Does Simulated Hydraulic Failure Explain Satellite-Observed Biomass Loss in the Amazon?

Satellites that measure low-frequency microwave emission from forests show that the Amazon's standing forests are losing carbon even where nobody cuts them down. Some of that loss comes from logging, edges and fire, and some may come from trees dying of drought. No one can currently say how much of each, and that split determines whether the Amazon sink is a land-use problem or a climate problem. Separately, vegetation models now simulate how drought drains water from a tree's roots, stem and leaves until the water-conducting tissue fails and the tree dies. Those models have only been tested at a single experimental site.

This proposal tests the drought explanation against the satellite record. For each roughly 25 km pixel and year, we take the model's predicted hydraulic damage and ask whether it predicts satellite-observed carbon loss one to two years later. The test is a partial correlation, meaning the correlation that remains after removing what three groups of variables already explain: the raw climate anomalies, the human-disturbance covariates, and the forest's same-season water status. If the predicted damage still carries signal after all three are removed, that is direct continental-scale evidence for drought-induced mortality in the satellite record. It also gives an observational constraint on mortality parameters that are currently calibrated at one site. If no signal survives, the model's drought-mortality pathway lacks support from the satellite record. Either outcome is informative.

Core Claim

The model in question is ORCHIDEE-CAN-NHA, a global vegetation model with an explicit tree-hydraulics module. It converts weather into stem "percentage loss of conductance" (PLC), the fraction of the stem's water transport capacity lost to air bubbles blocking the water-conducting vessels. This quantity is a strongly nonlinear, threshold-shaped, size-dependent transform of climate.

Our claim is that a cumulative-PLC exposure index from this model is an unconfounded, lagged predictor of aboveground carbon loss as measured by L-band vegetation optical depth (L-VOD). L-VOD is a passive microwave signal at about 1.4 GHz that penetrates dense canopies and tracks woody biomass and vegetation water content. We further claim the index carries information that the climate indices it is built from do not.

The design has two parts. The hydraulic model supplies the nonlinearity, so the statistic applied on top can stay a simple, interpretable rank partial correlation. Time lags supply the separation between temporary dehydration and actual death. Trees that are only dehydrated rewet within a season, while dead trees stay gone. So a drop in L-VOD that persists one to two years after a high-PLC year, and is not explained by concurrent soil water, is the signature of mortality.

Define, for pixel $i$ and year $t$, $\Delta A_{i,t+k}$ as the L-VOD-derived carbon change $k$ years later and $P_{i,t}$ as the cumulative PLC exposure. Let $Z_{i,t}$ be the confounder set: maximum climatological water deficit (MCWD), vapor pressure deficit (VPD), temperature, terrestrial water storage, human-disturbance fractions, and same-season soil moisture. The central quantity is

$$\rho_k = \mathrm{corr}_{\text{rank}}\big(\Delta A_{i,t+k} - \widehat{\Delta A}(Z_{i,t}),\; P_{i,t} - \widehat{P}(Z_{i,t})\big),$$

computed on pixel-demeaned anomalies. We predict three things. First, $\rho_1$ and $\rho_2$ will be negative and significant. Second, $\rho_0$ will be largely absorbed once soil moisture is in $Z$. Third, $|\rho_k|$ will grow with standing biomass.

Prior Work and Its Role Here

Work that motivates the gap. Satellite-observed pantropical carbon dynamics used SMOS L-VOD to show pantropical degradation losses are about twice deforestation losses. Global carbon balance of the forest extended the record to 2019 and found wet tropical forests are a net loss of $-0.07 \pm 0.02$ PgC/yr. Both attribute the loss jointly to "logging, fire, and climate-induced mortality" without separating the three. Human influence on Amazon's aboveground carbon dynamics disaggregated L-VOD to 100 m and estimated degradation loss at $-0.42 \pm 0.08$ PgC over 2010–2020. It states explicitly that it cannot distinguish human disturbance from drought mortality.

That product also constrains our design. Its downscaling uses climate covariates from ERA5-Land and CHIRPS, so correlating its output with climate would be partly circular. We therefore work with native-resolution SMOS-IC L-VOD and represent sub-pixel land use as fractional covariates.

Work that supplies the mechanism. The ORCHIDEE-CAN-NHA model description simulates organ-level water potentials and triggers cohort mortality after sustained high PLC. Its exposure threshold and daily mortality fraction are calibrated only at the Caxiuanã throughfall-exclusion experiment. How drought events during the last century have impacted biomass carbon in Amazonian rainforests ran the model from 1901 to 2019 and identified the 2015/16 El Niño as the largest single-event loss. Future drought-induced tree mortality risk projects sink-to-source switches in the Guiana Shield and the East-central Amazon. None of these papers is evaluated against a satellite biomass time series. We reuse their historical runs as the source of $P_{i,t}$.

Work that supplies the statistical lens. Emerging negative impact of warming on summer carbon uptake used partial correlation, controlling for precipitation and cloudiness, to expose an emerging warming penalty north of 50°N ($R = -0.65$, $P < 0.01$ for 1996–2012). Regional and seasonal partitioning of water and temperature controls used partial correlation to explain why inversions and process models attribute global flux variability to water storage while the data-driven FLUXCOM product attributes it to temperature. Both studies work on bulk net flux at zero lag and never touch standing biomass or mortality. We move the lens to structural carbon loss, add lags, and replace raw climate predictors with a mechanistic hydraulic index.

A paper that pulls the other way. Field-based tree mortality constraint found that vegetation models overestimate tropical mortality by up to 5 Mg ha⁻¹ y⁻¹ against 2,676 plots. Its finding appears to contradict the satellite story of large standing-forest losses. Our design can partly adjudicate this, because the regression slope behind $\rho_k$ converts hydraulic exposure into a carbon-loss magnitude that we can compare against both sources.

Research Program

Building the hydraulic exposure index. We rerun the ORCHIDEE-CAN-NHA historical simulation for 2010–2020 on the SMOS grid, using the forcing configuration of the 2022 historical study. From it we extract annual cumulative days of stem PLC above threshold, weighted by cohort biomass. A size-resolved version keeps the tallest cohorts separate, because the model description and the historical study both show tall trees dominate hydraulic mortality.

Building the response. Annual aboveground carbon comes from SMOS-IC L-VOD. Following the African SMOS-IC evaluation, we prefer SMOS-IC because it correlates more strongly with biomass maps ($R = 0.85$–0.94) than the other retrievals. We compute the response two ways: from annual-median L-VOD, and from wet-season L-VOD only. The wet-season version minimizes the dehydration artifact identified in the global L-VOD balance paper.

Building the confounders. Climate covariates are MCWD and precipitation from CHIRPS, temperature and VPD from ERA5-Land, and GRACE terrestrial water storage. Human-pressure covariates are per-pixel fractions of new deforestation, new degradation and edge length from the tropical moist forest (TMF) land-cover product used by Fendrich et al., plus a burned-area fraction. Water status is captured by same-season soil moisture and storage.

Estimation. We compute rank partial correlations on pixel-demeaned anomalies, separately within the three Amazon regions named in the projection study. Uncertainty comes from a spatial block bootstrap, which resamples contiguous blocks of pixels so that neighboring pixels are not treated as independent. A semi-partial variant, which removes $Z$ from only one of the two variables, partitions the unique variance in carbon loss attributable to hydraulic exposure, following the use in K. He et al..

The secondary deliverable: recalibrating the mortality parameters. We sweep the model's drought-exposure duration threshold and its daily mortality fraction. The calibration objective is the out-of-sample lagged partial correlation and slope. We calibrate on 2010–2014 and test on the 2015/16 El Niño and its aftermath, so the event that matters most is never used for fitting. This gives the first observation-based, multi-region estimate of the parameters that are currently tied to Caxiuanã.

Baselines

The naive reference is a bivariate rank correlation between the MCWD anomaly and lagged L-VOD carbon change, with no conditioning. This is how drought impacts are usually read off satellite budgets, and it sets the floor. There is no reported result on Lacuna for this correlation in the Amazon, so we will report it ourselves on identical pixels and years.

The climate-only partial correlation is the strongest statistical competitor. It uses the same conditioning set and the same lags, but replaces the hydraulic index with MCWD, VPD or storage directly, each partialled on the others. It is a direct extension of the method in the two Wang et al. papers. It is the real threat to the proposal, because the hydraulic index is itself a function of these inputs.

To keep the comparison fair, both versions share the pixels, years, confounder set, bootstrap blocks and lag windows. The climate-only version also gets nonlinear expansions of its inputs, namely threshold exceedance days and splines. This gives it the same chance to capture threshold behavior. The hydraulic index beats this baseline only if it adds out-of-sample partial correlation beyond the expanded climate set, and the gain survives the block bootstrap. The out-of-sample check uses the held-out 2015/16 event.

The strongest existing attribution system is the Fendrich et al. disaggregation model. It reports an Amazon degradation loss of $-0.42 \pm 0.08$ PgC and a net loss of $-0.37 \pm 0.17$ PgC for 2010–2020. It reports no attribution split between drought and human disturbance, and no reported result was found on Lacuna for such a split. Our comparison is a reconciliation rather than a competition. The drought share implied by our fitted slopes, plus the human share, must sum to within its degradation uncertainty. That is a consistency test of both approaches.

Two ablations isolate the key ingredients. The first removes the lag, using only $\rho_0$, to test whether persistence is what separates mortality from dehydration. The second replaces size-resolved PLC with a stand-mean PLC, to test whether the tall-tree mechanism matters.

Why not just…?

Why not just correlate drought indices with satellite biomass change? Climate drivers co-vary too strongly. The northern-latitude study found that the warming effect on carbon uptake only became clear once precipitation and cloudiness were controlled for. In the Amazon, drought years also coincide with fire and with increased clearing, so a raw drought correlation absorbs human losses. Our naive-baseline run will show directly how much the drought–biomass relationship changes once human-pressure fractions are partialled out.

Why not just use the climate indices in a partial correlation and skip the vegetation model? This is the honest threat, and the experiment is designed to settle it. If threshold-expanded MCWD and VPD match the hydraulic index out of sample, the mechanistic model adds nothing at this scale, and we will report that. We expect the index to win for two reasons. First, it integrates soil supply and atmospheric demand through tree size, which no single index does. Second, it predicts a specific interaction, a stronger effect in tall, high-biomass forests, that the climate indices do not encode.

Why not just use a random forest with feature attributions? Zheng et al. show that random forests can upscale mortality from plots. However, the Wang et al. global-flux analysis shows that the data-driven FLUXCOM product misidentified the dominant driver of flux variability. With about ten years per pixel and strongly collinear inputs, a flexible learner will spread credit arbitrarily across correlated features. We will still run a gradient-boosted model with the same inputs as a check. If its attribution to hydraulic exposure disagrees with $\rho_k$ in sign, we treat the result as unresolved.

Why not just restrict the analysis to intact-forest pixels and call any loss climatic? At 25 km, few pixels are truly intact. Restricting to them also biases the sample toward remote, wetter forest, which is less exposed to drought. We run this as a robustness check, using pixels with at least 95% intact forest cover, and require the sign of $\rho_k$ to match the full-sample result.

Proposed Experiments and Decisive Checks

The main test estimates $\rho_0$, $\rho_1$ and $\rho_2$ by region, comparing the hydraulic index against the threshold-expanded climate baseline, with calibration on 2010–2014 and testing on 2015–2018. The proposal's central claim fails if $\rho_1$ and $\rho_2$ are not significantly negative once human pressure and soil moisture are conditioned out.

The dehydration check has a clear prediction. The concurrent-year effect $\rho_0$ should shrink sharply when same-season soil moisture enters $Z$, while $\rho_1$ should not. It is also run on the wet-season-only response. If the lagged signal disappears in wet-season data, the "mortality" was rehydration noise.

The size check stratifies by baseline biomass, using benchmark biomass maps from Saatchi, Avitabile and ESA CCI to define the strata. It tests whether $|\rho_k|$ increases with biomass, as tall-tree hydraulic vulnerability predicts. A flat or reversed gradient would argue against the hydraulic mechanism even if the overall correlation is significant.

The site anchor runs the recalibrated mortality parameters back at Caxiuanã. They should still reproduce the reported pattern of large-cohort PLC exceeding 80% in the dry year 2005, and the observed mortality under experimental drought. A continental parameter set that breaks the one site where mortality was actually observed would be rejected.

The magnitude reconciliation converts the fitted slopes into a basin-wide drought-mortality carbon loss for 2010–2020. It then places that figure against the Fendrich degradation total and against the plot-based mortality constraint from Zheng et al. The question is which of the two the satellite-consistent hydraulic signal sides with.