response values denote stronger monitoring-support evidence; this orientation is set before ordinal encoding and before model fitting, preventing the target from being defined retrospectively from classifier outputs. Within this upstream layer, the ECDF-based construction also defines an upstream four-state ordinal monitoring-support register through three ordered thresholds, τ₁, τ₂, and τ₃: this register is therefore used to organize evidence before model fitting; it is not raw coherence, not a DEMATEL score, and not a field-validated heritage-condition class. In the retained TabICLv2 holdout evaluation, however, only three ordinal labels were represented after filtering and split construction. Therefore, the reported confusion matrix documents the observed holdout labels only, while the four-state register remains the upstream evidence-construction framework.

Table 1. Four-state ordinal monitoring-support target. The four states are ordered evidence categories derived after ECDF-based stabilization; they are not diagnostic heritage-condition classes.

Ordinal State	Condition	Interpretation Retained in the Manuscript
State 0	y ≤ τ1	lowest retained monitoring-support evidence
State 1	τ1 < y ≤ τ2	weak-to-moderate monitoring-support evidence
State 2	τ2 < y ≤ τ3	elevated monitoring-support evidence
State 3	y > τ3	highest monitoring-support evidence stratum

reports the operational interpretation of the four evidence states.

For each valid point, the encoded ordinal sequence records how monitoring-support evidence evolves across the 12 intervals, as summarized in Equation (5). This representation keeps three levels distinct: the reliability-filtered coherence value, the ECDF-aligned response, and the final ordinal target. That separation is central to the revision because it prevents model inputs, target definition, and later attribution from being conflated.

$$O_p=\left\{o_{p,1},\hspace{0.17em}{o}_{p,2},\hspace{0.17em}\dots ,\hspace{0.17em}{o}_{p,12}\right\},$$

(5)

2.3. Sector-Level Descriptor Construction

After four-state encoding, the analysis moves from point histories to sector-level descriptors. Such aggregation is deliberate: the learning and prioritization layers are not designed to classify isolated points, but to summarize operational monitoring sectors. The same archive of 1193 valid fixed-support points was then aggregated within these sectors, so that each sector retained a measurable support base before learning and decision transfer. Each sector is therefore described through a compact set of support and composition variables, including valid-point proportion, four-state composition, persistence of elevated states, temporal intermittence, highest-state share, and retained RAW/stabilized radar-meteorological descriptors while the sector-level interpretation was retained only where the main ordinal-composition and driver-priority patterns remained directionally stable. At descriptor level, Equation (6) reports the compact sector vector; Equation (7) then shows the resulting descriptor matrix obtained by stacking all sector vectors, with subsequent grouping and reduced-space interpretation read against standard clustering, silhouette, PCA, and biplot conventions [28,29,30,31]. Attribution-derived, prediction-derived, probability-derived, feature-importance-derived, and global shortcut variables are not included at this stage; the learning task therefore remains separated from the subsequent SHAP interpretation layer.

$${\mathbf{x}}_j=\left({\rho }_j,\hspace{0.17em}{f}_{j,0},\hspace{0.17em}{f}_{j,1},\hspace{0.17em}{f}_{j,2},\hspace{0.17em}{f}_{j,3},\hspace{0.17em}{P}_j,\hspace{0.17em}{I}_j,\hspace{0.17em}{A}_j,\hspace{0.17em}{\mathbf{m}}_j\right),$$

(6)

$$\mathbf{X}={\left[{\mathbf{x}}_1,\hspace{0.17em}{\mathbf{x}}_2,\hspace{0.17em}\dots ,\hspace{0.17em}{\mathbf{x}}_{N_s}\right]}^{\mathrm{\top }}\in {\mathbb{R}}^{N_s\times K},$$

(7)

where $N_s$ is the number of monitoring sectors and $K$ is the number of retained descriptors. This matrix constitutes the input space for the leakage-controlled TabICLv2 protocol described in the next section without reintroducing attribution, prediction, probability, feature-importance, or shortcut variables into the learner input space. The four-state composition terms are retained as upstream sector descriptors, whereas the reported holdout performance is computed only over the ordinal labels observed in the retained evaluation split.

2.4. Leakage-Controlled TabICLv2 and Post Hoc SHAP Protocol

Only after the sector-level descriptor matrix has been defined is the learning layer introduced. Recent SAR machine-learning literature shows the breadth of learning-based SAR interpretation; here, however, learning is used more narrowly to test ordinal separability within a stabilized heritage-monitoring descriptor matrix [32,33,34,35,36]. TabICLv2 is retained as the tabular learning component because the revised workflow requires a compact model operating on structured sector descriptors while keeping prediction separate from attribution. It is not claimed to be universally superior to conventional tabular classifiers; rather, it is used as a compact in-context tabular learner suitable for small structured sector-level matrices, with simpler comparators retained as baseline checks rather than as alternative decision architectures. The input space is restricted to RAW/stabilized descriptors. SHAP-derived, prediction-derived, probability-derived, feature-importance-derived, and global mean-coherence proxy variables are excluded before model fitting, reducing the risk of leakage and circular interpretation. Performance is summarized through accuracy, balanced accuracy, macro-F1, Cohen’s kappa, and cross-validation stability, so that agreement, class balance, and resampling bias are explicitly delimited [37,38,39]. In the retained holdout configuration, TabICLv2 reported accuracy = 0.9799, balanced accuracy = 0.9798, macro-F1 = 0.9798, and Cohen’s kappa = 0.9699. These values indicate strong ordinal separability within the stabilized descriptor matrix; they support the model task only and do not validate heritage condition in the field. After model fitting, SHAP is used only as a post hoc explanation layer for the fitted model-internal contribution structure.

2.5. SHAP-Informed DEMATEL Transfer Across H1–H3

At the final transfer step, SHAP-informed driver families are converted into a three-horizon DEMATEL prioritization register, implemented through the pyDEMATEL computational environment [40], which is not applied to raw raster products or to the full high-dimensional feature matrix. Rather, it is applied to a reduced, interpretable driver system derived from the stabilized sector matrix and the post hoc SHAP contribution structure; formally, Equations (8) and (9) report the total-relation construction and the prominence/relation summaries.

$${\mathbf{T}}^{\left(h\right)}={\mathbf{N}}^{\left(h\right)}{\left(\mathbf{I}-{\mathbf{N}}^{\left(h\right)}\right)}^{-1}, h\in \left\{H_1,H_2,H_3\right\},$$

(8)

$$R_i^{\left(h\right)}={\displaystyle \sum _{j=1}^m}{t}_{ij}^{\left(h\right)}, C_i^{\left(h\right)}={\displaystyle \sum _{j=1}^m}{t}_{ji}^{\left(h\right)}, P_i^{\left(h\right)}=R_i^{\left(h\right)}+C_i^{\left(h\right)}, Q_i^{\left(h\right)}=R_i^{\left(h\right)}-C_i^{\left(h\right)}$$

(9)

Three monitoring horizons are retained. H1 represents the immediate coherence-led screening horizon. H2 represents the seasonal horizon, where hydro-meteorological modulation is read together with persistence-sensitive descriptors. H3 represents the strategic horizon, oriented toward longer-term verification. In all three cases, DEMATEL outputs are interpreted as prioritization structures and narrow ROI candidates, not as causal mechanisms or field-confirmed diagnoses.

2.6. Sensitivity and Comparator Checks

For robustness control, two levels were retained. At the modelling level, the retained TabICLv2 run is reported through holdout metrics, five-fold cross-validation summaries, and the observed-label confusion evidence reported with the learning results. No unverified comparator is used as a performance claim. At the interpretation level, SHAP rankings and DEMATEL structures are evaluated for cross-horizon coherence. Any instability is treated as an interpretive limitation, not as evidence against the radar-side evidence construction.

Table 2. Compact robustness register. The table reports the retained TabICLv2 holdout performance and the directional aggregation/sensitivity reading used to delimit interpretation. Comparator metrics are not included because no verified baseline run is used as part of the main performance claim.

Configuration	Accuracy	Balanced Accuracy	Macro-F1	Cohen’s Kappa
TabICLv2 retained holdout	0.9799	0.9798	0.9798	0.9699
Aggregation/sensitivity reading	directional check	directional check	directional check	not applicable

condenses this robustness reading into the retained model metric row and the directional aggregation/sensitivity row.

3. Results

Six result blocks follow the operational sequence: cross-epoch coherence comparability, ECDF-stabilized ordinal evidence, sector-level evidence composition, leakage-controlled TabICLv2 performance with post hoc SHAP attribution, H1-H2-H3 DEMATEL prioritization, and a compact sensitivity check.

3.1. Cross-Epoch Coherence Comparability

At the first result level, the retained 12-pair Sentinel-1 archive remains spatially interpretable after reliability filtering and fixed-support restriction. From a comparability standpoint, the key result is not that all pairs share the same distribution; rather, the archive retains enough structured coherence evidence to justify rank-based stabilization before learning. Quantitatively, pairwise dispersion is not uniform: the highest variability is associated with P01 and P02, whereas P03 shows the lowest internal dispersion. These diagnostics, shown in Figure 4, show that cross-epoch comparability cannot be assumed from aligned maps alone and must be established through explicit distributions.

On this basis, the subsequent ordinal analysis is grounded in an observed need for stabilization rather than introduced as a purely formal transformation; distributional comparison is therefore treated as a support device, not as a standalone inferential endpoint [41,42]. For completeness, the extended pair-level and hydro-meteorological diagnostics are provided in Supplementary Figures S1 and S2 and Supplementary Table S1. These materials document cross-epoch variability, raw distributional structure, and ECDF alignment.

3.2. Four-State Ordinal Evidence and Sector-Level Composition

Once the pairwise distributions had been ECDF-aligned, the stabilized response was expressed as four ordered monitoring-support states. Across the archive, the state shares document how each interval contributes to the lower, intermediate, elevated, and highest evidence strata. Under this interpretation, upper-state evidence is a compositional signal requiring closer sector-level attention, not direct deterioration evidence.

At sector level, the pointwise ordinal histories were condensed into support and composition descriptors, including valid-point proportion, state frequencies, persistence of elevated states, temporal intermittence, and highest-state share. These descriptors provide the tabular bridge between radar-side evidence and the later learning layer, while preserving the distinction between evidence composition and diagnostic interpretation.

The ordinal-stabilization and threshold-adjacent diagnostics are reported in Supplementary Tables S2 and S3. These supporting materials detail the redistribution of reliability-filtered coherence evidence after ECDF-based rank alignment and should be read as target-construction documentation, not as heritage-condition confirmation.

3.3. Reduced-Space Organization of Stabilized Sector Descriptors

Beyond state composition alone, the stabilized sector descriptors also show a reduced-space organization. Within the reduced evidence space, the semantic contrast profile separates a stable lower-response group, a transitional/mixed group, and a persistent high-rank group. In parallel, the PCA loading structure indicates that coherence level and upper-rank persistence form the main reduced-space axis, whereas change count and mid-rank behaviour contribute to secondary variation. The result is not a final hotspot declaration, but a measurable organization of the stabilized evidence proxy before TabICLv2 and DEMATEL are introduced. Figure 5 provides the compact visual evidence for this result, combining ordinal support, semantic group contrasts, and variable-side reduced-space geometry.

This reduced-space result is retained in the main text because it explains why the learning layer operates on structured sector descriptors rather than on isolated raster values. Additional reduced-space and pointwise aggregate diagnostics are retained in the Supplementary Materials and cross-read with the pair-level summaries reported in Supplementary Figure S2 and Supplementary Table S2.

3.4. Leakage-Controlled TabICLv2 Performance and Post Hoc SHAP Attribution

At the learning stage, the retained RAW/stabilized descriptor matrix showed consistent ordinal separability within the bounded TabICLv2 task. For leakage control, model fitting excluded SHAP-derived, prediction-derived, probability-derived, feature-importance-derived, and global shortcut variables.

This restriction is essential for interpretation: the classifier operates on upstream sector evidence, while SHAP is applied only after fitting as a post hoc explanation layer; the retained model-task performance is first summarized in

Table 3. Leakage-controlled TabICLv2 performance summary. The outputs are interpreted as verification-oriented screening evidence, not as field-confirmed diagnosis or causal physical inference.

Metric	Value	Evaluation Layer	Reviewer-Safe Interpretation
Accuracy	0.9799	Stratified holdout	Ordinal separability only; not field validation
Balanced accuracy	0.9798	Stratified holdout	Class-balanced separability summary
Macro-F1	0.9798	Stratified holdout	Primary multiclass balance metric
Cohen’s kappa	0.9699	Stratified holdout	Agreement beyond chance within the model task
Macro-F1	0.9698 ± 0.0124	Five-fold CV	Stability check across folds
Cohen’s kappa	0.9547 ± 0.0185	Five-fold CV	Stability check across folds

through aggregate holdout and cross-validation metrics, so that the subsequent confusion-matrix evidence can be read against the same evaluation frame.

These aggregate values indicate strong ordinal separability within the retained model task. Because aggregate metrics alone do not show how errors are distributed across classes, the observed-label confusion structure is reported next.

At the observed-label level, the confusion matrix reported in

Table 4. Observed-label confusion matrix and per-class support for the retained TabICLv2 holdout evaluation. The upstream ECDF-based procedure defines a four-state ordinal evidence register; however, only three ordinal labels were represented in the retained holdout split after filtering and split construction, thus, the confusion matrix reports the observed holdout labels only.

True Class	Predicted State 0	Predicted State 1	Predicted State 2	Support
State 0	99	1	0	100
State 1	2	94	3	99
State 2	0	0	100	100

documents the support structure required to rule out a class-imbalance explanation of macro-F1.

Across the retained holdout split, observed labels are nearly balanced: State 0 = 100, State 1 = 99, and State 2 = 100. Numerically, the corresponding matrix produced accuracy = 0.9799, balanced accuracy = 0.9798, macro-F1 = 0.9798, and Cohen’s kappa = 0.9699. These values therefore indicate class-balanced ordinal separability within the observed-label model task, rather than a performance artefact driven by class imbalance. State 3 remains part of the upstream four-state evidence register, but it was not represented in the retained holdout split used for the reported performance metrics.

For visual inspection, the same observed-label confusion structure is plotted in Figure 6 as a compact heatmap, with the diagonal cells emphasizing correctly retained ordinal assignments.

Taken within this bounded evaluation frame, the tabular and visual confusion diagnostics support a restrained interpretation: the retained TabICLv2 result indicates class-balanced ordinal separability within the model task, not field validation, archeological confirmation, structural diagnosis, or causal physical inference. Only after this bounded performance evidence is established is the post hoc SHAP reading introduced.

Post hoc SHAP attribution was then used to summarize the fitted model-internal contribution structure. Under this constrained interpretation, SHAP values are treated as model-internal attribution summaries, not as causal, physical, archeological, or structural evidence. At the family level, the attribution pattern supports a joint radar-meteorological interpretation: stabilized coherence descriptors remain central, while rainfall, temperature, and soil-moisture modulation terms act as contextual contributors within the learned evidence space.

Extended post hoc SHAP diagnostics are provided in Supplementary Figures S3–S6. These plots summarize model-internal contribution structure after TabICLv2 fitting and are not used as causal, physical, archeological, or structural evidence.

3.5. H1-H2-H3 DEMATEL Prioritization

At the decision-transfer stage, SHAP-informed DEMATEL translates the post hoc attribution structure into a compact horizon-specific prioritization register; the driver system is reduced to eight operational families: MEAN_COH, COH_VARIABILITY, VERYHIGH_SHARE, LOW_COH_BURDEN, CHANGE_COUNT, RAIN_MODULATION, TEMP_MODULATION, and SOIL_MOISTURE_MODULATION. DEMATEL is therefore applied to a screened and interpretable driver system, not to raw raster products or to the full high-dimensional predictor matrix.

Across horizons, Figure 7 and

Table 5. Compact H1-H2-H3 DEMATEL decision register. P denotes prominence (D + R) and S denotes relation (D − R). Values are interpreted as structural positions within a SHAP-informed register, not as physically validated causal mechanisms.

Horizon	Result Regime	Key Driver	Prominence/Relation	Interpretive Role
H1 immediate	Coherence-led screening	MEAN_ COH	P = 4.056; S = +0.538	Highest positive relation-side position; immediate screening structure
H1 immediate	Coherence-led screening	VERYHIGH_ SHARE	P = 3.371; S = +0.101	Positive relation-side upper-evidence support
H2 seasonal	Maximum coupling	TEMP_ MODULATION	P = 5.488; S = +0.371	Highest overall prominence; seasonal amplification
H2 seasonal	Maximum coupling	MEAN_ COH	P = 5.247; S = −0.027	Highly central but nearly neutral/slightly receiver-side
H2 seasonal	Maximum coupling	SOIL_ MOISTURE_ MODULATION	P = 4.893; S = +0.064	Weakly cause-side seasonal context
H3 strategic	Coherence-selective prioritization	MEAN_ COH	P = 3.833; S = +0.602	Dominant positive relation-side strategic driver
H3	Coherence-selective prioritization strategic	VERYHIGH_ SHARE	P = 3.354; S = +0.405	Strategic stabilization of persistent high-response evidence
H3	Coherence-selective Prioritization strategic	TEMP_ MODULATION	P = 3.323; S = −0.218	Still prominent but reactive in the long horizon

summarize the decision-transfer result where H1 is coherence-led, H2 is the maximum seasonal coupling horizon, and H3 is a selective long-horizon configuration in which coherence-side descriptors recover positive relation-side positions within the decision-support register.

This horizon separation is the main decision-transfer result. H3 is not a delayed repetition of H2; it is a selective reorganization in which persistent coherence-side evidence becomes the primary basis for longer-term verification-oriented monitoring.

Beyond the compact main-text register, Supplementary Figures S7 and S8 report the extended H1-H2-H3 DEMATEL synthesis and the corresponding direct-relation and total-influence matrices. The supplementary board supports the compact decision-transfer register in the main text and should be interpreted as horizon-specific prioritization structure only.

3.6. Sensitivity and Comparator Check

At the final robustness stage, sensitivity is read at two levels: model-task stability and interpretation-boundary control. At this final robustness layer,

Table 6. Sensitivity and interpretation-boundary register. The checks preserve the distinction between radar evidence construction, learning, post hoc attribution, and decision-transfer prioritization.

CHECK	RETAINED EVIDENCE	FACING IMPLICATION
LEAKAGE CONTROL	Excluded SHAP-derived, prediction-derived, probability-derived, feature-importance-derived, and shortcut variables before fitting	Prevents circular interpretation
PERFORMANCE STABILITY	Holdout metrics supported by five-fold CV: macro-F1 = 0.9698 ± 0.0124; kappa = 0.9547 ± 0.0185	Supports ordinal separability within the task
ATTRIBUTION BOUNDARY	SHAP applied after model fitting only	Model-internal explanation, not causality
DECISION-TRANSFER BOUNDARY	DEMATEL applied to eight driver families across H1-H2-H3	Prioritization structure, not diagnosis

summarizes the boundary checks. At the modelling level, the retained TabICLv2 result is supported by aggregate metrics, five-fold cross-validation summaries, and the observed-label confusion evidence reported with the learning-performance results. No unverified baseline-comparator values are used as part of the main performance claim. At the interpretation level, SHAP rankings and DEMATEL structures were evaluated for cross-horizon coherence. Instability in a driver ranking was treated as an interpretive limitation, not as a failure of the radar-side evidence construction.

In a restrained synthesis, the results support a bounded evidence-to-decision claim: comparable ordinal evidence, separable sector descriptors, model-internal driver summaries, and horizon-specific monitoring priorities can be organized within one transparent workflow. It does not close the evidential chain as field validation, archeological confirmation, structural diagnosis, or causal physical inference. At reporting level, the retained main-text tables provide the compact performance, confusion-support, decision-transfer, and sensitivity registers; the Supplementary Materials preserves the extended diagnostics needed for transparency and reuse.

4. Discussion

Within this bounded case-study setting, the results are compatible with the working hypothesis that reliability-filtered multi-epoch Sentinel-1 coherence can be transformed from pairwise raster outputs into an interpretable evidence-to-decision structure for heritage-oriented screening [1,3,4,43,44,45,46,47,48]. In a restrained reading, the interpretation is bounded by 12 short-baseline intervals, 1193 valid fixed-support points, ECDF-based four-state stabilization, and leakage-controlled TabICLv2 evaluation; the contribution concerns the SAR-to-decision step between pairwise coherence rasters and verification-oriented sector prioritization [49,50,51,52,53,54,55,56].

At the learning level, the revised results indicate that the RAW/stabilized radar-meteorological feature space preserves strong ordinal separability once processed through TabICLv2 [57,58,59,60,61]. This interpretation is supported by the retained holdout statistics: accuracy = 0.9799, balanced accuracy = 0.9798, macro-F1 = 0.9798, and Cohen’s kappa = 0.9699. At the observed-label level, retained confusion diagnostics show balanced support across State 0, State 1, and State 2, confirming that the macro-F1 result is not driven by class imbalance alone. These values are interpreted as separability within the pre-defined ordinal learning task, not as field validation. TabICLv2 is therefore used as a leakage-controlled tabular learner embedded in the evidence chain, while attribution-derived values, predictions, probability outputs, feature-importance summaries, and prioritization scores remain outside the learner input space. This design reduces circularity between evidence construction, learning, attribution, and decision transfer [61].

At the attribution level, post hoc SHAP analysis shows that explanatory relevance is problem- and configuration-dependent rather than predefined by any single feature family [62,63,64,65,66,67]. In the present setting, SHAP is used to interpret the fitted TabICLv2 decision structure, not to pre-construct the input space and not to provide field validation. The attribution layer identifies which coherence descriptors and coherence-meteorological terms carry model-internal explanatory weight after ordinal learning. This directly addresses operation (Figure 1(3)): SHAP is downstream of prediction, whereas the learner input remains restricted to RAW/stabilized descriptors. In this model-internal reading, SAR coherence and environmental modulation jointly structure the fitted evidence space [11,68,69,70,71,72,73,74,75], while the continuous radar response remains separated from the upstream four-state ordinal monitoring-support register before learning and attribution are introduced.

At the decision-transfer level, SHAP-informed DEMATEL results further support the working hypothesis that preservation priority is horizon-dependent not only in magnitude but also in influence structure. At the immediate horizon, H1 remains primarily coherence-led: MEAN_COH occupies the strongest positive relation-side position in the DEMATEL register, with prominence P = 4.056 and relation S = +0.538. Within the H2 regime, hydro-meteorological amplification reaches its maximum, with temperature, rainfall, and soil-moisture modulation becoming more structurally active: within this coupling regime, TEMP_MODULATION reaches P = 5.488 and S = +0.371, while the largest total-influence paths include TEMP_MODULATION -> MEAN_COH = 0.42, TEMP_MODULATION -> SOIL_MOISTURE_MODULATION = 0.39, and TEMP_MODULATION -> VERYHIGH_SHARE = 0.37. In H3, this seasonal amplification contracts toward a more selective coherence-centred configuration, with MEAN_COH -> TEMP_MODULATION = 0.40, MEAN_COH -> SOIL_MOISTURE_MODULATION = 0.37, and MEAN_COH -> VERYHIGH_SHARE = 0.34. These cross-horizon shifts indicate a temporally conditioned prioritization structure within the modelled driver register, not a causal physical mechanism [49,50,51,52,53,54,55,56].

This architecture addresses the gap between geospatial evidence generation and operational heritage interpretation [1,3,4,48] by converting stabilized radar evidence into sector-level descriptors, model-internal attribution summaries, and horizon-specific prioritization registers. It should therefore be read as a heritage-sensitive case-study benchmark developed under the environmental, archeological, and acquisition conditions of Masseria Antalbo, not as a cross-site performance claim already validated across multiple landscapes [1,3,4,48,76,77,78,79].

Consistent with these limitations, the present study should be framed as a methodological demonstration whose broader transferability remains to be validated [13,78] across different sites, surface conditions, and SAR acquisition settings. First, the results remain tied to the present 12-pair archive size, target formulation, and feature space [50,59,71,72,73,74,75,80,81]. Second, the hydro-meteorological block is introduced through temporally aggregated pair-level descriptors [82,83,84,85,86,87,88,89] and may underrepresent shorter-lag environmental effects. Third, ROI and sector products remain screening-oriented outputs [76,77,78,79,88,90,91] and should be treated as support for verification and prioritization rather than as field-confirmed diagnosis.

Future validation should proceed through a denser and more seasonally balanced archive, alternative target definitions, lag-sensitive environmental descriptors [26,27,82,83,84,85,86,87,88,89,92], independent field or expert conservation assessment [79,90,91,93], and transferability tests across different heritage-sensitive landscapes.

5. Conclusions

Overall, the present study demonstrates, within a bounded case-study setting, how reliability-filtered multi-epoch Sentinel-1 coherence can be transformed from pairwise raster outputs into cross-epoch comparable sector evidence for heritage-oriented screening and prioritization [1,3,4,8,9,10,11,12,13,14,15,48,78,80,81,94]. Quantitatively, the demonstration is bounded by 12 short-baseline intervals, 1193 valid fixed-support points, an upstream four-state ordinal evidence register, and a retained TabICLv2 holdout evaluation with three observed ordinal labels. In a conservative interpretation, the archive remains case-bounded and does not provide temporally dense validation of persistence, recurrence, or long-term hotspot stability [69,70,92,95,96,97,98,99,100]. In a restrained reading, the contribution is a reliability-first SAR evidence-to-decision architecture, not a claim of definitive deformation retrieval, field-confirmed heritage diagnosis, structural diagnosis, or causal physical inference [69,70,92,95,96,97,98,99,100].

At the operational level, the architecture establishes a reproducible baseline from which a denser and more seasonally balanced interferometric sequence can be expanded. The present temporal spacing remains suitable for first-pass monitoring logic [88,101], but not for long-term recurrence validation. Its value lies in documenting a traceable chain from terrain-corrected coherence evidence to admissible support control, fixed-support extraction, ordinal stabilization, and sector-level evidence construction, rather than in providing a final deformation verdict [88,95,96,97,98,99,100,101,102].

From a heritage-monitoring perspective, the main benefit is that descriptive coherence outputs are converted into cross-epoch comparable sector indicators capable of supporting transparent verification priorities [48,76,77,78,79,88,90,91]. The framework structures recurrent departures, persistence-sensitive changes, low-observability conditions, and radar-meteorological modulation within one interpretable evidence chain.

At the pixel and ROI scale, caution remains essential. Isolated high-response pixels [69,70,103] are not stand-alone anomalies; they become meaningful only when supported by spatial consistency, temporal recurrence, and geometry-aware quality screening [11,21,68,80,81,94,95,103,104,105,106]. In this respect, ROI-level products should be read as pre-validation spatial syntheses that narrow inspection space and guide targeted verification, not as field-confirmed diagnostic maps.

Methodologically, TabICLv2, post hoc SHAP, and DEMATEL are retained because they preserve separation among learning, attribution, and decision transfer. TabICLv2 provides the leakage-controlled ordinal learning layer; SHAP summarizes fitted model-internal contributions after prediction; and DEMATEL organizes aggregated driver families into horizon-specific prioritization structures [49,50,51,52,53,54,55,56,57,58,59,62,63,64,65,66,67]. Operation (Figure 1(3)) is supported by the retained holdout metrics: accuracy = 0.9799, balanced accuracy = 0.9798, macro-F1 = 0.9798, and Cohen’s kappa = 0.9699. The corresponding confusion matrix and per-class support counts are reported with the learning results and document that the retained macro-F1 is not driven by class imbalance alone. Operation (Figure 1(4)) is supported by the H1–H2–H3 separation: H1 remains coherence-led, H2 is hydro-meteorologically amplified, and H3 is strategically coherence-selective. These are prioritization regimes, not validated causal mechanisms.

More broadly, the novelty of the study lies less in the introduction of a single unprecedented algorithm than in the operationalization of a complete evidence-to-decision architecture [1,3,4,48,49,50,51,52,53,54,55,56,57,58,59,62,63,64,78,94]. At methodological level, the framework demonstrates the feasibility of linking reliability-filtered multi-epoch Sentinel-1 coherence, ECDF-stabilized sector evidence, leakage-controlled TabICLv2 ordinal learning, post hoc SHAP attribution, and SHAP-informed DEMATEL prioritization within one continuous chain, while remaining methodologically distinct from earlier site-protection, coherence-gated InSAR, and DEMATEL-based planning applications [107,108,109]. It should therefore be understood as a verification-oriented SAR evidence architecture, not as a validated field-diagnostic system [1,3,4,48,49,50,51,52,53,54,55,56,57,58,59,62,63,64,65,66,67,76,77,78,79,88,93,94,101,102,103]. Future work should now test whether the same architecture remains stable under longer archives, alternative landscapes, different acquisition geometries, and independently verified conservation evidence.