Urban Expansion and Satellite–Administrative Divergence in Mongolia: Multi-Sensor Evidence from Ulaanbaatar and Pastoral Provinces

Abstract

Urban monitoring in rapidly growing dryland cities is difficult because satellite signals and administrative records capture different aspects of settlement change and can perform differently across land-cover contexts. Mongolia provides a useful case because urban growth is concentrated in Ulaanbaatar while pastoral provinces remain exposed to ecological and livelihood shocks. We integrate MODIS NDVI, CHIRPS-derived SPI_12, a harmonized DMSP-OLS/VIIRS nighttime-light series, and administrative livestock and migration data for 22 spatial units over 2000–2024 (N = 550). A planned Landsat BuiltUp component was tested against Dynamic World for 2016–2024 (N = 39,600). The frozen NDBI > 0 and NDVI < 0.2 rule failed transfer validation (balanced accuracy = 0.474, Kappa = −0.052, F1 = 0.408; bare-to-built false-positive rate = 0.949), and enriched spectral models retained material province-specific errors. No repaired BuiltUp classifier was therefore frozen, and the original UEI, Mismatch, Ulaanbaatar M3, and spatial UEI analyses are not retained as principal evidence. In the retained origin-side panel, lower lagged NDVI anomaly is associated with greater livestock loss (β = −0.0148, p = 0.016), dzud severity with greater loss (β = 0.00454, p = 0.034), and higher prior-year livestock loss with higher standardized registered out-migration pressure (β = 1.113, p < 0.001). Nighttime-light harmonization shows strong level agreement but weaker rank continuity around the sensor transition. The study therefore advances a validation-first multi-source monitoring framework that withdraws unsupported urban composites after a failed measurement gate.

1. Introduction

Climate-linked migration in pastoral drylands is an important research and policy issue, yet the links among ecological stress, livelihood disruption, and urban expansion remain insufficiently understood, especially where official statistics may miss peri-urban growth or delayed migrant registration [1,2,3]. Projections across six world regions suggest that up to 216 million people could move internally by 2050 under climate-related pressures, with drylands among the highly exposed settings [4,5]. We measure precipitation deficits with the 12-month Standardized Precipitation Index (SPI_12) and official dzud severity classifications, and ecological stress with MODIS NDVI anomalies reflecting realized pasture conditions shaped by climate, vegetation response, grazing pressure, and land use.

Mongolia is a strong empirical setting because about 83% of its territory is rangeland, rural livelihoods remain closely tied to mobile pastoralism, and urban growth is concentrated in Ulaanbaatar. Comparable climate–mobility dynamics occur across drylands with heterogeneous exposure and adaptive capacity [6]. The 2009–2010 dzud killed more than 10.3 million animals, about 20% of the national herd, and the 2023–2024 dzud again caused large-scale livestock mortality [7,8,9]. Over the same period, Ulaanbaatar’s population rose from 794,730 in 2000 to 1,768,151 in 2024 [10]. Many migrants settle in peri-urban ger districts, named for portable circular, felt-covered dwellings traditionally used by Mongolian pastoral households and situated among incrementally serviced plots and low-density settlement [11,12].

This study addresses three gaps. First, Mongolia-focused research on dzud, pastoral vulnerability, migration, and urban change remains largely separated across empirical strands rather than integrated in a long-run subnational design [1,7,13,14,15]. Second, satellite and administrative indicators capture different settlement and population constructs, so any divergence measure depends on valid satellite components [14,16,17,18,19,20,21,22]. Third, multi-sensor integration can create false confidence when a composite is retained after one component fails independent validation. The revised design therefore places measurement validation before downstream urban-composite, destination-side, and spatial inference.

The study uses a balanced panel of 21 aimags (first-level administrative provinces of Mongolia) and Ulaanbaatar from 2000 to 2024 (N = 550), combining ecological, livelihood, migration, and satellite indicators. Its retained remote-sensing contribution is a validation-first monitoring architecture: MODIS vegetation conditions and a locked DMSP/VIIRS nighttime-light series are retained, while Landsat BuiltUp serves as a methodological validation diagnostic. No replacement urban composite is imposed after the physical BuiltUp component fails transfer validation.

The original design specified a three-component Urban Expansion Index (UEI) and a Mismatch Index comparing it with administrative Urban_Share. Because the BuiltUp component failed transfer validation, the final analysis reports neither UEI nor Mismatch as empirical findings and does not substitute a post hoc two-component index. Satellite–administrative divergence is therefore treated as a measurement problem rather than inferred from an invalid composite.

The retained empirical design is modular. M1, M2, and M5 provide origin-side evidence on ecological conditions, livestock loss, and registered provincial out-migration pressure; nighttime lights remain a descriptive satellite signal with a cross-sensor warning. Prespecified M3 and M4 are not retained because their outcomes depended on the failed UEI/BuiltUp architecture. These modules are not interpreted as a single origin-to-destination causal chain.

2. Literature Review and Theoretical Framework

2.1. Pastoral Mobility Pressure and Urban Expansion Pathways

Environmental stress can shape migration, but mobility responses depend on household resources, institutions, livelihood systems, and adaptive capacity [1,2]. Cattaneo et al. [23] emphasize that stress may increase migration for some households while constraining it for others, a concern especially relevant in drylands [3,5]. Pastoral drylands combine high exposure with uneven adaptive capacity [6], and weather shocks can reduce livestock productivity [24]. In livestock-dependent settings, such livelihood disruption can increase migration pressure when asset losses weaken household resilience [7,13].

Mongolia illustrates the relevance of this origin-side livelihood context because rural mobility has long been shaped by dzud events, livestock losses, and changing pastoral livelihoods. Sternberg [7] documents the consequences of the 2009–2010 dzud, while Fernández-Giménez et al. [13,25] examine dzud-related social and ecological dynamics and rangeland co-management. Mongolia-specific evidence also shows that drought frequency, duration, and severity affect pasture production [26]. Together, this literature supports treating vegetation condition, winter hazard exposure, livestock loss, and mobility pressure as related but analytically distinct elements of pastoral vulnerability.

2.2. Remote Sensing of Urban Expansion and Satellite–Administrative Divergence

Remote sensing supports monitoring of urban land cover and settlement change. WSF 2015 and WSF 3D provide global built-up information at 10–30 m resolution [14,19], while Dynamic World v1 provides 10 m land-cover probabilities for recent years [27]. Mongolia-specific studies examine Ulaanbaatar land suitability and ger-district detection [11,12], and Landsat enables multi-decadal settlement monitoring [20]. Together, these sources motivate combining physical built-up mapping with other satellite indicators while keeping ecological, administrative, and settlement measures conceptually distinct.

Nighttime lights are widely used as proxies for settlement intensity and urban economic activity [21,22], but long-run analysis requires harmonization because DMSP OLS and VIIRS differ in resolution, radiometry, saturation, and sensor design. Prior studies use invariant-region calibration for intra-DMSP drift [28] and cross-sensor rescaling for DMSP–VIIRS continuity [29,30]. MODIS NDVI provides a complementary measure of ecological conditions in drylands and Mongolian grasslands [15,31,32]. These signals therefore offer distinct, not interchangeable, information for multi-sensor monitoring.

Satellite-based urbanization measures need not coincide with administrative statistics. Prior work documents gaps between remotely sensed settlement indicators and official classifications in data-constrained settings [16,17,18], reflecting differences in physical form, light intensity, population classification, temporary activity, and administrative boundaries. This study therefore treats satellite–administrative divergence as a measurement question. Because the planned BuiltUp-dependent composite failed transfer validation, no final Mismatch Index is used to infer informality, undercount, or unregistered migration.

2.3. Spatial Urban Dynamics, Panel Econometrics, and Mediation-Oriented Analysis

Spatial models are appropriate only when the dependent variable has an adequate measurement foundation. Anselin [33], LeSage and Pace [34], and Elhorst [35] provide the spatial-econometric framework originally specified for the urban-composite branch, while Driscoll and Kraay [36] motivate robust covariance estimation in the retained panel models. Weather-shock studies illustrate stronger identification under plausibly exogenous variation [37,38], and applied work shows the relevance of regional dependence [39]. Because UEI ultimately depended on a BuiltUp component that failed transfer validation, the final study does not report the UEI-based spatial model or alternative weight-matrix re-estimation.

Mediation analysis can assess whether an association is compatible with an intermediate pathway, but causal interpretation requires strong assumptions such as sequential ignorability [40,41]. Those assumptions cannot be verified in this observational provincial panel. Moreover, the retained lag structure estimates NDVI_(t − 1) to livestock loss_t and livestock loss_(t − 1) to out-migration_t in separate equations. M5 is therefore a mediation-oriented product diagnostic, not an exact causal decomposition or mediated share.

The conceptual framework combines push–pull, asset-based livelihoods, climate-mobility, and urban economic geography. Lee [42] links origin-side stressors to migration pressure, while Carter and Barrett [43] show how loss of productive assets can constrain recovery; in pastoral systems, livestock mortality is the relevant asset shock [7,13]. Black et al. [44] emphasize that climate stress can produce both mobility and immobility depending on resources and institutions. Urban economic geography separately motivates persistence in settlement and infrastructure concentration [45].

These perspectives motivate a modular rather than causal-chain design. Push–pull and asset-based theories support testing whether poorer ecological conditions are associated with greater livelihood loss and whether loss is associated with registered out-migration [42,43], while the broader mobility literature allows heterogeneous responses [1,44]. Urban economic geography remains relevant to the original destination-side question [45], but the final manuscript does not estimate a BuiltUp-dependent urban-composite pathway after that measurement component failed validation.

3. Materials and Methods

3.1. Study Area, Panel Structure, and Hypotheses

The study covers Mongolia’s 21 aimags and Ulaanbaatar from 2000 to 2024, forming a balanced panel of 22 spatial units with 25 annual observations per unit (N = 550). Satellite indicators are aggregated to the official administrative boundaries obtained from the Mongolian Administrative Boundary Information System (https://khil.gazar.gov.mn) [46]. Mongolia’s extensive rangelands, climate-sensitive rural livelihoods, and concentration of urban growth in Ulaanbaatar make the provincial scale suitable for examining ecological stress, livestock loss, migration pressure, and urban monitoring. Figure 1 shows the study area and analytical units. Dzud refers to a severe winter disaster involving extreme cold, deep snow cover, and pasture shortage, often following summer drought and causing large-scale livestock mortality.

Samples differ by empirical role. M1 and M2 use the 21 aimags and exclude Ulaanbaatar; after the one-year lag, N = 504. M5 uses the same origin-side sample. Prespecified M3 and M4 retain their model numbers for traceability but are not retained because their urban outcomes depended on the failed BuiltUp/UEI branch.

Table 1. Model-specific sample restrictions and empirical role.

Model	Spatial Sample	Observations	Empirical Role
M1: Ecological stress and livestock loss	21 aimags, Ulaanbaatar excluded	504	Tests whether lower NDVI anomaly, greater precipitation deficit, and higher dzud severity are associated with livestock mortality
M2: Livestock loss and out-migration pressure	21 aimags, Ulaanbaatar excluded	504	Tests whether lagged livestock loss is associated with subsequent out-migration pressure
M3: Ulaanbaatar persistence model	Ulaanbaatar only; prespecified, not retained	Not reported	Prespecified H3 module; not evaluated because its primary urban outcome depended on the failed BuiltUp/UEI branch
M4: Spatial dependence model	All 22 spatial units; prespecified, not retained	Not estimated	Prespecified M4 spatial module; not re-estimated because the UEI dependent variable failed the input-validity gate
M5: Mediation-oriented product diagnostic	21 aimags as primary sample; all 22 units as robustness check	504/528	Examines whether the two origin-side associations are compatible with livestock loss as an intermediate livelihood variable; no exact causal mediated share is claimed

summarizes these roles, and Figure 2 shows the retained pathway and measurement gate. Provincial Out_Migration records registered departures but not destinations.

Based on Section 2, four hypotheses were prespecified while keeping origin-side, destination-side, and spatial/measurement questions separate. H1, H2, and H4 remain evaluable under the retained origin-side architecture. H3 is preserved verbatim for traceability but is not evaluated in the final Results because its primary satellite urbanization outcome depended on the BuiltUp/UEI branch that failed the measurement gate.

H1. 

Lower NDVI anomaly, lower SPI values indicating greater precipitation deficit, and higher dzud severity are expected to be associated with higher provincial livestock mortality.

H2. 

Higher lagged livestock loss is expected to be associated with greater subsequent out-migration pressure.

H3. 

Ulaanbaatar’s satellite-derived urban expansion is expected to exhibit temporal persistence. The direct annual association between registered in-migration and the Urban Expansion Index is examined as exploratory, given the limited statistical power of the available city-level time series.

H4. 

The pair of origin-side associations is expected to be consistent with livestock loss acting as an intermediate livelihood variable between NDVI-based ecological conditions and registered out-migration pressure; no exact causal mediated share is hypothesized.

3.2. Data Sources and Key Variables

This study integrates satellite indicators, administrative socioeconomic and migration variables, and climate-hazard records. Satellite indicators were processed in Google Earth Engine [47] and aggregated to provincial boundaries. Socioeconomic and migration data come from the National Statistics Office of Mongolia [10], dzud severity from the National Agency for Meteorology and Environmental Monitoring [48], and CHIRPS v2.0 precipitation is used to calculate SPI [49].

A key distinction is made between NDVI_anom and ndvi_mean. Canonical NDVI_anom is computed separately within each province over 2000–2024 as (ndvi_mean_it − province mean_i) divided by the province-specific population standard deviation. It is therefore a within-province standardized ecological-condition measure used in the origin-side models. By contrast, ndvi_mean is the unstandardized growing-season mean retained as a descriptive vegetation measure; it is not used in a final UEI because that composite was withdrawn after BuiltUp validation. NDVI_anom is a proxy for realized ecological conditions rather than an exogenous climate shock. It can reflect precipitation, vegetation response, grazing pressure, pasture management, and land-use dynamics. SPI_12 and Dzud_Severity are complementary hazard indicators that are less directly determined by local land use. Out_Migration is interpreted narrowly as registered provincial out-migration pressure: the records do not identify whether departures were directed to Ulaanbaatar or elsewhere. BuiltUp is retained only as a validation diagnostic, while UEI and Mismatch are classified as non-retained planned constructs.

Table 2. Key variables, data sources, and analytical roles.

Variable	Measurement/Source	Role in Analysis
NDVI_anom	MOD13A3 v6.1; June–August ndvi_mean standardized within each province over 2000–2024 using the province-specific population SD	Within-province ecological-condition proxy in M1, M2, and M5
ndvi_mean	MODIS MOD13A3 v6.1; absolute growing season NDVI mean	Absolute vegetation measure; not used in a final UEI
SPI_12	CHIRPS-derived 12-month Standardized Precipitation Index	Precipitation deficit proxy; lower values indicate drier conditions
Dzud_Severity	NAMEM official dzud severity classification, 0 to 3	Winter hazard indicator capturing snow, cold, and pasture shortage
LivestockLoss_rate	Abnormal livestock mortality in year t divided by total livestock in year t − 1	Livelihood disruption ratio; dependent variable in M1 and origin-side intermediate variable in M5
Out_Migration	Registered provincial out-migrants from NSO records; pooled z-standardized over all 550 observations for M2/M5; destination not identified	Origin-side registered migration pressure; dependent variable in M2
In_Migration_UB	Registered migrants to Ulaanbaatar from NSO records	Prespecified M3 variable; not used in final inferential results
NTL	Canonical DMSP-OLS/VIIRS nighttime lights on a common log scale: mapped DMSP 2000–2012 and observed log1p(VIIRS) 2013–2024; no smoothing	Descriptive satellite proxy for settlement/activity intensity; fixed pooled z-score used only for longitudinal descriptive comparison and not in a final composite index.
Built-Up	Historical Landsat NDBI–NDVI threshold rule; audited against Dynamic World 2016–2024	Methodological validation diagnostic only; failed transfer gate
UEI	Original planned composite of NTL, BuiltUp, and inverted ndvi_mean	Not retained after BuiltUp validation failure
Urban_Share	Administrative urban population share from NSO	Administrative urbanization measure; retained descriptively, not used in a final Mismatch Index
Mismatch Index	Original planned difference between standardized UEI and Urban_Share	Not computed/reported in final evidence
RealGDP_pc	Provincial real GDP per capita from NSO	Economic control variable
Road_Density	Road length divided by provincial area	Infrastructure control variable
Mining_Province	Binary indicator for major mining-intensive provinces	Structural control for mining-related differences

summarizes the key variables, sources, and analytical roles.

This variable structure separates the retained empirical domains from the failed measurement branch. SPI_12, Dzud_Severity, and NDVI_anom describe precipitation conditions, winter hazard exposure, and realized ecological conditions. Livestock loss and registered out-migration describe the origin-side livelihood and mobility domain. Canonical nighttime lights provide a descriptive settlement/activity signal subject to a cross-sensor warning. Landsat BuiltUp, the planned UEI, the Mismatch Index, M3, and M4 are not used as final inferential evidence after the BuiltUp transfer-validation failure.

3.3. Satellite Processing and Measurement Validation

This subsection documents the satellite and climate processing retained in the final study and the validation audit that determined the empirical scope. All spatial aggregation uses the same GADM v4.1 ADM1 frame. NDVI/SPI and the canonical nighttime-light architecture are retained. The Landsat BuiltUp rule is documented because its external validation failure is a methodological finding; the planned BuiltUp-dependent composite indices are not retained. NDVI anomaly and SPI: Growing season vegetation conditions were measured using MODIS MOD13A3 v6.1 NDVI [50]. Monthly June to August composites were filtered using the Pixel Reliability layer. Pixels classified as Good or Marginal were retained, while Snow or Cloud pixels were excluded. Province-level NDVI was calculated by averaging valid pixels within each provincial polygon. The NDVI anomaly was calculated as:

$$NDV{I}_{{anom}_{i,t}}={\displaystyle \frac{{NDVI}_{i,t}-{\mu }_i}{{\sigma }_i}}$$

where ${NDVI}_{i,t}$ is the growing season NDVI for province i in year $t$, and ${\mu }_i$and ${\sigma }_i$are the province-specific means and standard deviations for 2000–2024. This transformation removes long-run baseline differences across provinces and captures within-province deviations in vegetation. Because NDVI reflects both climate variability and long-run baseline differences across provinces and captures within-province use pressure, it is interpreted as a realized ecological stress indicator rather than a purely exogenous climate shock. Precipitation deficits were measured using CHIRPS v2.0. The 12-month Standardized Precipitation Index was estimated as:

$$SPI\_{12}_{i,t}={\varphi }^{-1}\left[G\right(P_{i,t};{\alpha }_i,{\beta }_i\left)\right]$$

where $P_{i,t}$ is 12-month accumulated precipitation, G (⋅) is the province-specific gamma cumulative distribution function with parameters α_i and ${\beta }_i$ and ${\mathsf{\varphi }}^{-1}(\cdot )$ is the inverse standard normal cumulative distribution function. Following standard SPI interpretation, SPI_12 < −1 indicates moderate drought, SPI_12 < −1.5 indicates severe drought, and SPI_12 < −2 indicates extreme drought. SPI_3 and SPI_6 are used in robustness checks.

Night-time light harmonization: The canonical nighttime-light series uses DMSP-OLS Version 4 stable_lights and VIIRS VCMCFG under a frozen two-stage calibration. DMSP images are first inter-calibrated to the F18-2013 reference using the fixed-Sicily quadratic procedure; raw DN = 0 and DN = 63 are retained at their bounds, while DN 1–62 are transformed and bounded to [0, 63]. Calibrated same-year DMSP composites are then merged. Province-level DMSP values are mapped to the VIIRS log scale using the 2013 cross-sensor relation log1p(VIIRS_i,2013) = α + β log1p(DMSPcal_i,2013) + ε_i, estimated with an intercept across the 22 spatial units.

The canonical 2000–2012 branch is α + β × log1p(inter-calibrated DMSP), while 2013–2024 uses the observed annual province-level log1p(VIIRS) branch. VIIRS annual values are formed from cf_cvg-weighted monthly radiance, with negative radiance set to zero before annual aggregation. No additional temporal-light threshold is imposed. No temporal smoothing or three-year rolling mean is applied. The locked pooled NTL reference is used only for fixed longitudinal standardization and descriptive comparison; it is not combined with BuiltUp in a final composite index.

The locked pooled NTL reference has a mean of 0.2673 and a sample standard deviation of 0.2124 (N = 550; ddof = 1). This affine standardization preserves the ordering of the canonical NTL series but does not remove the material cross-sensor measurement warning: the 2013 calibration shows strong linear agreement and materially weaker rank agreement, while the 2012–2013 transition shows a larger rank discontinuity. These diagnostics are reported transparently as measurement limitations rather than used to justify ex-post recalibration.

Built-up classification and validation: The historical mapping branch draws on the multi-decadal Landsat archive [51]. The historical rule classified a pixel as built when NDBI > 0 and NDVI < 0.2, following the NDBI formulation of Zha et al. [52]. Rather than treating that rule as validated by assumption, the revision audited it against Dynamic World v1 [27] using the actual 2016–2024 fixed-sample development/diagnostic set (N = 39,600; 4400 observations per year; 22 ADM1 units each year). Dynamic World served only as the reference/diagnostic label and was never a model predictor.

$$\mathrm{N}\mathrm{D}\mathrm{B}\mathrm{I}={\displaystyle \frac{SWIR-NIR}{SWIR+NIR}}$$

Pixels were classified as built-up where

$$NDBI>0andNDVI<0.2$$

External validation determined the downstream scope. The frozen rule showed poor discrimination and severe bare-to-built error, while enriched spectral candidates improved average performance but retained material class-specific and province-transfer failures. Candidate specifications and cutoffs were fixed before outer evaluation; Dynamic World served only as reference information, and no repaired classifier was frozen. No Random Forest or open-ended black-box search was used as a substitute for transfer validation [53], and accuracy diagnostics follow standard remote-sensing assessment principles [54]. Detailed validation results are reported in Section 4.1.3 and Tables S9–S11.

Composite-index scope decision: The original design combined nighttime lights, BuiltUp, and inverted vegetation intensity in a UEI and compared that composite with Urban_Share through a Mismatch Index. Because BuiltUp did not pass transfer validation, the three-component UEI was not scientifically admissible. A two-component replacement was also not introduced post hoc because it would constitute a different construct without a prespecified validation target. Consequently, UEI and Mismatch are not reported as final empirical results, and no BuiltUp-dependent M3, Moran, SAR/SEM/SDM, or alternative-spatial-weight result is retained.

$${UEI}_{i,t}={\displaystyle \frac{z\left({NTL}_{i,t}\right)+z\left({BuiltUp}_{i,t}\right)+[1-z({ndvi\_mean}_{i,t}\left)\right]}{3}}$$

3.4. Empirical Strategy

The empirical strategy separates retained origin-side modules from non-retained urban-composite modules. M1 and M2 estimate associations among ecological conditions, livestock loss, and registered provincial out-migration; M5 assesses whether the two separately timed associations are compatible with an intermediate livelihood role. M3 and M4 retain their prespecified numbers for traceability but are not estimated as final evidence because their BuiltUp/UEI-dependent outcomes failed the input-validity gate.

Core push side models (M1 and M2): M1 tests whether ecological stress and climate hazards are associated with livestock mortality. The baseline model is a two-way fixed effects specification with Driscoll and Kraay standard errors:

$${LivestockLoss\_rate}_{i,t}={\beta }_0+{\beta }_{1}L{1}_{NDV{I}_{{anom}_{i,t}}}+{\beta }_{2}SP{I}_{{12}_{i,t-1}}+{\beta }_{3}Dzud\_{Severity}_{i,t-1}+X_{i,t}\delta +{\mu }_i+{\mathsf{\lambda }}_t+{\epsilon }_{i,t}$$

Here, LivestockLoss_ratio_it is abnormal livestock mortality in year t divided by livestock stock in year t − 1. L1_NDVI_anom is the one-year lagged within-province standardized ecological-condition measure. The specification also includes lagged SPI_12 and Dzud_Severity together with time-varying controls such as livestock per capita and real GDP per capita. Province and year fixed effects absorb time-invariant provincial factors and national year shocks; time-invariant covariates therefore do not supply separately identified within coefficients. Driscoll–Kraay inference uses the locked two-lag Bartlett-kernel specification. The main sample contains 21 aimags from 2001 to 2024 (N = 504).

M2 tests whether lagged livestock loss is associated with subsequent out-migration pressure:

$${Out\_Migration}_{i,t}={\mathsf{\gamma }}_0+{\mathsf{\gamma }}_{1}L{1}_{LivestockLos{s}_{{rate}_{i,t}}}+{\mathsf{\gamma }}_{2}SP{I}_{{12}_{i,t-1}}+{\mathsf{\gamma }}_{3}NDVI\_{anom}_{i,t-1}+X_{i,t}\delta +{\mu }_i+{\mathsf{\lambda }}_t+{\epsilon }_{i,t}$$

where the dependent variable is a pooled standardized version of registered provincial Out_Migration. The standardization is fixed over the full 550-observation panel using the historical mean 2216.078 and population standard deviation 1918.477; it is not a per-1000-population rate. The coefficient on lagged livestock-loss ratio therefore describes the association between prior-year proportional livestock mortality and standardized registered out-migration pressure, conditional on the controls and fixed effects. The migration records do not identify destination.

Ulaanbaatar persistence model (M3), prespecified but not retained: M3 was originally intended to examine persistence in a satellite-derived urban expansion outcome and to treat same-year registered in-migration as exploratory. Because the primary UEI depended on the failed BuiltUp component, the final manuscript does not report the historical M3 estimates. This decision also avoids drawing low-powered inference from approximately 22–24 annual city observations

$${DV}_{UB,t}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1{\mathrm{D}\mathrm{V}}_{UB,t-1}+{\mathsf{\alpha }}_2{\mathrm{I}\mathrm{n}\_\mathrm{M}\mathrm{i}\mathrm{g}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}_{UB,t}+X_t\delta +{\epsilon }_t$$

Spatial dependence model (M4), prespecified but not retained: M4 was originally intended to test spatial structure in the UEI. The dependent variable failed the input-validity gate because its BuiltUp component did not pass transfer validation. The historical Moran/SDM results and alternative-W extensions are therefore removed from principal evidence rather than re-estimated on an invalid outcome. Spatial econometric analysis can be reconsidered after a defensible historical urban outcome is reconstructed.

$$UE{I}_{z_{i,t}}=\rho WUE{I}_{z_{i,t}}+{\mathsf{\beta }}_{1}L{1}_{LivestockLos{s}_{{rate}_{i,t}}}+{\mathsf{\beta }}_{2}L1\_NDVI\_{anom}_{i,t}{+X}_{i,t}\beta +{\mu }_i+{\mathsf{\lambda }}_t+{\epsilon }_{i,t}$$

Mediation-oriented product diagnostic (M5): M5 evaluates the origin-side pair of associations involving NDVI-based ecological conditions, livestock loss, and registered out-migration pressure. Because the retained equations use different lag positions, the product a × b is reported as a diagnostic of compatibility with an intermediate livelihood role rather than as an identified causal mediation effect or exact mediated proportion.

$${LivestockLoss\_rate}_{i,t}={\alpha }_a+a\cdot \mathrm{L}1\_\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}\_{anom}_{i,t}+X_{i,t}{\delta }_a+u_{i,t}$$

$${Out\_Migration}_{i,t}={\alpha }_b+b\cdot L1\_LivestockLoss\_{rate}_{i,t}+\mathrm{c}\prime \cdot \mathrm{L}1\_NDVI\_{anom}_{i,t}+X_{i,t}{\delta }_b+v_{i,t}$$

The product diagnostic is calculated as a × b. Its sampling uncertainty is summarized with the Sobel statistic and an aimag-cluster bootstrap.

An exact identity between a × b and c − c′ is not imposed because path a uses NDVI_(t − 1) to livestock loss_t, whereas path b uses livestock loss_(t − 1) to out-migration_t. Consequently, an exact mediated share is not reported under this lag structure.

The primary M5 sample contains 21 aimags (N = 504). Path a links lagged NDVI_anom to contemporaneous livestock loss, and path b links lagged livestock loss to standardized registered out-migration. Their product and cluster-bootstrap interval summarize compatibility with an intermediate livelihood role; they do not establish sequential ignorability, destination-specific migration, or a causal mediated proportion. M1, M2, and M5 thus constitute the retained origin-side evidence, while M3/M4 remain non-retained prespecified analyses.

3.5. Identification Risks and Robustness Checks

The main identification risks are reverse causation, residual time-varying confounding, and measurement error. Lagging ecological and livelihood variables reduces contemporaneous feedback but does not create exogeneity. NDVI_anom may respond to precipitation, grazing intensity, land management, and changes in livestock pressure, so it is treated as a realized ecological-condition proxy rather than an instrumental variable or randomized shock. SPI_12 and Dzud_Severity provide complementary hazard measures, but none of these observational specifications identifies a causal effect. The measurement audit adds a separate rule: a downstream analysis is retained only when its satellite input passes its own validation gate.

Residual time-varying confounding can affect both the magnitude and direction of the origin-side estimates. Disaster assistance or non-agricultural employment could weaken a positive loss–migration association, whereas remoteness, poor market access, adverse labor demand, or insecure pasture-use conditions could strengthen it. Province and year fixed effects absorb time-invariant geography and common national shocks but not local time-varying factors. Administrative migration records may miss temporary or delayed registration, and satellite measurement is separately constrained by the NTL cross-sensor rank warning and the failed BuiltUp transfer gate.

Robustness checks are interpreted only when they use the same locked variable definitions and pass the same input-validity gate as the corresponding main specification. The origin-side sensitivity analyses therefore remain on the canonical NDVI, livestock-loss-ratio, and standardized out-migration scales. The BuiltUp/UEI/Mismatch/M3/spatial robustness chain is not retained because its measurement foundation failed validation. This is a scope reduction, not a reinterpretation of the historical outputs.

4. Results

4.1. Descriptive Patterns and Satellite Measurement Diagnostics

4.1.1. Descriptive Statistics

The panel shows substantial variation in livelihood and migration outcomes across provinces and years, while the retained satellite variables represent distinct measurement domains.

Table 3. Descriptive statistics of key variables, 2000–2024.

Variable	Mean	Std. Dev.	Min	Max	Source
LivestockLoss, thousand animals	103.5	201.5	0.0	1825.0	NSO
Out_Migration, persons	2216.1	1920.2	199.0	16,820.0	NSO
In_Migration, persons	2223.6	5986.8	1.0	41,592.0	NSO
Net_Migration, persons	7.6	4877.3	−8136.0	40,246.0	NSO
SPI_12	0.000	0.915	−2.030	3.340	CHIRPS
NDVI_anom, ecological stress proxy	0.000	0.017	−0.067	0.059	MODIS
ndvi_mean, UEI component	0.178	0.074	0.058	0.366	MODIS
Dzud_Severity, 0–3	1.069	0.999	0.000	3.000	NAMEM
NTL, Sum-of-Lights	0.446	1.020	0.001	7.188	DMSP-VIIRS

Notes: N = 550. NDVI_anom is used in the retained ecological and migration analyses. ndvi_mean is retained as a descriptive vegetation measure and is not used in a final UEI. BuiltUp is retained only as a validation diagnostic; the planned UEI was withdrawn after BuiltUp failed transfer validation.

provides the full distributions: SPI_12 is centered near zero by construction, NDVI_anom captures within-province vegetation deviations, and canonical nighttime lights describe settlement/activity intensity. BuiltUp, UEI, and Mismatch are not summarized as final panel outcomes because the BuiltUp measurement gate failed.

4.1.2. Nighttime-Light Measurement Diagnostics

The main nighttime-light result is a contrast between strong cross-sensor level agreement and weaker rank continuity. In the 2013 overlap, DMSP-to-VIIRS calibration has R² = 0.973 and Spearman = 0.639; rank continuity falls to 0.555 across the 2012–2013 transition. This supports descriptive longitudinal-level use but cautions against assuming seamless province ranking across sensors. Full diagnostics are reported in Table S6 and Figure S2. The performance of the frozen and enriched spectral BuiltUp candidates is summarized in Figure 3.

4.1.3. BuiltUp Validation Gate

The historical BuiltUp rule did not provide a defensible measure of physical urban extent. Its most consequential error was classifying 94.9% of Dynamic World bare reference observations as built, despite weak overall discrimination. Enriched spectral models improved average performance but traded one error for another: the strongest aggregate benchmark retained substantial crop commission, while the two-stage candidate reduced crop and grass false positives at the cost of lower built recall and persistent bare-soil error. Figure 4 shows these class-specific trade-offs.

The decisive weakness is spatial transfer rather than temporal stability. Performance remained comparatively stable across held-out years but deteriorated materially for held-out provinces, with Govisümber the worst case. The same pattern persisted under stricter Dynamic World confidence screens, so low-confidence reference labels alone do not explain the failure.

No spectral-only candidate therefore met the combined requirements for class-specific error control, useful recall, and stable province transfer, and no repaired classifier was frozen. A defensible historical redesign would require local spatial context, multi-year persistence, cross-sensor portability, and a prospectively specified independent holdout; these features are future requirements, not completed results.

4.2. Ecological Stress, Livestock Loss, and Out-Migration Pressure

The retained origin-side analysis uses 21 aimags (N = 504, 2001–2024) and separates ecological conditions, livestock loss, and subsequent registered out-migration pressure. Ulaanbaatar is excluded because this module concerns pastoral origins rather than destination-specific urban expansion. H1, H2, and H4 are evaluated as observational associations.

4.2.1. NDVI-Based Ecological Stress and Livestock Mortality

The M1 results support a temporal association between poorer prior-year vegetation conditions and greater livestock mortality (

Table 4. M1: NDVI anomaly and livestock mortality.

Variable	Spec. 1: Lagged Baseline	Spec. 2: + SPI_12 and Dzud_Severity	Spec. 3: Contemporaneous
L1_NDVI_anom	−0.01481 ** (0.00617)	−0.01370 ** (0.00621)
NDVI_anom, current year			0.00329 (0.00529)
L1_SPI_12/L1_Dzud_Severity		−0.00784 (0.00730)/0.00454 ** (0.00215)
Controls	Yes	Yes	Yes
Province fixed effects	Yes	Yes	Yes
Year fixed effects	Yes	Yes	Yes
Observations	504	504	504

Notes: The dependent variable is the unscaled livestock-loss ratio, defined as mortality in year t divided by livestock stock in year t − 1. NDVI_anom is standardized within province over 2000–2024. Driscoll–Kraay standard errors use the locked two-lag specification. Province and year fixed effects are included. N = 504, covering 21 aimags from 2001 to 2024. ** p < 0.05.

). The lagged NDVI association is negative in both the baseline and extended specifications, and dzud severity is also positively associated with loss when included. In contrast, lagged SPI_12 and contemporaneous NDVI do not provide comparable support. H1 is therefore supported for lagged vegetation condition and dzud severity, but not uniformly across all ecological indicators.

4.2.2. Livestock Loss and Out-Migration Pressure

M2 indicates that years following greater proportional livestock loss are associated with higher standardized registered out-migration pressure (

Table 5. M2: Livestock loss and out-migration pressure.

Variable	Push Model 1	Push Model 2
L1_LivestockLoss_rate	1.067 *** (0.314)	1.113 *** (0.319)
L1_SPI_12	0.018 (0.035)	0.003 (0.031)
L1_NDVI_anom	−0.039 (0.024)	−0.038 (0.024)
L1_RealGDP_pc		0.0000178 *** (0.0000062)
Province fixed effects	Yes	Yes
Year fixed effects	Yes	Yes
Observations	504	504

Notes: Dependent variable = pooled standardized registered Out_Migration; the fixed transformation uses all 550 observations and is not a per-1000 rate. L1_LivestockLoss_rate is the prior-year unscaled livestock-loss ratio. Driscoll–Kraay standard errors use the locked two-lag specification; province and year fixed effects are included. N = 504 for 21 aimags over 2001–2024. *** p < 0.01.

). This result is stable across both specifications; lagged SPI_12 and NDVI_anom are not statistically significant, while real GDP per capita is positive in the fully specified model. Substantively, livelihood loss is more directly aligned with subsequent registered mobility pressure than the ecological indicators in this specification. The administrative records do not identify destinations, so the result cannot be interpreted as movement to Ulaanbaatar. Figure 5 places this association beside M1 without implying a causal chain.

4.2.3. Mediation-Oriented Product Diagnostic

M5 is consistent with, but does not identify, an intermediate livelihood role for livestock loss (

Table 6. Mediation-oriented product diagnostics: NDVI_anom, LivestockLoss_ratio, and Out_Migration_z.

Measure	Estimate	Interpretation
Path a: L1_NDVI_anom to LivestockLoss_ratio	−0.01481 ** (SE = 0.00617)	Lower lagged NDVI_anom is associated with higher proportional livestock loss
Path b: L1_LivestockLoss_ratio to Out_Migration_z	1.115 *** (SE = 0.318)	Higher prior-year livestock loss is associated with higher standardized registered out-migration pressure
Total association, c	−0.03790	Reduced-form L1_NDVI_anom to Out_Migration_z association; p = 0.094
Direct association, c′	−0.03618	L1_NDVI_anom coefficient with lagged livestock loss included; p = 0.105
Product diagnostic, a × b	−0.01652	Product of separately timed origin-side coefficients; not an exact mediated effect
Sobel Z/p-value	−1.981/0.0476	Sampling diagnostic for a × b
Bootstrap 95% CI	[−0.02705, −0.00725]	10,000-draw aimag-cluster bootstrap; seed 42
Exact mediated share	Not reported	Different lag positions mean a × b ≠ c − c′; an exact mediation percentage is not identified

Notes: The product a × b is accompanied by a Sobel statistic and an aimag-cluster bootstrap with 10,000 draws and seed 42. Because path a and path b refer to different lag positions, the product is a mediation-oriented diagnostic rather than an exact decomposition of the total association. No mediated percentage is reported. *** p < 0.01, ** p < 0.05.

). The negative path-a and positive path-b coefficients produce a negative product whose cluster-bootstrap interval excludes zero, but the component equations use different lag positions. Because a × b is not an exact c − c′ decomposition and neither c nor c′ is statistically significant at the 5% level, H4 is interpreted narrowly: the two origin-side associations are compatible with an intermediate role for livestock loss, not an exact mediated share or causal pathway.

4.3. BuiltUp Transfer Validation and Downstream Scope Decision

4.3.1. Province Transfer and Error Trade-Offs

Table 7. BuiltUp classifier development diagnostics under leakage-safe leave-one-ADM1-out evaluation.

Candidate	Balanced Accuracy	MCC	Recall	Specificity	Key Error
Frozen original	0.4741	−0.0530	0.3631	0.5852	Bare→built FPR = 0.9491
Benchmark 1/Candidate 4	0.8156	0.6312	0.8215	0.8097	Crop→built FPR = 0.5677
Candidate B/two-stage	0.6677	0.3532	0.5114	0.8240	Bare = 0.3474; crop = 0.0592; grass = 0.0795

Notes: N = 39,600 development/diagnostic observations from 2016 to 2024. Dynamic World is a reference label only. Benchmark 1 is the strongest aggregate spectral model; Candidate B illustrates the commission–recall trade-off. Candidate specifications and cutoffs were fixed before outer evaluation, so no inner tuning was required. No model was frozen for historical BuiltUp reconstruction.

shows the central classifier-development trade-off: no spectral candidate combines acceptable class-specific error control with stable province transfer. Benchmark 1 improves overall discrimination but still confuses cropland with built surfaces; Candidate B reduces crop and grass false positives by sacrificing recall and still leaves substantial bare-soil error. Govisumber is the clearest held-out example (Figure 6), showing that the problem is not average accuracy alone but transfer to local land-cover conditions. Detailed metrics are reported in Supplementary Tables S9–S11.

4.3.2. Consequences for UEI, Mismatch, M3, and M4

Table 8. Downstream analytical decisions following the BuiltUp validation gate.

Output/Analysis	Dependency	Validation Status	Status in Final Analysis	Future Prerequisite
Historical BuiltUp panel	Direct BuiltUp measure	Failed transfer validation	Excluded from validated evidence	Context-aware historical reconstruction
UEI	Includes BuiltUp as a direct component	Invalid input	Not reported	Validated BuiltUp or a separately prespecified urban construct
Mismatch Index	Depends on UEI	Invalid input	Not reported	Valid satellite urbanization construct
M3/H3	UEI/BuiltUp primary outcome	Invalid primary outcome; short time series	Not evaluated	Valid urban outcome and prespecified small-sample design
M4 + alternative W	UEI dependent variable	Invalid dependent variable	Not re-estimated/not retained	Valid urban outcome, followed by spatial-weight robustness

Notes: “Not retained” indicates that the corresponding historical result depended materially on the BuiltUp/UEI measurement branch that did not pass transfer validation. Accordingly, these outputs are not treated as validated evidence in the final analysis. Re-estimation of UEI-, Mismatch-, M3-, and spatial-model outcomes is deferred until a context-valid historical BuiltUp measure or another prespecified and validated urban outcome becomes available.

records the final downstream scope decision. The failed historical BuiltUp series is not used as validated evidence. The original UEI is therefore withdrawn, and the Mismatch Index is not computed from that invalid composite. H3/M3 is not evaluated because its primary urban outcome depended on UEI/BuiltUp, and the UEI-based M4 spatial model and alternative-W robustness are not re-estimated. Retaining those historical estimates would give false precision after a failed measurement gate. H3 remains documented as a prespecified hypothesis that could not be evaluated under the final validated architecture.

4.4. Robustness and Sensitivity Checks

This subsection reports robustness checks only for retained results. The purpose is to assess whether the canonical origin-side findings are sensitive to sample composition or specification and to document the locked NTL measurement warning. BuiltUp/UEI/Mismatch/M3/spatial robustness is not reported because the downstream branch failed its input-validity gate.

4.4.1. Leave One Province out Sensitivity Analysis

Leave-one-province-out checks show that the two core origin-side relationships do not depend on any single aimag (

Table 9. Leave-one-province-out sensitivity analysis.

Relationship	Mean β	Min β	Max β	% p < 0.05	% Same Sign
L1_NDVI_anom → LivestockLoss_ratio	−0.01479	−0.01619	−0.01278	100.0%	100.0%
L1_LivestockLoss_ratio → Out_Migration_z	1.113	1.021	1.275	100.0%	100.0%
Reduced form: L1_NDVI_anom → Out_Migration_z	−0.03792	−0.04584	−0.03170	4.8%	100.0%

Notes: Each of the 21 aimags is excluded once from the Phase 1R canonical two-way fixed-effects specification. Driscoll–Kraay standard errors use the same locked covariance settings as the main origin-side models. Out_Migration_z is the pooled standardized registered out-migration count; LivestockLoss_ratio is an unscaled proportion. The table reports the mean and range of the target coefficient across 21 iterations, the share with p < 0.05, and sign stability.

). The lagged NDVI–livestock-loss coefficient remains negative and statistically significant in all 21 exclusions, and the livestock-loss–out-migration coefficient remains positive and statistically significant in all 21. By contrast, the reduced-form NDVI–migration association keeps the same negative sign but is statistically significant in only 1 of 21 iterations. The strongest robustness evidence therefore concerns the two linked origin-side associations, not the reduced-form relationship.

4.4.2. Additional Specification and Measurement Checks

Additional checks reinforce this hierarchy of evidence. The lagged NDVI result remains negative and statistically significant in the baseline and extended M1 models, whereas contemporaneous NDVI is small and not statistically significant; the fully specified M2 model retains a positive and statistically significant lagged livestock-loss coefficient (Supplementary Table S2). Historical local-projection, interaction, threshold, nonlinear feature-importance, UEI-composite, and spatial outputs are not treated as current evidence when they were not reproduced from the authoritative retained inputs.

Nighttime-light checks support a different conclusion: levels align strongly across sensors, but provincial ranks are less stable near the transition. The 2013 overlap has Pearson r = 0.987 but Spearman ρ = 0.639, and rank continuity weakens further across 2012–2013 (ρ = 0.555). These diagnostics remain a material measurement warning that pooled z-standardization does not remove.

5. Discussion

5.1. Urban Monitoring, Measurement Validation, and Satellite–Administrative Divergence

The measurement contribution is the validation-first treatment of satellite urban indicators. Nighttime lights remain useful for longitudinal settlement/activity monitoring, but strong cross-sensor-level agreement does not imply stable provincial ranking. The BuiltUp problem is more consequential: the frozen NDBI–NDVI rule systematically confuses bare and some cropland surfaces with built land, and enriched spectral candidates do not transfer reliably across provinces. This does not invalidate Landsat urban mapping generally; it shows that this tested spectral architecture is inadequate as a validated 2000–2024 BuiltUp measure for Mongolia’s heterogeneous drylands [14,19,20,27,51,52,54].

The NTL warning is consistent with known DMSP/VIIRS radiometric and sensor-design differences that motivate intercalibration and cross-sensor harmonization [28,29,30,55]. The canonical series is therefore retained for descriptive longitudinal level comparisons, while province ranking around the transition is interpreted cautiously.

The failed BuiltUp gate narrows destination-side conclusions. Historical UEI persistence, same-year migration, Mismatch, and spatial UEI results are not reported because their dependent variable no longer met the measurement requirement. Satellite–administrative divergence remains substantively important, but quantitative mismatch requires a validated satellite urban construct.

5.2. Origin-Side Livelihood Stress and Migration Pressure

The origin-side results align with pastoral-vulnerability research while adding a longitudinal provincial perspective. The lagged NDVI pattern is consistent with evidence that drought reduces Mongolian pasture production [26] and that weather fluctuations affect livestock productivity [24]; the positive dzud-severity term also accords with research on severe-winter livelihood disruption [7,13,25]. The subsequent loss–out-migration association is compatible with push–pull and asset-based livelihood accounts [42,43]. Because environmental stress can also constrain mobility [1,2,44], these aimag-level associations should not be read as household-level causal effects.

M5 links the two origin-side findings without establishing a causal mechanism. Its negative product is statistically distinguishable from zero under the locked diagnostics, but different lag positions prevent an exact c − c′ decomposition or mediated percentage. The practical implication is limited to monitoring: vegetation conditions and livestock loss help identify periods of elevated pastoral livelihood pressure associated with subsequent registered out-migration, not movement specifically to Ulaanbaatar or direct urban expansion.

5.3. Why the Urban-Composite and Spatial Branch Was Withdrawn

Spatial dependence remains relevant because settlement, infrastructure, and administrative classifications are geographically structured [33,34,35], but spatial robustness cannot repair an invalid dependent variable. After BuiltUp failed transfer validation, re-estimating Moran, SAR, SEM, SDM, or alternative weights on the original UEI would only analyze a construct that failed its measurement gate. Future spatial analysis therefore requires a context-valid historical urban outcome.

5.4. Urban Planning and Adaptation Monitoring Implications

The retained findings have implications for validation-first urban monitoring and adaptation planning. In the current study, canonical nighttime lights can complement administrative records as a descriptive settlement/activity signal, but quantitative satellite–administrative divergence should not be inferred until a context-valid physical BuiltUp measure is available. Operationally, retained satellite and livelihood indicators are screening inputs rather than automatic intervention triggers. This study does not estimate a universal satellite threshold, and it does not retain a UEI or Mismatch threshold after the composite was withdrawn. A defensible monitoring system would calibrate any future threshold against field, cadastral, infrastructure, and administrative evidence and would specify the decision-specific costs of false positives and false negatives. NDVI and livestock-loss monitoring can support origin-side situational awareness, while NTL can complement administrative monitoring with explicit sensor-transition caution.5.5. Limitations and Future Research.

Several limitations remain. Administrative migration records capture registered movement without a complete origin–destination matrix and may miss temporary or delayed registration. MODIS NDVI is coarse for fine peri-urban change, and nighttime lights retain cross-sensor rank instability despite strong linear calibration. Most importantly, the historical BuiltUp classifier and enriched spectral candidates fail the required transfer standard, so BuiltUp, UEI, Mismatch, M3, and UEI-based spatial results are not retained. Time-varying confounding may also persist despite fixed effects, including disaster assistance, non-agricultural employment, market access, land and pasture institutions, infrastructure investment, and COVID-period mobility constraints; its net direction cannot be identified from the available data.

Future work should resolve the measurement bottleneck before rebuilding the urban-composite branch. Historical BuiltUp reconstruction should add local spatial context, multi-year Landsat persistence, explicit cross-sensor portability, and a prospectively specified independent validation sample. Only after that gate is passed should UEI/Mismatch, Ulaanbaatar M3, and spatial models be reconsidered. Household-level origin–destination migration, cadastral and service-infrastructure data, and higher-frequency Sentinel-2 or Dynamic World products could then test the timing and mechanisms linking pastoral mobility pressure to physical urban change more directly.

6. Conclusions

This study re-evaluates multi-source monitoring of ecological stress, livelihood disruption, migration pressure, and urban change in Mongolia over 2000–2024. The retained evidence combines origin-side NDVI, CHIRPS, dzud, livestock, and registered migration analyses with an audited nighttime-light series; Landsat BuiltUp is retained only as a validation finding.

Canonical nighttime lights remain useful descriptively but carry a cross-sensor rank warning. The historical Landsat BuiltUp rule fails external transfer validation, and spectral-only repairs do not transfer stably across provinces. Consequently, the original UEI, Mismatch, M3, and UEI-based spatial results are not reported.

The satellite–administrative divergence question remains important, but the final study does not quantify it with the withdrawn composite. The methodological contribution is instead to show that disagreement between measurement systems should be investigated only after each satellite component passes an appropriate validation gate.

Origin-side results show that poorer lagged vegetation conditions are associated with higher proportional livestock loss, dzud severity with higher loss in the extended model, and greater prior-year livestock loss with higher standardized registered out-migration pressure. These are livelihood and mobility-pressure associations, not evidence of migrant destinations or a causal pathway to Ulaanbaatar.

Overall, the study demonstrates a validation-first monitoring principle: retain supported origin-side evidence and an explicitly qualified NTL signal, but withdraw a composite when its physical component does not transfer reliably. Context-aware historical BuiltUp reconstruction is therefore a prerequisite for future UEI, Mismatch, M3, or spatial urbanization analysis.