Many remote sensing algorithms infer continuous environmental variables by constructing a "feature space" from satellite observations and identifying its geometric boundaries. A prime example is the Land Surface Temperature–Vegetation Index (LST-VI or Ts-VI) triangle method used to estimate terrestrial evapotranspiration (ET). By plotting surface temperature against vegetation cover, observations typically form a triangular or trapezoidal distribution. The boundaries of this shape represent hydrological extremes: a "dry edge" corresponding to zero evapotranspiration (water-stressed, bare soil) and a "wet edge" corresponding to potential evapotranspiration (well-watered, dense vegetation).
The core vulnerability of this approach lies in a simple geometric truth: the boundaries of a scatter plot depend entirely on the data points included in the sample. When researchers define a spatial domain—the geographic bounding box of the analysis—they implicitly determine the range of land surface conditions sampled. If the domain is too small or homogeneous, it fails to capture the true physical extremes of the region, leading to a misidentified feature space and distorted downstream measurements.
Why the Dry Edge Shifts with Scale
To understand why domain size alters downstream calculations, consider how the dry edge is constructed. As detailed by Tian et al., automatic algorithms typically identify the upper envelope of the LST-VI space to define the dry edge. In a small spatial domain, the probability of capturing a completely dry, non-vegetated pixel experiencing maximum thermal stress is low, particularly during the early stages of the growing season when dense vegetation is sparse and temperature ranges are compressed.
Large Domain (Heterogeneous) Small Domain (Homogeneous)
LST ^ LST ^
| * . . [True Dry Edge] | . . [Truncated Edge]
| * * . . | * * . .
|* * * . . | * * * . .
+-------------------> +------------------->
VI VI
As the spatial domain expands, the diversity of the land surface increases. The larger area is more likely to encompass extreme surface conditions, which introduces higher maximum surface temperatures into the dataset. Consequently, the intercept of the dry edge shifts, and its slope typically becomes steeper. Tian et al. demonstrated that while the overall spatial pattern of estimated ET remains relatively consistent across scales, the absolute magnitude of the ET estimates varies significantly due to this boundary sensitivity.
Uncertainties in Heterogeneous Landscapes
This scale dependency introduces substantial uncertainty when applying feature-space methods across different geographic regions. Li et al. evaluated this effect by applying a remote sensing-based Priestley-Taylor method across nested windows of varying sizes in an arid/semi-arid region. The authors confirmed that the derived dry and wet boundaries fluctuate based on the chosen analysis area, directly impacting the reliability of ET calculations. Because an optimal, universal spatial scale for constructing the LST-VI feature space remains undefined, boundary definitions inherently conflict when moving between small, homogeneous patches and large, heterogeneous landscapes.
Recognizing these limitations, Zheng et al. noted that the domain dependence of the LST-VI method poses a significant challenge for very large-scale or global applications. When the analysis window becomes too vast, the assumption of uniform meteorological forcing across the feature space breaks down; conversely, when it is too small, the edges collapse. This trade-off suggests that large-scale monitoring may require alternative algorithms designed to be entirely independent of domain size.
A Pervasive Challenge in Spatial Modeling
The sensitivity to spatial range is not unique to thermal-vegetation feature spaces; it reflects a broader mathematical challenge in spatial correlation and resolution modeling. For instance, when measuring sea ice density from high-resolution synthetic aperture radar (SAR) images, Wang et al. found that spatial correlation models are highly sensitive to the accurate definition of the correlation range and the characteristic size of discrete objects. Just as an undersampled LST space miscalculates a climate boundary, an incorrectly defined spatial range in image processing degrades the precision of feature density measurements because it misjudges the distance at which pixels cease to be statistically correlated.
Furthermore, the scale at which data is aggregated introduces its own layer of uncertainty. Sun et al. highlighted that global ET products exhibit varying uncertainty levels depending on the spatial resolution at which they are evaluated. Whether adjusting the bounding box of an empirical feature space or aggregating pixels to a uniform grid, the spatial scale chosen by the researcher fundamentally dictates the statistical behavior of the underlying model, making scale optimization a critical prerequisite for accurate environmental monitoring.
Key Foundational Work
Impact of the Spatial Domain Size on the Performance of the Ts-VI Triangle Method in Terrestrial Evapotranspiration Estimation (2013) — contributed 10 concept elements to this direction.
Intrinsic mixture model-based measurement of sea ice density from a multi-look SAR image (2014) — contributed 2 concept elements to this direction.