Many headline findings about how urban vegetation cools cities rest on one statistical step: choosing between a simple and a more complex model with an information criterion. Examples include the claim that cooling only starts above a background air temperature, that greenery warms surfaces in winter, and that a spatial-lag regression beats a plain one. An information criterion is a score that rewards fit and charges a fixed price per parameter. The Akaike criterion (AIC) is the standard one. The price is set as if every satellite pixel or every field measurement were an independent observation. Thermal images and site measurements are anything but independent. Neighboring pixels share weather and soil moisture, and many measurements come from the same park. The result is that the price of complexity is too low, and extra structure such as thresholds, sign reversals, and spatial terms can be "selected" even when it only fits correlated noise.
This note proposes a dependence-aware replacement for the complexity price. Instead of charging $2k$ for $k$ parameters, it charges the measured amount by which a model chases noise. That amount is computed by simulating noise with the spatial or site-level correlation actually present in the data. The aim is to re-adjudicate a specific set of published vegetation-cooling claims. The selection rules are first validated on synthetic temperature fields where the true model is known. The scientific payoff is a clear answer to which non-linear vegetation effects are real. The methodological payoff is an audited comparison of information criteria against spatial cross-validation in the setting where they are most often misused.
Core Claim
Vegetation-cooling studies make two specific errors, and both inflate apparent complexity. The first is using the nominal sample size, for example the 16,312 Landsat pixels in Beijing, when correlated residuals carry far less independent information. The second is counting a searched breakpoint, such as a temperature threshold, as one parameter, when searching over candidate breakpoints lets the model fit noise much more than one free coefficient would.
For a Gaussian model with fitted values $\hat y$, the expected optimism of the in-sample log-likelihood is proportional to $\sum_i \mathrm{Cov}(\hat y_i, y_i)$. We therefore propose the criterion
$$\mathrm{AIC}_{\mathrm{dep}} = -2\ln L(\hat\theta) + \frac{2}{\hat\sigma^2}\sum_{i=1}^{n}\widehat{\mathrm{Cov}}(\hat y_i, y_i).$$
Here $L(\hat\theta)$ is the maximized likelihood and $\hat\sigma^2$ is the residual variance. The covariance is estimated by refitting the full model-building procedure, including any breakpoint search or bandwidth search, on many simulated responses. Each simulated response is the fitted mean plus noise drawn from a dependence model estimated from the residuals: a spatial random field for rasters, or a site random effect for clustered field data.
For a linear model with independent noise the penalty reduces exactly to $2k$, so the criterion only departs from AIC when dependence or search actually matters. For a linear smoother with hat matrix $H$ and noise covariance $\Sigma$, the penalty becomes $2\,\mathrm{tr}(H\Sigma)/\sigma^2$. This exceeds $2k$ when the correlation in the noise lines up with the spatial pattern of the predictors. That is the case for vegetation indices, which are themselves spatially smooth.
We predict two outcomes. First, under this criterion, some published complexity choices reverse in favor of the simpler model. Second, it recovers the true model on realistic synthetic fields more often than nominal AIC and BIC (the Bayesian information criterion, which uses a heavier fixed price), and about as often as spatial block cross-validation. It should do so without the arbitrary choice of block size.
Positioning Against Prior Work
The motivating evidence comes from three studies that use information criteria to adopt richer structure.
Decoupling of the Municipal Thermal Environment Using a Spatial Autoregressive Model ranks OLS (ordinary least squares), spatial-lag, spatial-error and combined spatial autoregressive models of Landsat land surface temperature over Beijing by AIC and BIC. It reports the combined model winning with $R^2$ of 0.803–0.872 against 0.074–0.509 for OLS, and it also reports a positive winter NDVI–temperature association. NDVI, the normalized difference vegetation index, is a satellite greenness measure. That study prices complexity with $n=16{,}312$. Our work keeps its candidate models and asks whether the ranking holds under a dependence-aware price.
Identifying Surface Urban Heat Island Drivers and Their Spatial Heterogeneity in China's 281 Cities selects a separate spatial bandwidth per driver in multiscale geographically weighted regression by minimizing finite-sample corrected AIC (AICc). It concludes that the urban–rural NDVI difference acts at very local scales. Its AICc uses the trace of the hat matrix under independent errors. We substitute the dependence-aware penalty inside the same bandwidth search and test whether "local" survives.
Phenology acts as a primary control of urban vegetation cooling and warming pools 3,634 measurements from 77 sites. From these it reports cooling thresholds at background air temperature 10.0 °C, evapotranspiration (water evaporated from soil and leaves) 61.7 mm/month, and patch area 35.2 ha. This is the sharpest target. The effective sample is closer to 77 sites than to 3,634 points, and the thresholds were located by search. That study used variance partitioning rather than formal selection, so our contribution is the first criterion-based test of whether the thresholds are justified.
Other work constrains the design. The Interpolating Information Criterion for Overparameterized Models handles models with more parameters than data, where classical AIC fails because of singular curvature. Our models are small and regular. Their failure mode is dependence and search, not overparameterization, so we deliberately do not adopt that criterion. A Bayesian Information Criterion for Singular Models is set aside for the same reason. The method lineage comes from Principled Architecture Selection for Neural Networks, which estimated prediction risk with effective-parameter penalties under independent noise. We generalize that idea of an effective complexity measured from the fitted procedure to correlated noise.
The neural microclimate surrogates in Climate-sensitive Urban Planning through Optimization of Tree Placements and UrbanGraph are out of scope. Their datasets are simulator outputs where noise dependence is not the binding issue. The obvious combination of applying AIC to prune their networks would mostly reproduce what validation loss already does.
Research Program
The work has three stages.
The first stage builds the criterion and its dependence models. For rasters, we fit an exponential or Matérn covariance to the residuals of the most complex candidate model. Matérn is a standard family that controls how fast correlation decays with distance. Using the most complex candidate avoids underestimating noise. For the field synthesis, we fit a site random intercept plus residual noise. Each candidate's full fitting pipeline is then rerun on 500 simulated responses to estimate the penalty. The pipeline includes the threshold grid search, the bandwidth golden-section search, and spatial weight estimation. Uncertainty in the penalty is reported by bootstrap.
The second stage is a synthetic recovery study built on real covariates. We take actual Landsat-derived NDVI and water-index maps for Beijing and generate temperature as a known function plus spatially correlated noise. The noise correlation range is matched to the residual Moran's I in the real data; Moran's I is the standard measure of spatial clustering. There are four generating truths:
- linear cooling;
- linear cooling plus a true threshold;
- a true winter sign reversal;
- a true spatial-lag process.
The noise range is swept from near zero to several kilometers. The field-synthesis analogue resamples the real site structure and covariate values, with truths either with or without the 10 °C threshold.
The third stage is a re-analysis. All selection rules are applied to the three real datasets, and we report which published choices flip.
Baselines
The naive reference point is nominal AIC and BIC with the raw sample size and raw parameter counts, exactly as used in the Beijing spatial-autoregression study. This is what the field currently does, so it defines the error rate to be reduced. It is run on identical candidate sets and identical fits. The only thing that changes is the penalty.
The second baseline is AICc with an effective sample size computed from Moran's I: shrink $n$ to $n_{\mathrm{eff}}$ and plug it into the standard formula. This is the cheapest dependence fix a careful practitioner would try. It is the most direct test of whether our simulation-based penalty is needed at all. It shares our dependence model, so the comparison isolates whether measuring optimism beats rescaling $n$.
The strongest competitor is spatial block cross-validation on rasters and leave-one-site-out cross-validation on the field synthesis. It estimates out-of-sample error directly and needs no likelihood. Both methods get the same refit budget: 500 refits for our penalty against fold counts chosen to match total compute. Block size is tuned over a grid, and we report both the best block size, chosen by an oracle using synthetic truth, and the block size a practitioner would choose from the variogram range.
The key ablation is the same simulated-optimism penalty computed with independent noise. This isolates dependence as the active ingredient, as opposed to merely accounting for breakpoint search.
Beating the baselines means a higher probability of selecting the true generating model on synthetic fields, across noise ranges, than nominal AIC/BIC and the $n_{\mathrm{eff}}$ correction. It also means matching or exceeding block cross-validation at the practitioner-chosen block size. On selection accuracy, no reported result was found on Lacuna for any of these baselines in vegetation-cooling data. The only reported numbers are fit statistics. The Beijing study reports $R^2$ of 0.803–0.872 for the spatial autoregressive model, and the 281-city study reports $R^2$ of 0.616–0.894 for the multiscale model. These are in-sample fits and say nothing about whether selection was correct, which is precisely the gap.
Why not just…?
Why not just use BIC, whose heavier penalty already favors simpler models? BIC's penalty grows with $\ln n$ using the nominal $n$. It is therefore uniformly harsher regardless of whether dependence is present. It will suppress true thresholds when the noise is nearly independent and may still under-penalize when correlation is strong and aligned with NDVI. The synthetic sweep over noise range shows this directly. If BIC matches our criterion across the whole sweep, we will say so and recommend it.
Why not just shrink the sample size using Moran's I and keep AICc? This corrects for dependence but not for breakpoint or bandwidth search, which nominal parameter counts also miss. The independent-noise ablation and the $n_{\mathrm{eff}}$ baseline together separate the two effects. If the $n_{\mathrm{eff}}$ correction alone recovers the truth as well as our criterion on the raster data, the simpler fix wins there. We would then claim an advantage only for searched-threshold models.
Why not just use spatial block cross-validation? This is a real threat. For these small regression models refitting is cheap, so the usual cost argument for information criteria does not apply. The case for our criterion rests on two things the experiments can check. The first is that block cross-validation accuracy depends on block size, which practitioners pick ad hoc. The second is that leave-one-site-out with 77 sites, as in the phenology synthesis, gives a high-variance error estimate. If block cross-validation at the variogram-chosen block size matches our recovery rates, the honest conclusion is that the field should adopt it. Our re-analysis of published claims would then use it as the arbiter. The scientific question of which vegetation effects are real would still be answered either way.
Why not just fit random forests and read off feature importance? Tree ensembles do not test whether a threshold or sign reversal is needed. Their splits produce step functions whether or not a threshold exists. On the synthetic "no threshold" truth we will report how often partial-dependence plots from a forest display a spurious breakpoint, as a concrete demonstration.
Proposed Experiments and Decisive Checks
The synthetic recovery study is the gate. We report selection accuracy for each rule as a function of noise correlation range, separately for each of the four truths. We also report the false-threshold rate, meaning how often a threshold model is chosen when the truth is linear. Our criterion should keep the false-threshold rate near the target set by independent-noise AIC at short ranges, while nominal AIC's rate climbs with range. If nominal AIC does not degrade with range, the premise fails and the project stops there.
The real-data checks are three.
On the phenology synthesis, we test whether the 10.0 °C background-temperature threshold, the 61.7 mm/month evapotranspiration threshold and the 35.2 ha area threshold are each preferred over smooth or linear alternatives once site clustering and threshold search are priced.
On Beijing Landsat-8 imagery, we re-derive the four 2017 dates from public scenes. We then test whether the combined spatial autoregressive model still beats the simpler spatial-error model, and whether the positive winter NDVI effect survives against a model with no seasonal sign change.
On the 281-city heat-island data, we check whether the bandwidth for NDVI difference stays local when the bandwidth search is penalized for dependence.
A finding that the evapotranspiration threshold survives while the area threshold does not would be a substantive result for planners, who currently size parks partly on that 35.2 ha figure. Robustness is checked by repeating each conclusion under exponential and Matérn noise models and with block cross-validation as an independent arbiter. A claim is reported as overturned only when both our criterion and block cross-validation agree.
References
- Decoupling of the Municipal Thermal Environment Using a Spatial Autoregressive Model
- Identifying Surface Urban Heat Island Drivers and Their Spatial Heterogeneity in China's 281 Cities: An Empirical Study Based on Multiscale Geographically Weighted Regression
- Phenology acts as a primary control of urban vegetation cooling and warming: A synthetic analysis of global site observations
- The Interpolating Information Criterion for Overparameterized Models
- A Bayesian information criterion for singular models
- Principled Architecture Selection for Neural Networks: Application to Corporate Bond Rating Prediction
- Climate-sensitive Urban Planning through Optimization of Tree Placements
- UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction