Reading Mineral Residence Times from River Carbon Records

Carbon budgets treat carbon that leaves the land through rivers as a fixed, averaged correction. Land that absorbs carbon from the air does not keep all of it. Some dissolves or erodes into rivers, part of it is breathed back out as CO2 from the water surface, and part reaches the sea. Global and regional budgets subtract this "lateral export" as a lumped scalar. Soil science meanwhile argues that how long carbon stays in the ground depends mainly on how tightly it is held by minerals: clay surfaces, iron and aluminium oxides, and soil aggregates. Soils that bind carbon strongly should release it to rivers slowly and smoothly. Soils that bind it weakly should release it in sharp, storm-driven pulses. We propose to test this with a tool from speech processing. A river's carbon flux record is treated as a signal made by passing a driving input (rain events and fresh litter mobilizing soil carbon) through a filter (the soil-mineral and channel storage of the catchment). We separate the two and ask whether the filter's time constants are predicted by the catchment's mineral properties.

If they are, the result is a mineral-conditioned transfer function. It can be carried to basins with no gauge, and it yields time-varying lateral export estimates for regional budget reconciliation in place of static fractions. It also gives an independent, data-driven check on the sorption time scales that process models such as MIMICS-BC assume but have never had measured in the field. If the filter's time constants are not predicted by mineral covariates, that is also informative. It would mean river records cannot see mineral protection, and budget corrections should not be built on that premise.

Core Claim

Source-filter decomposition treats a speech waveform as a glottal pulse train convolved with a vocal-tract resonance. Linear prediction (LP) estimates that resonance as an all-pole filter, a recursive filter whose output depends on its own recent past. What the filter cannot predict, the LP residual, is an estimate of the driving pulses. We apply the same logic to daily catchment carbon export $F(t)$, meaning dissolved plus particulate organic carbon flux at a basin outlet:

$$F(t) = \sum_{k=1}^{p} a_k(t)\,F(t-k) + g(t)\,e(t), \qquad a_k(t) = f_\theta\big(\mathbf{m},\, w(t)\big),$$

Here $a_k(t)$ are time-varying prediction coefficients, $p$ is the filter order, $e(t)$ is the excitation (mobilization pulses), and $g(t)$ is a gain. The coefficients come from a small network $f_\theta$ that sees static mineral descriptors $\mathbf{m}$ (clay fraction, a proxy for reactive Fe/Al oxides, aggregate-relevant tillage class, the paddy redox regime) and dynamic soil moisture $w(t)$. The poles of the filter translate directly into residence times of distinct storage pools.

The claim is specific and falsifiable. Across basins, the slow pole's time constant should increase with mineral binding capacity after conditioning on discharge. The excitation should concentrate into sparse, detectable events whose timing matches hydrological drivers. And a mineral-conditioned filter should predict held-out basins better than an unconditioned one of equal capacity.

Grounding in Prior Work

The mineral side supplies the hypothesis. Xiao et al. re-analyzed 132 global points and found that microbial carbon use efficiency explains little soil organic carbon (R² = 0.111) without mineral preservation. Sun et al. showed in 40-year Yangtze trials that alternating flooding and drainage, which precipitates poorly crystalline Fe/Al oxides, gave a 109.5% larger fertilizer-driven relative gain in soil organic carbon than continuous flooding. Han et al. built MIMICS-BC_v1.0 with explicit dissolved-carbon sorption. It is calibrated only on bulk carbon stocks, not on field sorption kinetics. These papers motivate the idea that mineral binding sets time scales, but none measures those time scales from a flux record. That is the gap we target.

The continuum models constrain and anchor the work. Zhang et al. (ORCHIDEE-Clateral) routes particulate carbon with fine clay sediment and estimates European lateral transfer at 47.3 ± 6.6 Tg C yr⁻¹. It finds that omitting this transfer inflates Europe's simulated sink by 4.5%. However, it treats particulate carbon as a static fraction of clay wash load and uses one virtual channel per 0.5° cell. Lauerwald et al. (ORCHILEAK) and Tian et al. (DLEM-TAC) show that export is large and changing: loading into inland waters rose 25% since preindustrial times. Both use generalized leaching rather than mineral protection. We use these models as sources of synthetic ground truth and as baselines. We do not replace them.

The budget side supplies the application. Bastos et al. define the RECCAP-2 adjustment that adds riverine export to inversion fluxes, and acknowledge that lateral flux data are sparse and mechanistically unconstrained. Friedlingstein et al. apply a lateral export correction within revisions that shrink the 2014–2023 imbalance to −0.02 ± 1.3 GtC yr⁻¹, while interannual imbalance persists. Borges et al. show European river export is a material fraction of continental exchange.

We are deliberately modest about the budget payoff. Global lateral export is small relative to the ±1.3 GtC yr⁻¹ imbalance uncertainty, so we do not claim to close the global imbalance. The target is regional reconciliation, where Zhang et al. show the correction is several percent of the sink.

The signal processing side supplies the method. Yu & Fazekas provide differentiable sample-wise time-varying LP, which avoids frame-boundary artifacts and trains stably. Prasanna & Yegnanarayana show that the Hilbert envelope of the LP residual locates excitation instants better than energy derivatives (88% vs 75% within ±30 ms). We reuse this idea to locate mobilization events. Murty et al. show the residual stays informative when additive noise destroys the envelope. They also note LP degrades under white noise, which matters because river concentration measurements are noisy. Loweimi et al. show that separated source and filter streams help downstream models, and that a fixed cepstral cutoff blurs across speakers. In our setting, the analogue of that fixed cutoff is a fixed band-pass split, which we test as an ablation.

Research Program

We first establish identifiability on synthetic data where truth is known. We run ORCHIDEE-Clateral-style and DLEM-TAC-style daily outputs, and we also run a simplified two-pool leaching model with MIMICS-BC sorption kinetics. In each case we inject known pool residence times and known storm mobilization pulses. We then ask whether the time-varying LP recovers the slow-pole time constant within a stated tolerance and detects pulse onsets. We report recovery error as a function of observation noise and sampling sparsity. This step decides whether real data are worth using: if residence times are unrecoverable at realistic sparsity, we stop.

We then fit the real records. Daily discharge comes from GRDC, as used by Zhang et al. and Lauerwald et al. Organic carbon concentrations come from GLORICH. GLORICH is irregular and sparse, which is the main practical obstacle. We therefore embed the LP recursion in a state-space model, in which the filtering step skips missing days, so the recursion runs on daily discharge-informed latent flux and is scored only where concentrations exist. Mineral descriptors come from soil texture maps. The Yangtze sites of Sun et al. provide a natural contrast between basins with alternating-redox rice-wheat systems and continuous double-rice systems. We train across European and Yangtze basins and hold out whole basins for testing.

Finally, we turn the transferred filters into daily lateral export for ungauged cells over Europe. We aggregate these to the regional totals used in RECCAP-2 style reconciliation and report both the mean and the interannual variance.

Baselines

The naive reference is a static lateral fraction. Lateral export is set to 4.7% of net ecosystem production, the European average reported by Zhang et al., and distributed in time in proportion to discharge. This is essentially what lumped budget corrections do. It is run on exactly the same held-out basins and days. Beating it means a higher Nash-Sutcliffe efficiency (NSE, one minus squared error relative to the observed variance) on held-out basin carbon flux and better reproduction of observed interannual variance.

The strongest existing method is ORCHIDEE-Clateral itself, driven identically over the overlapping European domain. It reports explaining 83% of spatial variation in European sediment discharge and 94% in water discharge, with a continental carbon delivery of 47.3 ± 6.6 Tg C yr⁻¹. No reported result was found on Lacuna for ORCHIDEE-Clateral's skill on per-basin daily or seasonal organic carbon flux against GLORICH, so we compute its NSE ourselves on our held-out split. Beating it means matching or exceeding its held-out flux NSE while producing a regional total within its uncertainty band. Matching it at a fraction of the cost, with interpretable time constants, would also count as a useful outcome.

The strongest data-driven competitor is a sequence model with the same inputs and the same hyperparameter-search budget. We use gradient-boosted trees on lagged features and an LSTM. Fortier et al. report that EcoPerceiver beat XGBoost in 12 of 15 ecosystem types on CarbonSense, but no raw numerical score was found on Lacuna. We therefore rerun both on our split rather than import a number.

Two ablations isolate the key ingredient. The first removes mineral descriptors from $f_\theta$ while matching parameter count. The second replaces time-varying LP with frame-wise stationary LP and with a fixed Fourier band-pass split into fast and slow components. The central claim fails if the mineral-free ablation matches the full model on held-out basins.

Why not just…?

Why not just use gradient boosting or an LSTM on climate, soil and discharge features? On pure prediction this is a real threat, and it may match us. But it does not produce residence times or event timings that can be checked against MIMICS-BC sorption kinetics. Xiao et al. illustrate how correlational fits across global data can point at the wrong driver. The LSTM baseline above settles the prediction question directly. If it wins clearly on held-out basins, the contribution shrinks to interpretability, and we would say so.

Why not just keep a static lateral fraction? Tian et al. show loading rose 25% since preindustrial times under changing land use and fertilizer. Sun et al. show management changes mineral retention substantially. A fixed fraction cannot follow either. The naive baseline quantifies how much this matters for regional totals. If the static fraction reproduces regional interannual variance nearly as well, the budget application is unneeded, and the scientific residence-time question stands alone.

Why not just couple MIMICS-BC sorption into ORCHIDEE-Clateral, as a process modeler would? That is the natural next step, and it is complementary rather than competing. Its weak point is the same as MIMICS-BC's: sorption parameters calibrated to stocks, not fluxes, as Han et al. acknowledge. Our fitted poles supply exactly the flux-derived time constants that such a coupling would need for validation. The synthetic stage also tests whether those constants are recoverable at all.

Why not just use a classical transfer function or unit-hydrograph model? Honestly, an all-pole LP filter is a transfer function. The difference lies in three places: coefficients that vary in time and are conditioned on minerals, differentiable training across many basins at once, and event detection from the residual envelope. The frame-wise and band-pass ablations measure how much each of these adds.

Decisive Checks

The work is decisive on four checks. The synthetic stage must recover known slow-pole time constants and pulse timing under realistic sparsity and noise. On real held-out basins, the mineral-conditioned model must beat the mineral-free ablation. The fitted slow time constant must correlate with clay and oxide proxies after controlling for discharge. In particular, alternating-redox Yangtze basins should show longer slow poles than continuously flooded ones, as the Fe/Al mechanism of Sun et al. predicts. Finally, the regional European total must fall within the ORCHIDEE-Clateral uncertainty while carrying interannual variance that the static fraction lacks. Failure on the first check ends the program. Failure on the second or third is a publishable negative result about what river records can reveal about mineral protection.