Recent global decline in rainfall interception loss due to altered rainfall regimes

Rainfall intercepted by vegetation (canopies, stems, litter) evaporates back to the atmosphere as interception loss (Ei), directly influencing how much rainfall reaches the soil to support vegetation growth or generate runoff. While site-level throughfall–stemflow studies show Ei can range widely (∼10–50% of rainfall) depending on vegetation and rainfall characteristics, these measurements are sparse and difficult to scale to regional or global levels. Eddy-covariance (EC) towers provide ecosystem-scale latent heat flux (LE) but do not directly partition Ei. Consequently, global estimates have relied on process or physically based models (e.g., Gash model) with simplifying assumptions and limited observational constraints. This study aims to develop an observationally constrained, globally upscalable approach to estimate Ei using hybrid physics–machine learning models trained on EC observations under wet versus dry conditions. The research questions are: (1) Can Ei be inferred robustly from EC LE by contrasting wet and dry conditions? (2) What are the primary global drivers of Ei and its fraction of rainfall (Ei/P)? (3) How has Ei changed globally over the last two decades in response to changes in rainfall regimes? Addressing these questions is important for constraining water partitioning between interception, soil moisture recharge, and runoff under a changing climate.

Site-based studies report large Ei variability tied to plant functional type (PFT), leaf area index (LAI), rainfall regime, and evaporative demand, but they lack scalability due to spatial heterogeneity of rainfall events and canopy structure. The Gash analytical model partitions interception into moistening, saturation, and drying phases but relies on assumptions (e.g., complete canopy drying between events) that may not hold universally. Empirical Ei formulations often scale linearly with rainfall and canopy cover/LAI, lacking explicit canopy energy and water budget constraints, which can bias estimates especially where rainfall intensity dominates interception dynamics. A previous EC-based method estimated Ei as excess evaporation during/after rain relative to a radiation-scaled dry baseline, validated at a single Amazon site, but generalizability across climates and ecosystems remained unproven. These gaps motivate a physics-informed, data-driven framework that leverages EC networks, Earth observations, and reanalysis to better constrain Ei at large scales.

Data and preprocessing: EC measurements were drawn from FLUXNET2015 Tier 2, using half-hourly to hourly aggregated variables: LE, H, air temperature (Ta), precipitation (P), vapor pressure deficit (VPD), net solar radiation (Rn), ground heat flux (G), wind speed (WS), plus site metadata (PFT, canopy/tower height). After quality filtering (exclude poor-quality or negative LE, Ta < 0°C to avoid snowfall interception, and PFT representativeness checks), 76 sites across 8 PFTs remained. LE was corrected for rainfall/high-humidity bias via a site-specific neural network predicting latent energy ratio (LER = LE/(R−G−H)) as a function of relative humidity (RH) and log(P); LE was then adjusted to a reference LER at RH=50% and P=0. A Bowen ratio method was applied to address residual energy balance non-closure. Wet/dry classification and event definition: Wet hours were defined within identified rain events: start when hourly P ≥ 0.5 mm; include the following 6 hours (12 h if nighttime) after rain cessation to capture post-rain drying; extend events if P ≥ 0.5 mm occurs within the trailing window; split events longer than 60 h into sub-events starting with P ≥ 1 mm separated by ≥6 h. Non-rain intervals (P=0) within events were allowed and consecutive rainy hours separated by non-rain intervals were treated as pulses. Dry hours were all hours outside wet events. This yielded 146,608 wet samples from 29,985 events and 287,764 dry samples. Hybrid physics–ML models: Two physics-constrained hybrid models (feedforward NN fused with a quadratic Penman–Monteith constraint) were trained at hourly scale: HMdry on dry samples and HMwet on wet samples. Inputs common to both: Ta, Rn, VPD, WS, LAI, and categorical PFT. HMwet additionally used a latent predictor, canopy water storage (CWS), inferred via a neural network from vegetation attributes and eight rainfall descriptors within events. The hybrid architecture predicts surface resistance (Rs) in log-space, with the PM-based loss enforcing energy balance conservation and diffusion physics. HMdry, trained without wet conditions, estimates the baseline LE from transpiration and soil evaporation. HMwet estimates LE under wet-canopy conditions (transpiration + soil evaporation + interception evaporation). The interception flux Ei is inferred as LE(HMwet) − LE(HMdry) at hourly resolution. Upscaling to global grids: Models were driven globally (0.5°×0.5°) for 2000–2020 using ERA5 hourly meteorology (Ta, VPD, Rn, WS, P, G), MODIS LAI (MOD15A2H, 8-day, interpolated hourly), and MODIS IGBP PFT (MCD12C1). Wet hours on grids were identified by scanning the preceding 60 h P time series. Grid CWS was approximated from LAI, PFT, and rainfall event characteristics, then normalized using statistics from 5°×5° windows around flux sites to align scales between site-based and grid-based CWS. All other predictors were normalized using site-based means/SDs with range checks. Predicted Ei was multiplied by vegetated cover fraction (MOD44B) to exclude non-vegetated surfaces. Treatment of light rain outside events: Because coarse grids include substantial light rain (P < 0.5 mm) not classified as events, Ei/P for light rain was linearly interpolated between full interception as P→0 (approximated by vegetated fraction) and the modeled Ei/P at P≈0.6 mm (0.5–0.7 mm) for each PFT; this ratio was used to compute Ei under light-rain conditions outside events. Validation and comparisons: Event-level Ei/P from the hybrid approach was compared against 10 geographically proximate in situ interception datasets (community-level) across four PFTs, yielding strong agreement (r=0.76, p<0.05). Global Ei and Ei/P were compared with GLEAM (Gash-based) and an ensemble of six TRENDYv7 LSMs (2000–2018), with evaluation across five Köppen–Geiger climate zones. Attribution metrics: Rainfall frequency (Frain) was defined as the yearly fraction of wet hours with 0 < P < local 90th percentile of rainy-hour P; rainfall intensity (Irain) as the yearly fraction of hours with P exceeding the local 90th percentile. Trends were computed for 2000–2020, excluding areas with significant precipitation trends when attributing Ei trends to Frain/Irain changes.

Model performance: The hybrid models reproduced hourly LE at sites under wet and dry conditions with r² > 0.75 and RMSE ≈ <60 (units as in plots). With identical predictors except CWS, LE(HMwet) exceeded LE(HMdry) by ~42% at hourly and ~32% at event scales, reflecting the additional interception evaporation during/after rainfall. The modeled Ei response exhibited physically consistent timing: small during rainy hours with low radiation, peaking a few hours after rainfall when both canopy water and energy are available. Event-level Ei/P and drivers: Event Ei/P spanned ~0–100% with median 18.3% and mean 24.4%, consistent with a global meta-analysis. Forest biomes generally had higher Ei/P than non-forest; evergreen needleleaf forests showed the highest mean Ei/P (~32%), while deciduous broadleaf forests were lowest among forests (~19%). Ei/P was primarily controlled by rainfall characteristics: it decreased with increasing average and maximum hourly rainfall rates; roughly three-quarters of events with mean rain rate >1.0 mm h−1 had Ei/P <5%. Relationships with vegetation attributes were weak at event scale: Ei/P vs LAI was positive but weak overall (r≈0.13) and often insignificant within PFTs; wind speed effects were mixed, with a significant negative correlation for evergreen broadleaf forests. Global magnitude and spatial patterns (2000–2020): Mean global Ei was 84.1 ± 1.8 mm yr−1, accounting for 8.6 ± 0.2% of gross land rainfall. Ei spatial patterns followed precipitation and LAI climatology, peaking in wet, densely vegetated tropics. Ei/P increased with P in relatively dry regions (P < 800 mm yr−1) but decreased with P in humid regions (P > 800 mm yr−1), reflecting a trade-off between vegetation cover and higher rainfall intensity/shorter duration in humid climates. Model intercomparison: The data-driven Ei exceeded GLEAM (60.2 mm yr−1) and the LSM ensemble mean (73.9 mm yr−1; range 18.9–100.3 mm). Both GLEAM and LSMs showed higher Ei/P in humid tropics and lower in arid/boreal zones than this study, suggesting an overreliance on canopy cover/LAI and underrepresentation of rainfall intensity effects. Trends and controls (2000–2020): Despite negligible global precipitation trend (+0.81 mm yr−1, p>0.1), global Ei declined significantly by 4.9% (−0.20 mm yr−1; p<0.01), and Ei/P declined by 6.7% (p<0.01). Rainfall frequency decreased (Frain −2.3%, p<0.01) and intensely raining hours increased (Irain +2.3%, p=0.09). Interannually, Ei correlated with Frain (r=0.55, p<0.05) but not with Irain; Ei/P correlated negatively with Irain (r=−0.61, p<0.05) but not with Frain, indicating that frequency controls average Ei magnitude while intensity controls partitioning. Spatially, Ei decreased across most tropical (74.2%) and >50% of extratropical lands, mainly where rainfall became less frequent; Ei/P decreased where rainfall became more intense (e.g., Amazon, E Africa, India) and increased where intensity lessened (e.g., W Australia, E Europe).

The study demonstrates that interception loss can be robustly inferred from EC observations using a physics-constrained hybrid ML approach that contrasts wet and dry conditions. By explicitly learning wet-canopy dynamics (via a canopy water storage proxy) while enforcing energy-balance constraints, the framework yields observationally constrained, spatially explicit Ei estimates. The findings reveal that rainfall regime characteristics dominate Ei behavior at large scales: frequency governs the average Ei volume, while intensity governs the fraction of rainfall partitioned to interception. The observed global shift toward less frequent and more intense rainfall since 2000 explains a significant decline in Ei and Ei/P, implying a reallocation of rainfall toward soil moisture and runoff. This partitioning shift may benefit ecosystem water availability and human water use, but it also concentrates excess water during intense events, increasing flood risk. The results challenge model parameterizations that scale interception mainly with LAI or canopy cover and underscore the need to represent sub-daily rainfall variability and intensity effects. The framework provides an empirical benchmark to evaluate and improve land-surface and evaporative flux models, especially in humid regions where intensity-driven limits on Ei/P are most pronounced. The implications are significant for forest and water management: interception must be considered when assessing hydrologic impacts of deforestation (potentially increasing available water) and afforestation (reducing water yield via interception and transpiration).

This work provides an observationally constrained, globally upscaled estimate of rainfall interception loss using hybrid physics–ML models trained on EC data under wet versus dry conditions. Ei averages 84.1 mm yr−1 globally (8.6% of rainfall) for 2000–2020, with Ei/P largely regulated by rainfall intensity rather than vegetation attributes. Since 2000, less frequent and more intense rainfall has driven a significant global decline in Ei (−4.9%) and Ei/P (−6.7%), implying increased partitioning toward soil moisture and runoff. The approach offers a new benchmark to evaluate models such as GLEAM and LSMs, which appear to underweight rainfall intensity effects. Future research should: (1) better represent sub-daily rainfall variability and intensity in land models; (2) integrate slowly varying vegetation changes (e.g., LAI dynamics, structural traits) into hybrid frameworks; (3) quantify interactions among interception, transpiration, and soil evaporation pulses after rainfall, particularly in water-limited systems; and (4) incorporate interception explicitly in assessments of afforestation/deforestation impacts on regional water resources and flood risk under projected future intensification of rainfall extremes and reduced frequency.

Key limitations include: (1) the assumption that dry-period relationships extrapolate to wet conditions; wet leaves may reduce transpiration by stomatal blockage, potentially biasing Ei low during storms; (2) confounding of Ei with post-rain pulses of soil evaporation and transpiration in water-limited biomes; (3) difficulty of ML models to capture slowly evolving vegetation effects (e.g., LAI/structural changes), leading to a weaker apparent LAI influence; (4) limited spatial/temporal representativeness and period mismatch in validation against in situ interception measurements; (5) treatment of sub-grid and light-rain variability relies on assumptions (linear Ei/P scaling for P<0.5 mm); (6) potential uncertainties from LE corrections for rain/humidity bias and energy balance closure; and (7) differences between open- and closed-path EC systems and reanalysis/satellite forcing errors. These factors may affect the magnitude and spatial patterns of Ei and Ei/P, especially in climates with complex rainfall regimes and in strongly water-limited regions.