Few-Bit Alarms for Dryland Carbon-Uptake Crashes

Year-to-year swings in how much CO₂ the land absorbs are driven mostly by a small number of droughts in dry grasslands and savannas, and these are the places with the fewest instruments. This proposal asks a design question for the cheap, low-power sensor networks that could fill those gaps. Suppose each sensor can send only one to four bits per season. Can a central server still tell whether a carbon-uptake crash is happening? We argue it can, provided the sensors share a random seed. The shared seed lets every sensor compress its local seasonal anomaly onto the same randomly chosen directions, so no bits are spent saying which direction was used. The directions are drawn from the anomaly patterns that vegetation models consider plausible, not from all directions equally.

The opportunity comes from two literatures that have never been connected. Carbon-cycle work shows that 7% of the global land-time domain explains 78% of the interannual variation in gross primary production (GPP, total plant photosynthesis), in spatially contiguous, drought-driven "craters" (Zscheischler et al.). Semi-arid regions supply 57% of detrended GPP variability (Zhang et al.), and Africa and other drylands are badly under-observed (Weber et al.; Dolman et al.). Information-constrained statistics proves a useful asymmetry. Testing whether data depart from a reference is far cheaper under a bit budget than estimating the whole distribution, but only when nodes share randomness (Acharya, Canonne, Tyagi). Shared randomness is also strictly needed once there are more nodes than local dimensions (Szabó, Vuursteen, van Zanten). We put the two together for a detection problem whose structure, spatially clustered anomalies with season-dependent sign, the theory has not yet handled.

Core Claim

Some dryland and transitional anomalies are seasonal contrasts rather than simple mean drops. A wet-season deficit can be partly offset by a late-season recovery, and a warm spring gain can be cancelled by a hot summer loss; the latter is the seasonal compensation that Wang et al. identify as the reason data-driven and process models disagree. For such anomalies, a sensor that reports only the sign of its seasonal mean is nearly blind. A sensor that reports "my local anomaly energy is high" is optimal only while the number of sensors is below the local dimension. Our claim has three parts. First, in the regime where sensors outnumber local dimensions, shared random sign projections recover close to the detection power of a centralized test at one to four bits per sensor per season. Second, drawing those projections from a model-ensemble anomaly subspace, mixed with a small isotropic component for safety, beats isotropic projections when sensors are few. Third, the gain over local methods appears at roughly the sensors-equal-dimensions elbow that theory predicts. A further prediction keeps us honest: for one-signed mean drops, our scheme should not beat the simple mean-sign bit, and if it does, something is wrong.

Mechanism

Each sensor $i$ forms a whitened seasonal anomaly vector $z_i \in \mathbb{R}^d$. It does this by subtracting its own climatology and rescaling by its climatological covariance, using monthly values of GPP proxy, soil moisture and temperature over a season, so $d$ is roughly 12 to 36. The shared seed assigns sensor $i$ to a group $g$ and a direction $u_g$ drawn from $\mathcal{N}(0,\Sigma_\lambda)$. Here $\Sigma_\lambda = (1-\lambda)P + \lambda I/d$, where $P$ projects onto the leading whitened anomaly modes of the TRENDY vegetation-model ensemble and $\lambda\in[0,1]$ sets how much isotropic coverage we keep. The sensor sends $b_i=\operatorname{sign}(u_g^\top z_i)$, a single bit. With $\ell$ bits it sends signs for $\ell$ independent directions. For a candidate spatial cluster $C$, the server computes

$$T_C=\sum_{g}\Big(\tfrac{1}{\sqrt{|g\cap C|}}\sum_{i\in g\cap C}(2b_i-1)\Big)^2-G_C,$$

where $G_C$ is the number of groups meeting $C$. The statistic has mean near zero under no anomaly. It grows when cluster members share a displacement that has a component along the groups' directions.

The server scans $T_C$ over contiguous clusters of many sizes, because craters follow a power-law size distribution (Zscheischler et al.). Thresholds are set by resampling whole years, not by i.i.d. theory. Spatial correlation and temporal autocorrelation in real geophysical data break the independence assumptions of Acharya et al. and the Gaussian-equivalence conditions of Vuursteen, so we calibrate empirically and report the effective number of independent sensors alongside nominal counts.

Relation to Prior Work

The method comes from three related results, all for unstructured alternatives. Szabó et al. prove the minimax rate $\rho^2\asymp\min\{\sqrt{dm}/n,\,d/n\}$ for one-bit Gaussian signal detection: local chi-square tests are optimal when machines $m\le d$, and common random directions are required when $m>d$. Random flattening (Acharya, Canonne, Tyagi) and domain compression (Acharya, Canonne, Han, Sun, Tyagi) hash large discrete alphabets into few bins with shared coins while preserving distance. Domain compression adds the exchange rate that one communication bit is worth two shared random bits.

We reuse their core move, coordinating compression through a public seed. We change three things. The alternatives are spatially clustered and drawn from a physically informed prior rather than worst-case. The data are continuous and autocorrelated, which addresses the continuous-extension gap that domain compression leaves open. The null is calibrated empirically. Vuursteen bridges discrete testing to the many-samples-per-node Gaussian regime but provides no implementation. Our emulation is a first test of whether those large-sample predictions survive on real flux records.

On the carbon side, Zscheischler et al. motivate the target and supply the event definition we reimplement as ground truth. Wang et al. supply the seasonal-compensation anomaly class. Zhang et al. show that ensemble members must be weighted against observations rather than averaged, which is why our prior is a mixture with an isotropic floor rather than the ensemble mean. Liu et al. constrain the design: CMIP6 models miss the observed 35% intensification of tropical water-carbon coupling, so a prior built purely from models could be confidently wrong. We therefore test performance off the model subspace explicitly.

We are not proposing to fix inversion artifacts such as the Amazonian dipoles in Molina et al.. Those come from atmospheric transport, not bit budgets. They motivate the general idea that answering a testing question can avoid the instability of estimating everything from sparse data.

One honest caveat on the application. Existing eddy-covariance towers are not bit-limited. The regime we target is dense, low-cost dryland sensor deployments on power- and uplink-limited radios. The resulting curve of bits against detection power is useful as a network-design tool even before such networks exist, because it says how much a coarse report retains.

Research Program

We build an emulation testbed and then a theory-to-data check. For the emulated dense networks, each grid cell of the four 1982–2011 GPP data cubes is treated as a sensor: two empirical products, MTE and MOD17+, and two process models, LPJmL and OCN (Zscheischler et al.). Each cell is paired with CRU/GPCC precipitation and temperature. We restrict to semi-arid masks and subsample to network sizes $m$ from 20 to 2000.

Ground-truth crater labels come from reimplementing the 3D connectivity event detection on the full-resolution cube. For the prior, whitened anomaly modes are fit on TRENDY outputs (Wang et al.) and evaluated on the empirical cubes. This train-on-models, test-on-observations split is deliberate. For independent observations, GOME-2 SIF anomalies for 2007–2015 (Zhang et al.) serve as a second test domain, including the 2015/16 El Niño (Impact of the 2015/2016 El Niño).

For real point data, FLUXNET2015 sites with at least ten years of record (Liu et al.) test whether site-level whitening and the sign channel behave as the gridded emulation predicts under real measurement noise and gaps. We vary $d$, $m$, the bits per sensor $\ell\in\{1,2,4\}$, the mixing weight $\lambda$, and the number of shared random bits, using the domain compression exchange rate to compare equal-resource configurations.

Baselines

The reference point is a centralized oracle. It computes the full-precision whitened scan statistic with every sensor's complete $z_i$, giving the power ceiling that every few-bit scheme is reported against as a relative efficiency. The naive one-bit reference is the mean-sign bit: each sensor reports whether its seasonal mean GPP anomaly is negative, and the server runs a clustered sign test. This is what a practitioner would deploy first.

The local-optimal baseline is the local energy bit: each sensor thresholds $\|z_i\|^2$. Szabó et al. prove this optimal when sensors are fewer than local dimensions. The private-coin baseline has each sensor project on its own random direction without coordination, matching the setting whose sample-complexity penalty Acharya et al. quantify.

The strongest existing method is isotropic shared-randomness sign projection, the super-critical scheme of Szabó et al. adapted to clusters. The only published figure attached to it is its theoretical minimax rate above. For empirical detection power on FLUXNET2015 or gridded GPP, there is no reported result found on Lacuna, so we will establish its number ourselves.

The most dangerous practical baseline is deterministic round-robin over a fixed basis. Each sensor reports signs of fixed TRENDY leading modes, with modes assigned to sensors in rotation. This captures coordination without randomness.

Our key ingredient is isolated by sweeping $\lambda$ from 1 (isotropic) to 0 (pure model subspace) at matched bits. All schemes share the same sensors, years, whitening, cluster scan and resampled 5% false-alarm threshold. All scheme parameters, $\lambda$ and group sizes, are tuned on 1982–1999 and evaluated on 2000–2011 plus SIF years. Beating the strongest baseline means a paired, year-block-bootstrapped gain in detection power at 5% false alarm of at least five points for seasonal-contrast craters at $m\le d$. We also require no loss larger than two points anywhere on the $m$ grid and no loss when the anomaly lies off the model subspace.

Why not just…?

Why not just send a few quantized deterministic summaries, like the seasonal mean or fixed leading-mode coefficients? For one-signed crashes, which are the dominant crater type by Zscheischler's asymmetry, the mean-sign bit is likely near-optimal, and we predict our scheme will tie it there. The case for randomness rests on seasonal-contrast anomalies and on robustness. Deterministic coding loses distance information that randomized coding preserves (Acharya et al.). A fixed basis has permanent blind spots wherever the true anomaly is orthogonal to it, which is exactly the risk Liu et al. raise for model-derived structure. Round-robin is a real threat for average-case anomalies. The off-subspace test settles it: if round-robin matches us there too, we will report that shared randomness buys only worst-case insurance.

Why not just let each sensor randomize privately or threshold its own energy? Private coins cost a factor up to $\sqrt{k/2^\ell}$ in samples for testing (Acharya et al.). Local energy tests plateau once sensors outnumber dimensions (Szabó et al.). Both claims are asymptotic and assume independence. Our $m$-sweep checks whether the predicted elbow appears on correlated real anomalies; if spatial correlation erases it, local energy bits are the right design.

Why not just centralize the data and run machine-learning upscaling such as XGBoost or FLUXCOM? Where bandwidth allows, centralization is our oracle ceiling, not a competitor. Under a few-bit budget the raw data never arrive. Even centrally, tree models underpredict extreme NEE and degrade under extreme heat and dryness in grasslands (Liu et al.). They also need roughly 2.5 to 5.4 years of daily data to perform well, which new dryland sites lack. FLUXCOM misattributes the main driver because it misses seasonal compensation (Wang et al.), the very anomaly class we target.

Decisive Checks

Five results would confirm or kill the idea. First, the efficiency gap between shared-randomness and local-energy bits should open near $m\approx d$ in the emulation, consistent with the Szabó et al. elbow. Second, on one-signed crashes, mean-sign bits should match our scheme. Third, on injected and observed seasonal-contrast anomalies, mean-sign power should collapse while ours stays near the oracle. Fourth, with the prior fit on TRENDY and tested on MTE, MOD17+ and SIF, a pure model subspace ($\lambda=0$) may fail, and an intermediate $\lambda$ must recover isotropic-level power. Fifth, on FLUXNET2015, whitening and calibration must hold the false-alarm rate at 5% despite autocorrelation.

If the second check fails, or the fourth shows no safe $\lambda$, the structured prior is not trustworthy. We would then fall back to reporting the bits-versus-power design curve for isotropic and deterministic schemes alone.