Satellites cannot measure evaporation directly, so a widely used workaround reads it off a scatter plot. Each pixel in a scene is plotted by its surface temperature against how green it is. The hot, bare upper boundary of the cloud of points (the "dry edge") marks surfaces that are barely evaporating. The cool, vegetated lower boundary (the "wet edge") marks surfaces evaporating at their potential rate. Every pixel's evaporation is then interpolated between the two edges. The weak point is that the edges depend entirely on which pixels the analyst puts in the plot. Choosing a basin rather than a county moved the fitted dry edge by up to 20 K and changed evapotranspiration by up to 66 W·m⁻² in Tian et al.. Different nested windows changed latent heat flux by 69.5 W/m² on average in Li et al.. Both papers end by saying that no objective rule for choosing the window exists.
We propose such a rule, built from two spatial correlation lengths that can be measured from the data. A semivariogram is a curve showing how different two measurements become as the distance between them grows. Its "range" is the distance beyond which they stop being related. The window must be several times larger than the correlation range of the land surface, so that it contains enough independent patches to sample both the driest and the wettest conditions. It must also be smaller than the correlation range of the weather forcing, meaning air temperature and incoming radiation, so that every pixel in the plot faces roughly the same atmospheric demand. That second assumption is what the triangle method rests on. The rule gives a feasible window for each location, or reports that no feasible window exists. That second outcome is itself a testable prediction about where the triangle method should fail.
Core Claim
The domain-size problem has an upper and a lower bound, and each is set by a different correlation length. Zheng et al. states the dilemma qualitatively: too large a window breaks uniform forcing, and too small a window collapses the extremes. Our claim is that the dilemma can be made quantitative. Let $a_s$ be the range of the land-surface fields (temperature and vegetation index) after removing elevation trends, and $a_f$ the range of the forcing fields. The window width $W$ should then satisfy
$$k\,a_s \;\le\; W \;\le\; c\,a_f,$$
where $k$ is the number of surface correlation lengths needed to sample the extremes reliably, and $c\le 1$ is the fraction of the forcing correlation length over which forcing can be treated as uniform. The lower bound comes from an effective-sample-size argument. A window holds roughly $(W/a_s)^2$ statistically independent patches, and the chance of containing a truly dry patch grows with that count. The upper bound comes from the triangle method's own assumption.
The non-obvious payoff is the upper bound. The alternative that suggests itself first is to enlarge the window until the edges stop moving. That has no natural stopping point, because Tian et al. shows the dry edge keeps steepening as more extreme surfaces enter the plot. Only a forcing-side length scale says when to stop.
Relation to Prior Work
The motivating evidence comes from the domain-sensitivity studies. Tian et al. ran the triangle method over five nested Heihe domains and found that the spatial pattern of evapotranspiration was stable (r > 0.98) while its magnitude was not. Li et al. found the same for the Priestley-Taylor variant, and compared the statistical edges with energy-balance (SEBI) theoretical edges. Neither paper measures any spatial correlation structure. We keep their testbeds and replace their administrative or arbitrary windows with measured ones.
The method comes from geostatistics, the statistics of spatially correlated fields. Wang et al. modeled soil-moisture residuals after a regression on land surface temperature, vegetation index and vegetation cover. We reuse that idea of fitting the variogram to residuals after removing a trend. It is our answer to the stationarity conflict described in the file: fitting a raw variogram over a large heterogeneous basin drifts instead of levelling off. That study used these variables as regression covariates, did not build a feature space, and did not vary the domain. The sea-ice work in Wang et al. and Wang et al. shows that a fitted range tracks a real physical patch size, here floe dimensions. It also shows that range estimates are delicate. We take that warning seriously: ranges are only estimated at lags below half the estimation window, and estimated ranges are bootstrapped. Zhang et al. supports the idea that range separates contiguous land-cover classes from fragmented ones. Note that it works on categorical maps, not continuous temperature fields.
One paper constrains the design. The HiWATER variability study argues that histogram entropy is the most stable measure of land-surface heterogeneity. If entropy alone predicted good window sizes, variograms would be unnecessary, so we test it directly.
Research Program
We first build the range estimator. We remove trends from MODIS 1 km land surface temperature and vegetation index by regressing on elevation and latitude within a large estimation region. We then fit spherical or exponential variograms to the residuals on moving tiles, which gives a map of $a_s$. We do the same for the forcing fields used by the Heihe studies, GDAS 1 km air temperature and solar radiation, to obtain $a_f$.
Next we calibrate $k$ and $c$ on a single site-season. For each pixel, we choose the window that satisfies the bounds and is centred on that pixel. Where the interval is empty, the pixel is labelled "contextually infeasible." For those pixels we report the SEBI-edge estimate instead, so that coverage stays complete and the label can be scored.
Finally we run the chosen windows through the unchanged triangle and Priestley-Taylor algorithms. Only the window changes, so any improvement can be attributed to the window choice.
Baselines
The naive reference is the fixed-window practice of the source papers. We use the five nested domains of Tian et al. and the large, intermediate and small windows of Li et al., rerun on the same MODIS and AATSR scenes and the same days of year. We report the spread across these fixed windows: up to 66 W·m⁻² in evapotranspiration and a mean of 69.5 W/m² in latent heat flux, as reported. We also report each window's error against eddy-covariance towers, which measure surface fluxes directly. The file gives no tower-validated error for the triangle method in Heihe, so we will produce those numbers ourselves. To beat this baseline, our window must have lower tower error than the best fixed window chosen after the fact, not merely the average fixed window.
The strongest heuristic competitor is edge-saturation growth: enlarge a centred window in steps until the dry-edge slope and intercept change by less than a tolerance. It needs no variograms and directly targets the extremes problem. We give it the same calibration site and the same tuning budget for its tolerance as we give $k$ and $c$.
The physical competitor is the SEBI theoretical edge, which Li et al. found differs from statistical edges by 0.11 in relative evaporative fraction at 1 km. On the Thailand network, the hybrid model ETMonitor is the strongest external reference: Zheng et al. reports a daily RMSE of 0.97 mm d⁻¹, against 1.54 mm d⁻¹ for MOD16. We will run our windowing inside the clear-sky soil moisture stress step of that pipeline. The question is whether reducing the window-induced bias moves RMSE below 0.97 mm d⁻¹ at the same towers.
Four ablations isolate the key ingredient. The first keeps only the lower bound and the second only the upper bound. The third uses raw variograms with no trend removal. The fourth is a shuffle control: pixels are permuted spatially, which destroys correlation structure while leaving histograms and entropy unchanged. If the shuffled ranges pick equally good windows, spatial structure is not doing the work.
Why not just…?
Why not just grow the window until the dry edge stabilizes? This is the most serious alternative, because it targets the lower bound directly. Its weakness is that larger domains keep adding hotter, barer pixels, and in Tian et al. the edges never clearly settle. The likely result is windows that cross forcing gradients. The upper-bound-only and lower-bound-only ablations settle it. If saturation matches the full rule on tower error, the lower bound contributes nothing new, and the claim shrinks to the forcing-range stopping rule alone.
Why not just subtract air temperature from surface temperature and use the whole basin? Normalizing by air temperature removes much of the forcing gradient, and that would weaken the case for an upper bound. It does not remove gradients in radiation or vapour-pressure demand. It also leaves the extremes-sampling problem untouched in small or homogeneous regions. We run it as an additional baseline, and also combined with our rule, where the forcing range is estimated from the temperature difference field. If normalization plus a large window ties our rule, we will report that the forcing bound is redundant under normalization.
Why not just use histogram entropy to judge whether a window is heterogeneous enough? The HiWATER variability study shows entropy is robust to shifts in the mean. However, it is identical for a clustered landscape and a scrambled one. The number of independent patches, which governs whether extremes are sampled, depends on clustering. An entropy-plateau window rule is included as a baseline, and the shuffle control tests this argument directly.
Why not just use SEBI edges everywhere? That removes the domain dependence entirely, but it requires local forcing data and differs systematically from the statistical edges. Its strength in data-rich sites is real. Our method is aimed at the regime the triangle method is used for: scenes where only the satellite data and coarse forcing are available.
Decisive Checks
We will calibrate $k$ and $c$ on Heihe, specifically the Yingke oasis tower and MODIS days 96 and 231. We then test without retuning on the HiWATER Yingke-Daman eddy-covariance stations and the six Thailand towers, using the same MODIS products.
Three results would support the claim. First, the variogram windows should give tower errors at or below the best fixed window chosen in hindsight, and below edge saturation. Second, estimates should be stable: the spread in latent heat flux across bootstrapped range estimates should be well below the 69.5 W/m² spread across fixed windows. Third, pixels labelled infeasible should have clearly larger triangle-method errors than feasible ones on held-out towers. This should hold most visibly in the early-season sparse-vegetation case (day 96), where Tian et al. saw estimates roughly double between domain groups.
The claim is refuted if the shuffled-range control or edge saturation performs as well as the full rule. It is also refuted if the infeasible label does not predict where errors are large.
References
- Impact of the Spatial Domain Size on the Performance of the Ts-VI Triangle Method in Terrestrial Evapotranspiration Estimation
- On Uncertainties of the Priestley-Taylor/LST-Fc Feature Space Method to Estimate Evapotranspiration: Case Study in an Arid/Semiarid Region in Northwest China
- Earth Observations-Based Evapotranspiration in Northeastern Thailand
- Upscaling In Situ Soil Moisture Observations to Pixel Averages with Spatio-Temporal Geostatistics
- Intrinsic mixture model-based measurement of sea ice density from a multi-look SAR image
- ESTIMATING SEA ICE PARAMETERS FROM MULTI-LOOK SAR IMAGES USING FIRST- AND SECOND-ORDER VARIOGRAMS
- Spatiotemporal Landscape Pattern Analyses Enhanced by an Integrated Index: A Study of the Changbai Mountain National Nature Reserve
- Analysis of the Spatial Variability of Land Surface Variables for ET Estimation: Case Study in HiWATER Campaign