Finding Missing Grazers in the Residual of a Mass-Balance Bottleneck

Rangeland carrying-capacity models estimate how much forage a landscape grows and subtract what livestock eat. When they leave out wild ungulates, rodents, insects, or plant disease, the books appear to balance while the grass is actually being overgrazed. A clear case is the Yellow River Source Park. There the annual budget is balanced under a livestock-only count, but adding wild ungulates turns it into critical overload, and the winter withering season into severe overload (Seasonal Herbage–Livestock Balance). Interpretable machine learning has a close analogue in the concept bottleneck model. This is a network that must first predict a small set of human-named intermediate quantities and then predict the output only from those quantities (Concept Bottleneck Models). When a needed concept is missing from that list, accuracy drops or the network sneaks the missing information through the bottleneck.

We propose to treat these as the same problem. We would build a bottleneck model whose final layer is the rangeland accounting identity itself rather than a learned linear map. Its named concepts are the physical terms of the budget: forage grown, livestock intake, wild herbivore intake, and known pest loss. It also has one deliberately narrow, uncertainty-aware residual channel that absorbs whatever depletion the named terms cannot explain. The claim is that, with leakage controlled and each named term carrying calibrated uncertainty, the spatial and seasonal structure of this residual is a usable detector of unaccounted biological stressors. The size of the residual channel is also the knob that traces the accuracy–interpretability trade-off in this domain.

Core Claim

Inspecting the residuals of a fitted budget is not new. What is new is being able to say when a residual means "a stressor is missing" rather than "one of the known terms was misestimated." A plain residual cannot make that distinction. Wild herbivore intake comes from sparse transects extrapolated by a habitat-suitability model (Survival Risk Analysis for Four Endemic Ungulates). Pest loss enters as a uniform literature multiplier. Forage production comes from a light-use-efficiency model. All three carry large, spatially uneven error.

We borrow the probabilistic concept representation of Probabilistic Concept Bottleneck Models. That paper maps each concept to a Gaussian rather than a point, and showed that concept-level variance tracks ambiguity and is useful for ranking where humans should intervene. With this, every budget term gets a per-pixel variance. A missing stressor is then flagged only where the residual is large relative to the combined uncertainty of the named terms, and where it is spatially coherent.

Formally, for location $s$ and season $t$, observed standing forage $B(s,t)$ from field clipping is modeled as

$$B(s,t) = N(s,t) - L(s,t) - W(s,t) - P(s,t) - R(s,t) + \varepsilon,$$

where $N$ is forage grown, $L$ livestock intake, $W$ wild herbivore intake, and $P$ known pest loss. Each is predicted as a Gaussian by a network reading satellite and climate rasters. $R$ is the residual channel and $\varepsilon$ is measurement noise. The subtraction is fixed, not learned, so every prediction can be decomposed exactly into named, auditable terms.

Research Program

The model ingests stacked rasters. These are NDVI time series (GIMMS 3g and MODIS MOD13A3), monthly temperature and precipitation, elevation, land-cover classes, and county livestock statistics, all as assembled for Xinjiang in Monitoring and influencing factors of grassland livestock overload in Xinjiang. For the Tibetan Plateau we add the transect-derived ungulate densities and the 30 m ecosystem maps.

Each concept head is weakly supervised by whatever exists. $N$ is supervised by CASA estimates, $L$ by livestock inventories converted to sheep units (a sheep unit is the daily intake of a 50 kg ewe), and $W$ by the transect-plus-habitat-model densities. $P$ is supervised by the literature multiplier, with wide prior variance. Supervision follows the "sequential" regime of the original bottleneck paper, so that the concepts keep their meaning.

The residual channel is the methodological core, and we constrain it three ways. First, it reads only a deliberately coarse input, such as a low-resolution embedding with a spatial-smoothness prior, so it cannot memorize pixel-level noise. Second, it pays a capacity penalty $\lambda \lVert R \rVert$. Third, we add a decorrelation penalty between $R$ and the predicted named terms. This penalty is the guard against the leakage failure of joint bottleneck training, in which the backbone silently encodes target information outside the concepts.

Sweeping $\lambda$ yields a curve. At one end is a pure accounting model, where $R$ is forced to zero: fully interpretable but possibly inaccurate. At the other end is a near-black box in which $R$ explains everything. The scientific object of study is where on that curve the residual stops being noise and starts carrying coherent, seasonally structured depletion.

Seasonality follows the phenological periods and nutrient dimensions used in the Yellow River Source Park work. There, crude protein rather than energy is the binding constraint in the withering season. We therefore fit separate budgets for dry matter and digestible crude protein, which lets an ecologist intervene on one stressor concept, such as raising rodent severity, and see the effect on both budgets.

Baselines

The naive reference point is the standard livestock-only accounting model: CASA forage minus livestock intake, with no wildlife or pest terms. It represents the "false sense of balance" that motivates the work. The only reported number here is for CASA's forage estimate on its own: Monitoring and influencing factors of grassland livestock overload in Xinjiang reports $R^2 = 0.746$ against 617 field biomass samples. Our model must at least match this on forage production when evaluated on the same samples and splits.

The strongest existing domain method is the "realistic" accounting used for the Tibetan Plateau and the Yellow River Source Park. It subtracts transect-derived wild ungulate intake and a literature pest multiplier (Survival Risk Analysis for Four Endemic Ungulates; Seasonal Herbage–Livestock Balance). No reported result on Lacuna was found for its predictive error on held-out field biomass, so we will compute it ourselves on identical splits. Beating it means lower held-out error on standing forage and fewer false "balanced" classifications in the seasons where field data show overload.

The strongest machine-learning competitor is an unconstrained network. We will use the same backbone, inputs, and training budget, regressing standing forage directly. On image benchmarks such networks are only slightly more accurate than bottleneck models: on CUB-200-2011, the black-box error was 17.5% against 19.9% for the joint bottleneck (Concept Bottleneck Models). However, they degrade more under distribution shift, rising to 62.7% error against roughly 48–49% on CUB + Places in the same paper. We count it as beaten if our model is within its error in-region and better under spatial transfer from Xinjiang to the Tibetan Plateau.

The key ablations isolate each ingredient. The first uses deterministic concepts without variances: same architecture, but missing-stressor flags are thresholded on residual magnitude alone. It tests whether the probabilistic representation matters. The second drops the decorrelation penalty and tests whether leakage contaminates the residual. All variants get the same hyperparameter search budget.

Why not just…?

Why not just fit the standard budget and map its residuals against field biomass? This is the most serious alternative, and our main ablation essentially is it. Its weakness is identification. A positive residual in the eastern Tibetan Plateau could come from missed rodents or from underestimated wild donkeys. Without per-term uncertainty, the two cannot be told apart. The omission test below is designed to settle this: if raw residuals recover a held-out stressor as well as our uncertainty-normalized residual does, the bottleneck adds nothing for detection. We would then reduce the contribution to its intervention and transfer properties.

Why not just use a Bayesian hierarchical accounting model with priors on each term? Such a model would also propagate uncertainty, and on county-level data it may do as well. The difference is that it cannot read raw rasters to produce per-pixel concept estimates, whereas a learned bottleneck can. We will run it as an additional baseline at county resolution. If it matches our detection performance there, the claim narrows to the gain from spatial resolution, which we test at the pixel level against the 617 Xinjiang samples.

Why not just train a black-box network and explain it with SHAP or saliency maps? Post-hoc attributions are not tied to the model's actual computation, and they break down when inputs are correlated. Sum-of-Parts Models proposes self-attributing feature groups for exactly this reason and reports 0.267 error against 0.306 for the per-feature baseline SHAP-F on ImageNet-S. Rangeland drivers are strongly entangled; precipitation and elevation jointly dominate overload patterns in Xinjiang. Moreover, an attribution over input pixels never yields a quantity in sheep units that can be compared with a budget term. We include this baseline in the omission test to show whether its attributions localize the held-out stressor at all.

Proposed Experiments

The decisive check is a controlled omission test on the Tibetan Plateau, where wild ungulates are measured. We train the model with the $W$ head removed, so the wildlife is "unaccounted," and ask whether the residual $R$ recovers it. Recovery would mean $R$ correlates with the transect-based wild-intake map and reproduces the reported south–east high-pressure versus north–west refuge pattern (Survival Risk Analysis for Four Endemic Ungulates). We score this by spatial rank correlation and by precision and recall of flagged cells against high-wildlife-pressure cells, compared across raw residuals, deterministic residuals, the hierarchical model, and SHAP maps.

A seasonal version repeats the test in the Yellow River Source Park. Removing wildlife should make the residual concentrate in the greening and withering periods, and in the crude-protein budget more than the dry-matter budget, matching the reported seasonal overload pattern.

Only if the omission test succeeds do we apply the model to the truly unmeasured case. We restore all known terms and map the remaining residual, then ask whether it aligns with land-cover classes and elevations associated with rodent colonies. We treat any such map as a hypothesis for targeted field sampling, not as a finding.

Across all runs we report the accuracy–interpretability curve as $\lambda$ varies. Accuracy is held-out forage error, both in-region and transferred across regions. Interpretability is measured as the fraction of variance carried by named terms, residual–concept correlation as a leakage measure, and intervention gain, meaning how much held-out error falls when an expert overwrites the most uncertain concept. That last measure directly parallels the knee-grading result, where correcting two concepts cut error from above 0.4 to about 0.3.

The main practical risk is data access. The field and transect data are described in the cited papers but may not be public. If they are unavailable, the omission test can still be run on Xinjiang by withholding the livestock term and checking recovery against county inventories.