Northern ecosystems that were expected to keep greening as they warmed have instead started to be held back by water. Across much of the land north of 45°N, satellite leaf area turned from greening to browning around 2000. Over the same period, the link between summer warmth and carbon uptake flipped sign. What this record cannot tell us is the direction of the key arrows. Does dry air (high vapor pressure deficit, the "thirst" of the atmosphere) shut down photosynthesis directly? Or is the effect routed through soil drying? Or do stressed canopies, by transpiring less, heat and dry the air above them, so that some of the arrows run backwards? These interactions happen within days. Monthly or 8-day satellite data therefore record them as simultaneous, and the standard tools of the field, partial correlation and its time-lagged extensions, are blind to the direction of simultaneous links.
There is a statistical way to read direction from simultaneous data. If the random shocks driving each variable are non-Gaussian (skewed droughts, heavy-tailed heat spikes), then regressing the effect on the cause leaves residuals that are independent of the cause, while the reverse regression does not. The obstacle is sample size. A single pixel offers perhaps 120–240 growing-season time steps, too few to estimate the third- and fourth-order statistics this method relies on. The usual fix of averaging pixels into regions destroys exactly the non-Gaussianity the method needs. This note proposes a different fix. We learn groups of pixels that share the same causal mechanism and pool their samples instead of averaging their signals. Within each group we orient the simultaneous links among dry air, soil water, temperature and photosynthesis, separately before and after the turn-of-century shift.
Core Claim
The claim has three parts. First, region-level averaging works against direction-finding. Excess kurtosis is a fourth-order cumulant, and cumulants add for independent sums. Averaging $m$ independent pixel shocks with excess kurtosis $\kappa$ therefore yields a shock with excess kurtosis $\kappa/m$; with spatially correlated pixels the reduction is milder and set by the effective number of independent pixels. Second, treating the pixels of a region as replicate draws from one shared causal model keeps the shocks' non-Gaussianity intact and multiplies the sample size. Third, which pixels share a mechanism is itself unknown, and can be learned jointly with the models.
Formally, for pixel $s$ in learned region $k$ and period $p$ (1982–1999 or 2000–2018), let $x_{s,t}$ be the vector of anomalies at time $t$ in vapor pressure deficit, root-zone soil wetness, air temperature, incoming shortwave radiation and a vegetation activity signal (LAI/NDVI, or SIF where available). The model is
$$x_{s,t} = B_{k,p}\, x_{s,t} + e_{s,t},$$
where $B_{k,p}$ is the matrix of direct simultaneous effects shared by all pixels in region $k$ during period $p$, and $e_{s,t}$ are mutually independent non-Gaussian shocks. Region labels carry a spatial smoothness penalty, but they are not restricted to round, smooth shapes. They can follow coastlines, wetland fringes or permafrost boundaries.
The scientific output is a map of oriented simultaneous edges with uncertainty, and a test of whether the post-2000 reversal corresponds to a change in direction, and not only in strength. One example would be the emergence of a vegetation→dry-air feedback edge, or a switch from a soil-water-mediated path to a direct dry-air path.
Relation to Prior Work
The motivating evidence comes from three partial-correlation studies.
Emerging negative impact of warming on summer carbon uptake in northern ecosystems controls for precipitation and cloudiness and finds that the summer CO2 drawdown–temperature partial correlation moved from $R=-0.40$ ($P=0.14$, 1979–1995) to $R=-0.65$ ($P<0.01$, 1996–2012).
Inhibitive Effects of Recent Exceeding Air Temperature Optima of Vegetation Productivity and Increasing Water Limitation on Photosynthesis Reversed Global Greening attributes northern browning after about 2000 to rising dry air and falling soil wetness.
Regional and seasonal partitioning of water and temperature controls on global land carbon uptake variability shows that process models and data-driven products disagree on water-versus-temperature control.
All three produce symmetric coefficients on pre-defined regions. None can say which way an arrow points. Observed strong atmospheric water constraints on forest photosynthesis using eddy covariance and satellite-based data across the Northern Hemisphere supplies a hand-built partition into synchronous and asynchronous climates, based on the sign of the precipitation–radiation correlation. We treat that partition as a serious competitor to our learned regions, not as background.
The method comes from the non-Gaussian literature. A Linear Non-Gaussian Acyclic Model for Causal Discovery proves that the effect matrix is identifiable when the shocks are non-Gaussian. Its own simulations report roughly a 10% false-positive rate for edges at 10 variables even with 10,000 samples, which shows how hungry the method is for data. Causal Discovery with Latent Confounders Based on Higher-Order Cumulants and ParceLiNGAM handle hidden common causes, but both are tested at $N\geq500$ on scalar variables. Within each region we reuse their machinery for the hidden regional weather shock.
None of these methods says where samples should come from on a spatial grid. That is the gap we fill. The closest representation-learning work is SPACY, described in Discovering Latent Causal Graphs from Spatiotemporal Data. It jointly learns smooth, round spatial factors and a latent causal graph. It does so by aggregating pixels into latent time series, assuming a Gaussian prior and a fixed causal structure over the whole record, and relying on time lags. Our design differs on each point: it pools pixels instead of aggregating them, uses the non-Gaussian shocks for identification, splits the record into periods, and targets simultaneous edges. Directed Cyclic Graph for Causal Discovery from Multivariate Functional Data does use non-Gaussian mixture noise inside a learned representation, but only for pre-defined channels and with costly sampling-based inference. Causal Mixture Models: Characterization and Discovery addresses mixtures of causal models in general. The file gives no details of its assumptions, and a spatially structured, period-split mixture of non-Gaussian models is the specific step we add.
Research Program
The model. We fit a spatially regularized mixture of linear non-Gaussian models with alternating updates. Each pixel is assigned the region whose model makes its residuals most mutually independent, with a smoothness penalty on neighboring labels. Given the assignments, each region–period effect matrix is re-estimated by ICA on the pooled samples, followed by pruning of weak edges. The number of regions is chosen by held-out residual independence rather than fixed in advance.
Confounding by shared weather. Pixels in one region share large-scale weather shocks, which act as a hidden common cause of their simultaneous anomalies. We model this shock explicitly as a single latent variable with many observed children. That structure is exactly the one-latent-component case that the cumulant method of Cai et al. resolves in closed form. Uncertainty is computed with a spatial block bootstrap, which respects the reduced effective sample size.
Data. Anomalies are computed after removing the mean seasonal cycle, because Estimating causal networks in biosphere–atmosphere interaction with the PCMCI approach shows that seasonal cycles inflate false links. The data sources are GIMMS LAI3g/NDVI3g (1982–2018), CRU TS climate and MERRA-2 root-zone soil wetness, following the greening-reversal study, plus GOME-2/TROPOMI SIF for the shorter recent window. The Tibetan Plateau marsh dataset is a stress test for sharp hydrological boundaries.
Diagnostics before modeling. Before any causal claim, we measure the skewness and kurtosis of the anomalies, at single pixels and at increasing levels of spatial averaging, and test the predicted $\kappa/m_{\text{eff}}$ decay. If growing-season anomalies turn out to be near-Gaussian, the project reports that negative result and stops. That is a useful finding in its own right, because it would mean simultaneous direction is unrecoverable from these products.
Baselines
The naive reference point is per-pixel partial correlation on anomalies, the method of the three ecology papers above. It recovers the adjacency pattern but orients nothing, so it fixes the floor of zero orientation information. It is run on identical anomalies and variables.
The direct ablation of pooling is per-pixel LiNGAM, the same estimator as ours applied to each pixel's 120–240 samples alone. Beating it means two things: higher orientation accuracy on the synthetic benchmark, and higher agreement with the reference edges described below.
The ablation that isolates our key argument is average-then-orient. It uses the same regions, but averages each region's pixels into one time series before running LiNGAM. If our claim is right, it should lose orientation accuracy as region size grows. If it matches pooling, our central argument is wrong.
The strongest fixed-region competitor is pooled LiNGAM within the synchronous/asynchronous partition of Observed strong atmospheric water constraints, and separately within standard biome classes. These are tuned with the same bootstrap and pruning thresholds.
The strongest lag-based competitor is PCMCI with partial-correlation tests, as in Runge et al. It is run on the same anomalies at monthly resolution, with contemporaneous links left as the method returns them. Its paper reports a false-positive rate "near the expected 0.01" on stationary synthetic data. Beating it means orienting edges it leaves undirected without exceeding that false-positive rate on our synthetic data.
The strongest joint-representation competitor is SPACY, run with its released settings on the same grids. No reported result found on Lacuna for SPACY's numerical F1 or latent-recovery scores, so we will measure them ourselves. For reference, the cumulant method reports F1 of 0.80–0.98 on its own small synthetic benchmark, against 0.0–0.4 for FCI. We include that method as the within-region confounder handler and do not treat it as a competitor. All methods get the same anomalies, the same compute budget and an equal hyperparameter search budget.
Why not just…?
Why not just use per-pixel partial correlation and then orient each edge with pairwise LiNGAM? This is the cheapest version of the idea. It fails on sample size: fourth-order statistics from about 200 dependent samples are noisy, and the original LiNGAM simulations already show 10% false positives with far more data. The experiment that settles this is bootstrap stability of per-pixel orientations against pooled ones. If per-pixel directions are stable, pooling is unnecessary, and we will say so.
Why not just pool within existing climate classes such as the synchronous/asynchronous split? This is the real threat. If the learned regions roughly reproduce that map, the learning step adds little beyond a confirmation. We will compare held-out residual independence and orientation agreement between the two partitions. If they tie, the contribution narrows to "pooling within known classes orients the edges," which is still new relative to the partial-correlation literature, but we would drop the claim that learned regions matter.
Why not just use time lags (PCMCI) or SPACY, which already handle spatial data? Both identify direction mainly through lagged dynamics. The dry-air/stomata/heating loop closes within days, far below the monthly or 8-day sampling interval. SPACY additionally averages pixels into latent series, which our cumulant argument predicts will flatten the non-Gaussian signal. The average-then-orient ablation and the head-to-head comparison on synthetic simultaneous-effect grids test this directly.
Proposed Experiments and Decisive Checks
The first experiment is synthetic. We build grids in the style of the SPACY benchmark, with irregular, non-convex regions that each carry a known simultaneous effect matrix, non-Gaussian shocks with controlled kurtosis, a shared regional weather shock, and a planted mid-record change in one edge's direction. Methods are scored on orientation F1, region recovery (adjusted Rand index) and detection of the direction change. Sweeps over kurtosis, region size and spatial correlation trace where pooling stops helping.
The second experiment uses real data with partial ground truth. Some edges are known physically: radiation is not caused by vegetation anomalies, and precipitation drives soil wetness and not the reverse. Orientation accuracy on these anchor edges is a checkable score.
The third experiment cross-checks against flux towers. At FLUXNET2015 northern forest sites, half-hourly data let lag-based PCMCI orient the dry-air–light-use-efficiency links. We test whether our monthly-scale simultaneous orientations agree with those high-frequency directions more often than the baselines do.
The decisive scientific check covers the two periods. If the post-2000 browning is water-driven, as the reversal studies suggest, then the learned northern regions should show a strengthened, correctly oriented dry-air→vegetation or soil-water→vegetation edge after 2000, with bootstrap intervals that exclude the pre-2000 value. A vegetation→dry-air feedback edge appearing only in the later period would be the result that partial correlation could never have shown.
References
- Emerging negative impact of warming on summer carbon uptake in northern ecosystems
- Inhibitive Effects of Recent Exceeding Air Temperature Optima of Vegetation Productivity and Increasing Water Limitation on Photosynthesis Reversed Global Greening
- Regional and seasonal partitioning of water and temperature controls on global land carbon uptake variability
- Observed strong atmospheric water constraints on forest photosynthesis using eddy covariance and satellite-based data across the Northern Hemisphere
- A Linear Non-Gaussian Acyclic Model for Causal Discovery
- Causal Discovery with Latent Confounders Based on Higher-Order Cumulants
- ParceLiNGAM: A Causal Ordering Method Robust Against Latent Confounders
- Discovering Latent Causal Graphs from Spatiotemporal Data
- Directed Cyclic Graph for Causal Discovery from Multivariate Functional Data
- Causal Mixture Models: Characterization and Discovery
- Estimating causal networks in biosphere–atmosphere interaction with the PCMCI approach