Tibetan Plateau biomass maps rewire by elevation: local R² rises 0.048, but beyond 30 km the gated graph falls behind
Synopsis
In aboveground biomass estimation on the southeastern Tibetan Plateau, a spatial graph built from geographic nearest neighbors links valley forest to adjacent alpine meadow, because horizontal closeness does not imply ecological similarity. The study first quantifies these misconnections, then rebuilds the graph by keeping only edges whose endpoint elevation difference falls below a threshold, raising local-interpolation R² from 0.362 to 0.409 across 6,638 samples and ten seeds. Under a strict 30 km buffer holdout the same construction scores 0.054 lower than the ungated graph, so the gain depends on training neighbors remaining near the test points.
Interpretation
Spatial graphs no longer assume that nearby means similar; each edge's elevation and biomass differences are checked before the edge is kept. The standard construction connects each sample to its k nearest neighbors by coordinates, implicitly assuming horizontal distance stands for target similarity, an assumption that breaks in terrain spanning 146 to 5,383 m. On 66,380 directed edges over 6,638 clean samples at k=10, 21.41% span elevation differences above 800 m and 12.24% span biomass differences above 70 Mg/ha, and 60.6% of the high-biomass-difference edges connect different land-cover types.
The filter uses one rule: keep an edge only if its endpoint elevation difference is within a threshold τ, and drop the rest. This corrects the topology at construction time rather than patching the model architecture, and elevation acts as a constraint on edges rather than as an extra input feature the model must learn from. Under a random split with ten seeds, τ=250 m retains 60,698 edges with 6 isolated nodes and reaches R²=0.409±0.023, against 0.362±0.019 for the ungated geographic k-nearest-neighbor graph, a paired difference of 0.048 (p=6.0×10⁻⁷, 95% CI [0.039, 0.056]) with all ten seeds in the same direction.
Under strict spatial extrapolation the same filter reverses sign, turning gain into loss. A random split leaves training neighbors near test points so the graph can pass local knowledge along; a buffer split removes all training samples within 30 or 50 km of each test point, cutting that channel. The gated graph scores 0.054 lower than the ungated graph at a 30 km buffer (p=0.027) and 0.083 lower at 50 km (p=0.007), with all ten seeds agreeing, while the local support index is zero at 30 km under the same protocol.
Perspective
The conclusions rest on one region, the southeastern Tibetan Plateau, a single 2021 snapshot and a fixed sampling density with a 1.19 km minimum spacing; the 250 m threshold is an empirical value for that region and density. The supported setting is local interpolation inside the sampled domain, such as within-region densification and gap filling; under strict extrapolation the gated graph does not outperform the ungated one.
The authors state that τ=250 m is specific to this region and sampling density and that cross-region transfer needs multi-region validation; with a GraphSAGE backbone the improvement is not significant (−0.001, p=0.642), so the gain depends on the message-passing mechanism. The biomass labels are not field measurements but a derived product from random-forest upscaling of radar and lidar reference data followed by spatial aggregation, so part of the gain could reflect alignment with the label-smoothing pattern.