Chapter Four · failure evidence
What Spatial Mapping & GIS got wrong, from 74 dissertations
Spatial mapping and GIS analyses frequently encounter modeling failures when spatial dependencies, resolutions, and boundary conditions are improperly specified. Across these records, standard non-spatial regressions, geostatistical interpolation methods, spatial autoregressive frameworks, and siting algorithms repeatedly break down due to autocorrelation, aggregation error, and spatial non-stationarity. These records come from PhD theses at 22 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Spatial proximity assumptions and distance metrics break down under spatial non-stationarity
Uniform spatial density thresholds, Euclidean distance assumptions, and kernel-based graphs struggle when underlying spatial processes exhibit severe localized clustering or complex geometries. Distance-based metrics often fail over large geographic distances and cannot resolve conflicting spatial configurations across heterogeneous regions.
Tried and failed
distance-based Gaussian kernel spatial graph construction applied to spatiotemporal time-series forecasting. Outcome: worse than baseline. Reason: failed to capture relational correlations between distant nodes
Graph-based Time-series Forecasting in Deep Learning · Virginia Tech
Lost to a baseline
RREuc model performed worse than simpler spatial models in simulation due to mismatch with rectangular plot geometry.
STABILITY OF TRAITS ACROSS ENVIRONMENTS USING IMAGE PHENOTYPING AND GENOTYPING · Cornell
Considered and rejected
Considered and rejected: Rejected analytical distance-based Gaussian spatial covariance functions for estuarine water level interpolation due to their inability to represent channel topology and hydrodynamic phase lags.
A Scalable and Adaptable Coastal-Urban Flood Modeling Framework for Changing Climates · Georgia Tech
Tried and failed
spatial subsetting to high-density regions applied to human mobility spatial flow models. Outcome: did not generalise. Reason: density filtering did not align regional parameter estimates or improve gridded-scale likelihood
The role of seasonal population movements in malaria transmission and control in sub-Saharan Africa · Imperial
Tried and failed
DBSCAN clustering applied to spatial trajectory clustering with varying density. Reason: fixed density threshold merged distinct clusters in regions of varying spatial density
A Data-driven Methodology for Aircraft Trajectory Analysis to Improve Mid-air Conflict Detection in Terminal Airspace · Georgia Tech
Tried and failed
network distance over Euclidean distance metric applied to spatial econometric regression models. Outcome: worse than baseline. Reason: reduced statistical significance and provided no meaningful predictive improvement over straight-line distance
Tried and failed
spatial correlation of agent-based simulation and historical usage applied to urban micromobility travel demand estimation. Outcome: no signal. Reason: historical data exhibited extreme localized spatial clustering uncaptured by citywide latent demand modeling
Pedal Power: Expanding Austin's Bikeshare Network · UT Austin
Tried and failed
threshold-based domain truncation with density renormalization applied to spatial risk assessment models. Outcome: worse than baseline. Reason: Renormalizing probability density after removing low-density areas artificially inflated exposure in remaining high-density regions
Integration and Effects of Three-Dimensional Terrain Models on Potential Crash Areas of Aircraft · Virginia Tech
Tried and failed
density-based spatial clustering of trajectory data applied to spatial trajectory anomaly detection. Outcome: did not generalise. Reason: dynamic airspace constraints and environmental perturbations degraded clustering performance
Tried and failed
spatial correlation of modeled exposure with demographic metrics applied to cross-regional socioeconomic exposure disparities. Outcome: did not generalise. Reason: local urban configurations produced conflicting positive and negative correlation trends across different metropolitan regions
An Analysis on Aircraft Overflight Noise Distribution on Airport Adjacent Communities · MIT
Tried and failed
global ordinary least squares on spatial data applied to spatial emergency demand modeling. Reason: spatial non-stationarity and confounding caused counterintuitive negative regression coefficients
Tried and failed
nearest neighbor spatial interpolation of meteorological data applied to hydrological watershed modeling. Outcome: worse than baseline. Reason: failed to capture spatial variability compared to distance-weighted methods, worsening discharge simulation metrics
Improving Watershed Models to Achieve a Better Prediction of Water Quantity and Quality · Virginia Tech
Tried and failed
similarity-based parameter regionalization applied to hydrological models on distant ungauged catchments. Outcome: did not generalise. Reason: spatial proximity breakdown over larger geographic distances (>20 km) led to poor predictive performance
Improving Watershed Models to Achieve a Better Prediction of Water Quantity and Quality · Virginia Tech
Spatial interpolation and kriging fail under data sparsity and boundary extrapolation
Gaussian processes and kriging methods produce large extrapolation errors outside training boundaries and generate spurious local extrema on coarse grids. These interpolation techniques also break down when stationary semivariogram assumptions are violated by non-stationary noise or when observation points are too sparse to estimate spatial structure.
Tried and failed
Gaussian process regression applied to sparse discrete spatial count data. Outcome: worse than baseline. Reason: failed to model spatial distribution of sparse discrete counts effectively
Adaptive Robotic Search and Sampling of Sparse Natural Phenomena · MIT
Tried and failed
Gaussian process regression with squared-exponential kernel applied to spatial magnetic field mapping for localisation. Outcome: did not generalise. Reason: Gaussian process regression suffered high extrapolation errors outside the spatial training boundary
Enhanced Navigation Using Aerial Magnetic Field Mapping · Virginia Tech
Tried and failed
Voronoi spatial interpolation and layer overlay applied to sparse spatial suitability mapping. Outcome: did not generalise. Reason: Low observation density caused artificial boundary generation, severely degrading validation map purity
Application of GIS and remote sensing for land use planning in the arid areas of Jordan · Cranfield
Tried and failed
universal kriging with Gaussian semivariogram applied to spatial groundwater level interpolation. Outcome: data insufficient. Reason: spatial smoothing and grid resolution failed to resolve small-scale, localized depressions
Tried and failed
Purely data-driven MLPs and Gaussian process regression applied to spatial impact localisation. Outcome: did not generalise. Reason: Models failed to extrapolate outside the spatial coverage of the training data
Tried and failed
sequential Gaussian kriging spatial interpolation applied to continuous field reconstruction on grids. Outcome: worse than baseline. Reason: Generated spurious local extrema and exhibited excessive sensitivity to grid coarseness compared to splines
Hydrocarbon expulsion and scaling · UT Austin
Tried and failed
kriging with stationary semivariograms applied to spatial prediction under non-stationary noise. Outcome: did not generalise. Reason: produced radially unbounded errors and zero predictive variance under non-stationary ambient noise
Communication-Aware, Scalable Gaussian Processes for Decentralized Exploration · Virginia Tech
Tried and failed
precision-weighted aggregation of local gaussian processes applied to spatial satellite temperature prediction. Outcome: did not generalise. Reason: aggregation degraded predictions across large unobserved spatial extrapolation regions
Precision Aggregated Local Models · Virginia Tech
Tried and failed
local search radius ordinary kriging interpolation applied to spatial prediction with sparse observations. Reason: Prediction locations lacked sufficient conditioning points within the fixed local search radius.
Worlds Collide through Gaussian Processes: Statistics, Geoscience and Mathematical Programming · Virginia Tech
Lost to a baseline
Under spatial leave-one-out cross-validation (SLOO-CV) on Malawi and Tanzania PHIA data, kernel models (CK/IK) were beaten by the simpler Besag model (Malawi SLOO CRPS: 19.3 Besag vs 29.0 CK / 28.3 IK).
Bayesian spatio-temporal methods for small-area estimation of HIV indicators · Imperial
Considered and rejected
Considered and rejected: Rejected ordinary kriging for spatial interpolation because irregular, sparse field-parcel centroids provided insufficient data points to compute usable semi-variograms.
Estimating field-scale soil moisture using SAR remote sensing and the COSMOS-UK network · Cranfield
Considered and rejected
Considered and rejected: Rejected Kriging spatial interpolation in favor of Inverse Distance Weighting (IDW) because Kriging required cross-variable spatial correlation not suited to the univariate layout.
Assessment of groundwater aquifer impact from artificial lagoons and the reuse of wastewater in Qatar · University of Nottingham Repository
Complex spatial autoregressive models suffer numerical instability and underperform simpler baselines
Incorporating conditional autoregressive priors, spatial random effects, or Markov random fields often leads to phase transitions, high error rates, and failure to partition spatial variance. In several evaluations, these complex spatial formulations proved harder to estimate and were outperformed by simpler non-spatial models or basic regression specifications.
Tried and failed
spatial-dependence occupancy modeling applied to transect-based species sign surveys. Outcome: worse than baseline. Reason: Spatial autocorrelation model formulation lacked empirical support over standard single-season occupancy baseline
Tried and failed
centered autologistic model applied to spatial binary data modeling. Outcome: unstable. Reason: experiences unintended phase transitions under strong spatial dependence, heavily penalizing configurations matching the mean trend
Bayesian spatial models for discrete data : methodological advances with applications in urban sociology · UT Austin
Tried and failed
Conditional autoregressive models for spatial inference applied to spatial regression with autocorrelated covariates. Reason: Strong autocorrelation and large spatial scale inflated Type I error rates across models
Conditional autoregressive models : implications for inference · UT Austin
Tried and failed
adding conditionally autoregressive spatial random effects applied to spatial causal effect estimation models. Outcome: worse than baseline. Reason: deteriorated treatment effect estimation accuracy in simulated spatial confounding environments
Tried and failed
full-rank multinomial Markov random fields applied to spatial categorical data modeling. Reason: The reference category baseline forces its spatial auto-correlation to near zero, creating category-relabeling asymmetry
Symmetry, stability, and spatio-temporal dependence in multinomial Markov random field models · Iowa State
Lost to a baseline
Gaussian CAR spatial random effects model had worse fit (BIC = 24225.730) than the non-spatial nested GLMM baseline (BIC = 19743.090).
An International Spatial Analysis of the Welfare Spending’s Influence on Measles Immunization · HARVEST
Lost to a baseline
ICAR model had substantially higher Type II error rates (up to 0.681) than the non-spatial baseline (0.374) in high autocorrelation conditions.
Conditional autoregressive models : implications for inference · UT Austin
Lost to a baseline
In spatial binary simulations with no repetition (λw = 1), the naive spatial binomial probit model yielded lower MSE for spatial dependence λη than the proposed SPPM (e.g., MSE 0.242 vs 1.427 at λη = 0.1, and MSE 0.511 vs 7.830 at λη = 0.6).
Spatial modeling of repeated events · MSpace - University of Manitoba
Considered and rejected
Considered and rejected: Occupancy modelling was abandoned in favor of logistic regression (GLM) for main landscape analyses due to failure of occupancy models to converge under complete separation and inability to account for spatial autocorrelation
Listening in on the forest: use of bioacoustics to preserve soundscapes and rare species · Imperial
Considered and rejected
Considered and rejected: Rejected analyzing spatial models with spatial random effects due to higher BIC and inability to partition spatial error variance.
An International Spatial Analysis of the Welfare Spending’s Influence on Measles Immunization · HARVEST
Considered and rejected
Considered and rejected: Rejected the Besag-York-Mollié (BYM) convolution model due to fixed spatial autocorrelation (rho_s = 1), non-identifiable separation of spatial and independent random effects, slow MCMC convergence, and unreliable parameter estimation.
Conditional autoregressive models : implications for inference · UT Austin
Coarse spatial resolution and regional aggregation obscure fine-scale geographic variations
Discretizing spatial models using coarse computational meshes or aggregating data to broad administrative units obscures local spatial patterns and dynamic flow processes. This scale mismatch introduces substantial numerical variance and causes models to smooth out localized physical, environmental, and socioeconomic phenomena.
Tried and failed
average pooling spatiotemporal sensor features applied to traffic incident impact duration prediction. Outcome: worse than baseline. Reason: simple pooling discarded fine-grained spatial density and temporal dynamics necessary for accurate prediction
On Modeling Dependency Dynamics of Sequential Data: Methods and Applications · Virginia Tech
Tried and failed
spatial distance regression with demographic controls applied to county-level business formation rates. Outcome: no signal. Reason: aggregation to county level obscured tract-level spatial patterns after controlling for demographics
Opportunity or Desperation: Investigating the COVID-19 Surge in Business Creation · Harvard
Tried and failed
coarse regional reporting data for spatial prediction applied to chemical contaminant leaching across geographic zones. Outcome: did not generalise. Reason: zip code level aggregation lacked sufficient spatial resolution and accuracy to predict localized chemical detections
Tried and failed
statistical downscaling of global climate models applied to regional precipitation variability in complex terrain. Outcome: did not generalise. Reason: models smoothed and flattened spatial precipitation variability compared to observed local station records
Projecting Planning-Related Climate Impact Drivers for Appalachian Public Health Support · Virginia Tech
Tried and failed
coarse spatial mesh discretization applied to transient porous media flow simulation. Outcome: worse than baseline. Reason: insufficient spatial resolution failed to capture transient flow dynamics accurately over time
Integrated modelling of the clogging processes of plastic grid permeable pavement · Cranfield
Tried and failed
numerical quadrature of pointwise error bounds applied to parametric model order reduction. Reason: local spatial discretization inaccuracies caused severe overestimation of integrated error bounds
Efficient 𝐻₂-Based Parametric Model Reduction via Greedy Search · Virginia Tech
Tried and failed
model equivalence testing with coarse spatial inputs applied to third-party spatial risk assessment. Outcome: did not generalise. Reason: coarse spatial resolution introduced variance exceeding equivalence bounds
Assessment of Multiple Third-Party Risk Models Incorporating Uncertainty Quantification and Sensitivity Analysis · Virginia Tech
Tried and failed
GIS digital elevation models for hydrological modeling applied to excavated terrain and plateau surface hydrology. Outcome: data insufficient. Reason: Coarse spatial resolution and missing historical imagery failed to resolve excavation micro-topography and hydrology
From Wasteland to Biocultural Heritage: Negotiation by Design in Khotale, Konkan, India · Cornell
Tried and failed
general circulation models for local climate projections applied to regional scale climate impact modeling. Outcome: data insufficient. Reason: spatial resolution was too coarse to capture regional variations accurately
Innovative methodology for the assessment of thermal comfort and climate resilience in social housing in Coahuila, Mexico · University of Nottingham Repository
Lost to a baseline
Graphical map generation outperformed computer-based digital mapping in spatial resolution and preservation of fine linear features (e.g., thin colluvium streams).
Graphical and digital slope stability analyses for Giles County, Virginia · Virginia Tech
Standard non-spatial regressions fail due to uncorrected residual spatial autocorrelation
Ordinary least squares and standard linear regressions consistently violate independence assumptions when applied to geographically distributed observations. Researchers reject non-spatial specifications because uncorrected spatial autocorrelation inflates coefficients, underestimates standard errors, and leads to severe pseudoreplication.
Tried and failed
generalized additive models with spatial weight correction applied to species distribution modeling. Reason: residuals continued to exhibit significant spatial autocorrelation despite spatial weight corrections
Tried and failed
non-spatial logistic regression applied to geocoded epidemiological disease data. Reason: significant spatial autocorrelation in model residuals violated independence assumptions
Enhancing Electronic Health Record Data For Population Health Studies · Penn
Tried and failed
difference-in-difference-in-differences regression without spatial correction applied to spatially distributed panel data. Reason: uncorrected spatial autocorrelation inflated regression coefficients and underestimated standard errors
Lost to a baseline
OLS regression underperformed SERM across all metrics (e.g., OLS R2 of 0.184 vs SERM R2 of 0.255, and higher AIC of 28,051.7 vs 27,984.1 for devch vs slrdevch) due to spatial autocorrelation
Risk and Development Along the California Coast: A Study of Projected Sea-Level Rise and Land-Use · TXST Digital Repository
Considered and rejected
Considered and rejected: Rejected non-spatial linear regression models for large-scale herbarium data because spatial autocorrelation caused pseudoreplication and more than doubled the estimated phenological shift (-1.34 vs -0.56 days/decade).
Effects of global change on plants: tracing the footprints of climate warming and land use from herbaria to forest understories. · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: National-level OLS regression was rejected due to significant residual spatial autocorrelation violating independence assumptions.
Greater Access to Recreational Resources is Associated with More Leisure-time Physical Activity Engagement in Counties Across the United States · Scholars' Bank
Considered and rejected
Considered and rejected: Rejected standard flat Ordinary Least Squares regression for regional correlational analysis due to significant spatial autocorrelation violating error independence; adopted spatial lag regression models instead.
Considered and rejected
Considered and rejected: Rejected standard linear regression without spatial consideration due to strong spatial autocorrelation (Moran's I = 0.62) in sub-district income data
Evaluating Industrial Activities and Its Social/Economic Impact in a Developing Country with Remotely Sensed Data: A Case Study of Eastern Economic Corridor, Thailand · Queens University Institutional Repository
Considered and rejected
Considered and rejected: Rejected non-hierarchical/unweighted OLS in favor of entropy-balanced hierarchical random-intercept-and-slope modeling to prevent model misspecification and account for spatial clustering.
ECONOMIC AND SOCIAL IMPACTS OF AIR POLLUTION: A CAUSAL INFERENCE APPROACH · Harvard
Spatial coordinate features and spatial cross-validation introduce artifacts and evaluation bias
Directly including geographic coordinates or spatial cluster centroids as features causes predictive models to overfit and generate unnatural spatial boundaries and map discontinuities. Standard cross-validation schemes also cause evaluation failures by either enabling spatial information leakage or generating overly pessimistic out-of-sample metrics.
Tried and failed
including spatial coordinates as tree model features applied to global spatial distribution regression. Reason: produced unrealistic spatial artefacts and discontinuities in continuous global reconstruction maps
Environmental pollution as a marine tracer: measuring and modelling anthropogenic lead in the ocean · Imperial
Lost to a baseline
Spatial filtering (2-km and 10-km) and kernel density bias files performed worse (higher commission/overfitting errors) than the baseline unmanipulated dataset in MaxEnt.
The conservation status of Owston’s Civet Chrotogale owstoni · University of Nottingham Repository
Tried and failed
adding spatial cluster centroids as features applied to tabular spatial regression models. Outcome: overfit. Reason: spatial centroid features caused overfitting, reducing out-of-sample predictive performance in tree and regularized models
City Services for a Smarter Boston: Constituent Driven Intelligence through BOS:311 · Harvard
Tried and failed
Cross-validation error estimation for spatial regression applied to spatial property prediction from covariates. Reason: Cross-validation produced more pessimistic error estimates than independent validation.
Tried and failed
rule-based regression with spatial features applied to soil texture mapping. Outcome: overfit. Reason: Standard cross-validation produced inflated metrics that inversely correlated with true independent spatial validation accuracy.
An approach to map soil texture class on the Iowan Erosion Surface · Iowa State
Tried and failed
multiscale spatial autocorrelation on geometric curvature features applied to terrain anomaly and feature detection. Outcome: did not generalise. Reason: produced excessive false positives and near-chance agreement when transferred across study areas
An Evaluation of DEM Generation Methods Using a Pixel-Based Landslide Detection Algorithm · Virginia Tech
Considered and rejected
Considered and rejected: Rejected standard random and spatial/block k-fold cross-validation because random CV yields overly easy predictions (information leakage) while block CV forces severe extrapolation
Spatially-informed machine learning for augmented subsurface characterization, modeling, and interpretation · UT Austin
Considered and rejected
Considered and rejected: Rejected using latitude and longitude features in hierarchical clustering for rainfall up-scaling because spatial coordinates dominated and degraded precipitation pattern grouping.
Underpinning research on the dynamical aspects of one-dimensional nonlinear lattices · Research Repository UCD
Multi-criteria spatial decision analysis and facility siting methods produce inconsistent rankings
Multi-criteria spatial evaluations and automated siting algorithms often generate candidate rankings that conflict with empirical project performance and alternative decision methods. Siting models also fail when required spatial real estate data are unavailable or regional hazard screening databases omit localized facilities.
Tried and failed
multicriteria weighted index with reduced demand weighting applied to spatial facility location selection. Outcome: worse than baseline. Reason: downweighting simulated demand produced site rankings inconsistent with empirical usage patterns
Pedal Power: Expanding Austin's Bikeshare Network · UT Austin
Tried and failed
multicriteria spatial analysis with real estate pricing applied to facility location selection. Outcome: data insufficient. Reason: reliable real estate pricing data was too difficult to access reliably
Locating mobility hubs in college towns: A case study of Ames, Iowa · Iowa State
Tried and failed
multi-objective evolutionary algorithm with geospatial techno-economic modeling applied to offshore wind farm layout and site selection. Outcome: did not generalise. Reason: real-world deployed project configuration was dominated and excluded from the Pareto-optimal set
Tried and failed
VIKOR multi-criteria decision making with equal weights applied to spatial multi-criteria site suitability ranking. Reason: produced rankings with low correlation compared to other multi-criteria decision making methods
Tried and failed
standardized environmental hazard screening tools applied to institutional sites and regional hazards. Outcome: data insufficient. Reason: spatial databases omitted specific facilities and lacked coverage of localized rural environmental hazards
Tried and failed
correlation-based feature selection for spatial siting applied to renewable energy resource placement. Outcome: worse than baseline. Reason: Produced less total generation and higher spot-market prices than cost or price metric baselines.
Left open by the authors
Problems the authors named and did not get to.
Left open
Incorporate spatial correlation structures into the Bayesian population synthesis model framework. Blocker: Lacks specific mathematical formulation, targets, and spatial structure definitions
Inference for Populations: Uncertainty Propagation via Bayesian Population Synthesis · Virginia Tech
Left open
Model the spatial variability of shear strength parameters within individual geologic units for regional landslide hazard assessment. Blocker: Lacks specific target geologic units, datasets, and spatial statistical methodology.
Regional earthquake-induced landslide assessments using a data-informed probabilistic approach · UT Austin
Left open
Extend the random parameter panel negative binomial model with heterogeneous overdispersion to incorporate spatial correlation across crash observation sites. Blocker: Lack of specific mathematical formulation or target spatial framework proposed for integration.
Random parameter models with nonlinear functional form and heterogeneous overdispersion · Texas Tech
Left open
Extend the hierarchical Poisson species-area model to include spatial dependence across archipelagoes using an irregular lattice CAR prior. Blocker: None
A hierarchical Poisson model for the species-area curve · Iowa State
Left open
Model spatial autocorrelation and implement hierarchical multi-level modeling across census tracts, counties, and metro regions using MGF sub-geographies. Blocker: None
Using land as a lens : multi-scalar analysis of metropolitan structure and growth · UT Austin
Left open
Incorporate spatial econometric models to account for spatial autocorrelation in soil organic carbon measurements across agricultural landscapes. Blocker: Requires the specific soil sampling spatial dataset and landscape field data from the study region.
Economic Analyses of Soil Carbon Management and Regenerative Agricultural Practices in Semi-Arid Ecoregions · Texas Tech
Left open
Investigate why the Q* fitness balance criterion derived from non-spatial models fails in specific spatial regions of aquatic stoichiometric PDE systems. Blocker: None
Left open
Develop distributed hypothesis testing algorithms under communication constraints for spatially correlated or non-independent node observations. Blocker: Lacks specific target error bounds, correlation model formulation, and concrete algorithmic approach
Time vs. Truth: Age-Distortion Tradeoffs and Strategies for Distributed Inference · EPFL
Left open
Extend the epidemiological model to include spatial dynamics and multi-year environmental variability. Blocker: Lack of specific mathematical formulation or data parameters for spatial and environmental variability
Using Modelling to Optimise the Use of Biological Control Agents Against Soil-Borne Plant Pathogens · Cambridge
Left open
Derive a formal mathematical relationship linking spatial autocorrelation and overdispersion across multiple length scales. Blocker: None
Experimental design and analysis for high-parameter spatial omics · MIT
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.