A comprehensive meta-analysis of previous interception loss field campaigns provides an extensive archive of data to parameterise and/or validate model estimates over multiple biome types. We search for peer-reviewed articles and academic dissertations reporting rainfall interception or rainfall partitioning published before September 2021 on Google Scholar, Web of Science and China National Knowledge Infrastructure and in reference lists of identified primary studies or review papers. In this study, we mainly focus on parameters related to vegetation storage capacity and wet canopy evaporation rate and field observations of interception, precipitation and rainfall rates.

We synthesis the partitioning of incident rainfall into interception, stemflow and throughfall by trees and shrubs at the global scale. In total, 268 observational records are collected from 169 independent publications. Most of them span up to 2 years. To ensure the representativeness of the observations and minimise their inconsistencies with estimations, records are discarded if (a) the campaign lasts less than half a year, (b) they include cloud and/or snow interception, (c) they are affected by abundant epiphytes, (d) they belong to city parks or (e) they are based on insufficient measurements (less than 10 throughfall gauges and no assessment of stemflow) or fixed rain gauges. After such screening, 193 observations from 125 sites are retained for validation. The locations of experimental sites are shown in Fig. 1. All the metadata collected from literature are given in the Supplement.

Several observational datasets are used to compute I at the global scale based on a global vD–B model (Sect. 3) parameterised and constrained using the in situ data (Sect. 4). To characterise canopy cover fraction (c), global vegetation continuous field (VCF) products from the Moderate Resolution Imaging Spectroradiometer (MODIS) MOD44B and the Making Earth System data records for Use in Research EnvironmentS (MEaSURES) are selected. Both products are generated on an annual basis and provide the percentage of each grid cell covered by each of the following land cover classes: tall vegetation (i.e. tree canopies), short vegetation (i.e. non-tree vegetation) and bare ground. The MEaSURES product (Hansen and Song, 2018) is created with a bagged linear model algorithm based on surface reflectance and brightness temperature from the Advanced Very High Resolution Radiometer (AVHRR) and MODIS, covering a 35-year record from 1982 to 2016. The MOD44B product (DiMiceli et al., 2017) is retrieved from MODIS on the basis of regression tree models created using machine learning and spans from 2000 to near present. In order to have a long and consistent data series, a cumulative density function matching approach of Reichle and Koster (2004) is applied. This removes systematic differences between the two and yields a merged VCF dataset covering 1982–2019. For the period 1980–1981, the VCF of 1982 is used. Moreover, the MODIS Land Cover Product (MCD12C1) (Sulla-Menashe et al., 2019), based on the International Geosphere-Biosphere Programme (IGBP) classification, is selected to extract the spatial distribution and fractions of forest (FF; including evergreen needleleaf forests, evergreen broadleaf forests, deciduous needleleaf forests, deciduous broadleaf forests, mixed forests, and woody savannas and savannas) and non-forest (closed shrublands, open shrublands, grasslands, croplands, cropland/natural vegetation mosaics and permanent wetlands) ecosystems per pixel for validation purposes (see Sect. 5.1).

The fraction of absorbed photosynthetically active radiation (fPAR) and leaf area index (LAI) retrievals are taken from the MODIS V6 MCD15A3H product. This newest version at 500 m resolution benefits from an improved biome map from the high spatial resolution MODIS Land Cover Product (MCD12Q1), which provides an accurate parameter estimation related to vegetation structural types for three-dimensional radiative transfer formulations (Yan et al., 2016a). To discriminate between tall-vegetation and short-vegetation fractional covers and obtain representative fPAR and LAI for each of these two fractions at 0.1∘ resolution, 250 m resolution MOD44B data are used to select the values of fPAR and LAI from the “purest” high-resolution pixels; for instance, the fPAR for tall vegetation in a certain 0.1∘ grid cell is the average of the 500 m resolution fPAR values for the pixels with a fraction of tall vegetation > 98th percentile in the 0.1∘ grid cell. This work is done in Google Earth Engine, and its quality flag is used to exclude low-accuracy observations contaminated by clouds and snow. The original 4 d resolution is temporally smoothed and gap-filled based on the temporal smoothing and gap filling (TSGF) method proposed by Verger et al. (2011). A 7-year climatology is applied to fill gaps with missing data longer than 64 d. The gap-free daily time series are achieved with linear interpolation. The daily climatology of fPAR and LAI based on 2003–2007 is used for the period prior to MODIS.

Taking advantage of the complementary strengths of gauge-, satellite- and reanalysis-based data, the Multi-Source Weighted-Ensemble Precipitation (MSWEP v2.8) data (Beck et al., 2019) are selected as the precipitation forcing in this study. The climatological rainfall rate (R) during P events is also derived from the 3 h MSWEP, by taking the maximum accumulated volume over the 3 h periods at the monthly timescale. To mask out snow periods, observations of snow-water equivalent (SWE) from the European Space Agency (ESA) GLOBSNOW product (Luojus et al., 2013) are used over the Northern Hemisphere; the monthly SWE climatology product from the National Snow and Ice Data Centre (NSIDC) (Armstrong et al., 2005) is used for the Southern Hemisphere. The Priestley–Taylor-based potential evaporation (Ep) from the Global Land Evaporation Amsterdam Model (GLEAM; Miralles et al., 2011b) version v3.5a (Martens et al., 2017) is selected as a proxy of mean wet canopy evaporation (EC) for short vegetation. Natural neighbour interpolation is applied in resampling the datasets from their original spatial resolution to a common 0.1∘ global grid. An overview of all gridded datasets used can be found in Table 1.

Generally in the literature, canopy storage capacity is expressed either per unit of total area (S), canopy area (SC) or leaf surface area (SL). In most rainfall interception studies, S is assumed to be linearly related to SC and c. SC is often assumed to vary per vegetation type and is dependent on climate conditions. In nature, SC is dependent on vegetation morphological characteristics such as leaf surface area, inclination and hydrophobicity (Garcia-Estringana et al., 2010; Holder, 2013; Ginebra-Solanellas et al., 2020), as well as meteorological variables like rainfall intensity, droplet size and wind (Hörmann et al., 1996; Klaassen et al., 1996; Sun et al., 2018; Gerrits et al., 2010). It may explain why the SC values collected in previous campaigns can vary widely, from 0.35 mm (Valente et al., 1997) to 4.47 mm (Shi et al., 2010) – see Fig. 2a. The concepts of static/dynamic storage (Keim et al., 2006) and minimum/maximum storage (Xiao and Mcpherson, 2016) have been proposed to account for the storage changes driven by meteorological variables during specific rainfall events. Some studies suggest that LAI can be a valuable variable to explain the variability in S and further study their potential relation using linear (Van Dijk and Bruijnzeel, 2001b; Deguchi et al., 2006; Wallace and McJannet, 2008), nonlinear (De Jong and Jetten, 2007; Mianabadi et al., 2019) and exponential (Wallace et al., 2013) regressions. Here, we revisit the relationship between S, LAI and c over multiple ecosystems based on previous studies (Fig. S1 in the Supplement). A linear relationship between S and LAI is only found for short vegetation (r= 0.73) and coniferous forests (r= 0.60), while S shows a weak linear correlation to c only in broadleaf forest (r= 0.54–0.59). Nonlinear regressions do not show a higher accuracy than linear regressions in the prediction of S (based on either LAI or c) over any ecosystems.

As the majority of studies focus on either S or SC, c and LAI are collected to derive SL indirectly under the assumption that canopy capacity is linearly related to LAI. As such, caution should be taken in calculating SL that canopy capacity and LAI should be expressed in uniform scales, as LAI can be given in per unit of total land area or just canopy area, which often have to be deduced from the context of the study. Based on traditional statistical analysis, needleleaf forest shows larger SL with a median value 0.29 mm (95 % confidence level 0.25–0.34 mm), while within other forest types SL is similar; 0.20 (0.16–0.24), 0.18 (0.16–0.21) and 0.20 (0.18–0.22) mm are found for evergreen broadleaf forests, deciduous broadleaf forests and mixed forests, respectively (Fig. 2b). The median value of 0.23 (0.20–0.27) mm for all forest types is much larger than the 0.10 (0.08–0.12) mm found for short-vegetation plant functional types (i.e. crops, grass and shrubs). Stem storage capacity (SS) is influenced by stem density, bark surface roughness, the arrangement of twigs and leaves, and epiphytes. Large discrepancies are shown in reported studies, with a range from 0.01 mm (Návar, 2013) to 0.83 mm (Chen et al., 2013). Often, SS is obtained from an indirect, regression-based method in I simulations based on field observations (Gash and Morton, 1978; Gash, 1979; Lloyd et al., 1988). Compared to other variables like SC, which can have a large influence, the sensitivity of I to SS is fairly low (Liu et al., 2018; Ma et al., 2019), being even ignored in some early studies (Lundgren and Lundgren, 1979; Lankreijer et al., 1993). Despite the strong range of variability in the values of SS reported in past field campaigns, the median value around 0.09 mm is found for all tall-vegetation types (Fig. 2c). Reviewing the limited literature on short vegetation SS, the values from mixed crops, i.e. maize, rice and cassava (Van Dijk and Bruijnzeel, 2001a), hedgerow (Herbst et al., 2006) and thorn scrub (Návar and Bryan, 1994; Návar et al., 1999), are remarkably similar, ranging from 0.01–0.05 mm (Table S1 in the Supplement). Based on the results of this comprehensive meta-analysis, the median value is used in the execution of the global vD–B model over different vegetation types (Sect. 3), as shown in Table 2.

EC is usually estimated from the canopy energy balance or the surface water budget. A conventional method is to derive E/R from the slope of the linear regression of observed evaporation (i.e. I) against observed P (Gash, 1979; Klaassen et al., 1998; Wallace and McJannet, 2006). Alternatively, based on meteorological data (e.g. net radiation, temperature, humidity and wind speed), the Penman–Monteith equation (PM) (Monteith, 1965) is often applied to estimate EC from wet canopies, with the surface resistance being set to zero, essentially equating to the original Penman equation (Penman, 1948). The main drawback in applying PM is systematic underestimation of EC due to the underestimation of the aerodynamic conductance and, to a lesser extent, the available energy for wet canopy evaporation (Holwerda et al., 2012; Van Dijk et al., 2015). Considering that EC is driven largely by water vapour pressure deficit and aerodynamic conductance, to a smaller extent by available energy, Pereira et al. (2009, 2016) suggested that a Dalton-type equation, a simple water vapour diffusion equation determined by air wet bulb temperature, could be used to estimate EC from wet sparse canopies. In addition, EC can be optimised by minimising the squared differences between the paired simulated and observed I (Ghimire et al., 2012; Wallace et al., 2013; Fan et al., 2014). Finally, less commonly, EC can be estimated on the basis of eddy-covariance or Bowen-ratio measurements (Hörmann et al., 1996; Holwerda et al., 2012; Ringgaard et al., 2014). All these methods suffer from their own potential issues and uncertainties (Van Dijk et al., 2015).

Before comparing the EC from previous studies published in the literature, it is essential to clear their units and scale them correctly. The evaporation obtained from the PM and Dalton-type equation represents the rate per unit area of canopy cover (i.e. EC), but the value derived from regression is expressed per unit of total area (E). When it comes to estimating EC on the basis of eddy-covariance or Bowen-ratio measurements, it is important to note the influence of all components of evaporation from canopies and bare soils. Although transpiration tends to be very low during rain (Gash and Stewart, 1977), Ringgaard et al. (2014) suggested restricting this method to canopies with sufficient cover when evaporation from soils approaches zero. Here, the value of EC is obtained by dividing E by c for the studies in which only E is given. The synthesis of all these studies shows that the values of EC predicted from regression and optimisation methods have greater fluctuations (Fig. 2d), and they can be several times larger than those based on PM and other energy-balance-based methods (i.e. Dalton equation, eddy covariance and Bowen ratio). This discrepancy is recognised and critically discussed by Van Dijk et al. (2015). For tall vegetation, the median value of EC is 0.32 mmh-1 with a 95 % confidence level of 0.29–0.36 mmh-1. For short vegetation, EC exhibits large variability, from 0.09 mmh-1 for Potentilla fruticosa in China (Zhang et al., 2018) to 2.96 mmh-1 for thorn scrub in Mexico (Návar et al., 1999), and is on average slightly higher than that for tall vegetation. In addition, in terms of the ratio of EC to R, low vegetation has a higher median value and a smaller range of variability (see Fig. 2e). We note, however, that the short-vegetation data only come from eight publications (Table S4 in the Supplement). These findings seem to contradict the expectations of lower evaporation rates over short-vegetation types (see, for example, Van Dijk et al. 2015), likely due to limitations in the number of short-vegetation campaigns and the lack of representation of grasslands (in particular) where interception measurements are impractical. In those ecosystems, EC is expected to be lower due to the higher aerodynamic resistance, presenting analogous rates to those of transpiration in similar weather conditions (David et al., 2006). Based on this assumption, potential evaporation (Ep) is selected as a proxy of EC for short vegetation in the execution of the global vD–B model (Sect. 3), despite the high EC from the eight short-vegetation campaigns. For tall vegetation, the median value from this comprehensive meta-analysis of 50 studies is used, as shown in Table 2.

The validation of the global vD–B model estimates of I is performed by comparison to the 193 field I observations. We note that while the parameterisation in Sect. 4 also uses the field campaign data, the calibration is not performed per site but globally, so the comparison against the field observations to evaluate model performance appears adequate. A major challenge is the need to account for differences in forest cover between the 0.1∘ resolution grid cells and the study sites, bearing in mind that field observations are usually taken in local forest or shrubland plots, whose density may not be representative of that of the 0.1∘ resolution grid cell. For most natural forest stands, gaps exist between and within tree crowns, so standardising the pixel I estimates by c might result in an overestimation with respect to the field data. Conversely, dividing the pixel estimates by FF (instead of c) might result in an underestimation, especially when the interception experiment is carried out only under specific trees. In order to allow for a fair comparison, we explore the characteristics of the individual field campaigns (e.g. vegetation types, observed c and LAI, and throughfall measurement method). Standardisation by c is used for campaigns based on individual tree observations, when throughfall gauges are positioned beneath tree canopies only or where c exceeds FF within the pixel; for all other sites, standardisation by FF is used. The correspondence between the observed and modelled I for all sites is shown in Fig. 3.

In tall-vegetation ecosystems, both I (mmd-1) and I/P (%) generally agree well with field observations, with correlation coefficients (r) of 0.70 and 0.73, respectively. A slight underestimation is shown by the mean bias error (MBE) of -0.05 mmd-1 for I and -2.09 % for I/P. This underestimation mainly occurs for high I values associated with high-advection coastal forests (Schellekens et al., 1999; Sadeghi et al., 2015; Fathizadeh et al., 2018) (Fig. S2 in the Supplement). Similar validation results are found over different forest types (Fig. S3 in the Supplement), except for mixed forests where the performance is lower. The accuracy of estimates is strongly influenced by P (Sadeghi et al., 2015; Fathizadeh et al., 2018), which may explain some discrepancies in I and be attenuated when expressing the results as I/P (Fig. S4 in the Supplement). The slight underestimation may also relate to the assumption of one storm per rainy day in the daily application of Gash-type models. A precipitation event-scale validation can also be performed using the few field campaigns in which P and I have been reported for individual events. Figure S5 in the Supplement shows the comparison between daily estimates from the global vD–B model and event-based observations reported by Link et al. (2004) in a temperate needleleaf forest in southwestern Washington, USA, and by Chen and Li (2016) in a subtropical evergreen broadleaf forest in Taiwan, China. Here, events spanning more than 24 h are not included. These two sites are well represented due to a good consistency of pixel-based vegetation cover compared to their site-level descriptions, even though I during the largest P events is underestimated by the model, probably affected by the daily scale of our simulations. A good agreement is found between the daily estimated I/P and event-based observed I/P, and significant negative logarithmic relationships are shown between I/P and P as described by Sadeghi et al. (2015).

For short-vegetation interception, the estimated I has a good consistency with observations (r= 0.81) but shows a larger underestimation (MBE = -0.29 mmd-1). Moreover, a low correlation is found between estimated and observed I/P (r= 0.36). This lower performance is likely related to the errors derived from the modelling, measurement and validation, in addition to the limited number of short-vegetation studies. From the modelling perspective, the underestimation of EC related to the lower values of Ep (Sect. 4.2) explains the lower estimates of short-vegetation interception. Further, although the study species (e.g. shrubs, sugarcane and maize) from limited publications are defined here as “short vegetation”, they are all tall enough to fit funnels or gutters under them. Hence, these studies normally report higher I and I/P and may not be representative of global short-vegetation ecosystems, especially grasslands, that have a weaker coupling to the atmosphere and may experience shelter effects from the overstorey tall vegetation (Carlyle-Moses et al., 2010). For example, the measured I/P around 24 % in hedgerows (Herbst et al., 2006) and sugarcane fields (Fernandes et al., 2017) is of similar magnitude with that typically reported in forests and much higher than our estimates of 10.48 % and 8.56 % at these sites (Fig. 3). Waterloo et al. (1999) found grass interception was only about 4.53 % of P in Fiji, which is, in fact, slightly lower than our estimate of 6.19 %. Note as well that most observations in past campaigns come from single species of shrubs (Zhang et al., 2018) and crops (Finch and Riche, 2010; Zheng et al., 2018; Nazari et al., 2020) and that past studies have found large variability in I for different short-vegetation species, even when exposed to the same climate. For instance, Z. S. Zhang et al. (2016) reported I/P values of 29.1 % and 17.1 % for Caragana korshinskii and Artemisia ordosica in the Shapotou Desert (China). Likewise, Zhang et al. (2017) reported 24.9 % and 19.2 % for two xerophytic shrub communities (dominated by Hippophae rhamnoides and Spiraea pubescens) in the Loess Plateau. Hence, rainfall interception may have high sub-grid heterogeneity due to the large spatial complexity of biome compositions. The observed I from certain species may, therefore, not be representative of the whole grid. Finally, the average I/P values over low vegetated regions compare well with the findings by Wang-Erlandsson et al. (2014) based on a hydrological land-surface model.

The global distribution of I is shown in Fig. 4, and its seasonal-mean latitudinal variations are presented in Fig. 5. During the 40-year period 1980–2019, the estimated global average I is 73.81 mmyr-1 or 10.96 × 103 km3yr-1, accounting for 10.53 % of continental P and representing 14.06 % of continental evaporation (taking GLEAM v3.5a evaporation as reference). As expected, most (68.70 %) of I comes from tall vegetation, with a global average of 50.69 mmyr-1 or 7.52 × 103 km3yr-1; this amounts to 6.12 % of the continental precipitation, in agreement with the values reported by Miralles et al. (2011a). Although short vegetation I is estimated to be substantially lower (bearing in mind the underestimation reported in Sect. 5.1 against past field campaigns), it still accounts for 4.20 % of the continental P and has a widespread influence across most of the land surface, deserving full consideration as a separate flux.

In general, the spatial patterns of I agree well with the distribution of vegetation and precipitation. The high I volumes shown in tropical rainforests occur due to the combination of high P, dense evergreen vegetation and high evaporation rates. High values of I, expressed in percentage of P, are estimated in both tropical and boreal regions, where c can approach 100 %. Moreover, the lower rainfall rates in high-latitude regions (Fig. S6 in the Supplement) contribute to increasing I as a percentage of P by delaying canopy saturation. Tall vegetation dominates I in tropical and boreal latitudes, while the magnitude of short vegetation I can be comparable or even exceed that of tall vegetation in mid-latitudes (15–40∘ N and 20–35∘ S) (Fig. 5a and b). This relates to low forest cover coverage of croplands, grasslands and shrublands over the south of Europe and North America, southeastern Asia, southern Africa and Australia. The highest annual I/P of short vegetation is shown in African drylands and the Tibetan Plateau (Fig. 4f). Note that the fluctuations around 40–60∘ S (Fig. 5) relate to the low fraction of land in those latitudes.

The mean seasonal patterns of I are represented in a latitudinal profile (Fig. 5) and globally (Fig. 6). Overall, the seasonal variability of I follows the annual cycle of canopy cover and rainfall volumes and intensity. The global averaged I and I/P are higher during boreal summer (June–August) and lower during austral summer (December–February) (Fig. 6). The largest seasonal variations in I are found in mid- to high-latitude regions (15–60∘ N and 10–30∘ S), with the highest values in summer and lowest in winter (Fig. 5c), following the seasonal green wave (Fig. S7 in the Supplement). In tropical areas, the seasonal I is the highest in March–May, but it is rather stable throughout the seasons (Fig. 5c). However, when expressed in I/P, the latitudinal average shows higher values in June–August in middle to high northern latitude due to the increased c and in the tropics south of the Equator, i.e. Amazon and Congo forests, as a consequence of the reduced P at this time of the year (Figs. 5d and 6f). Similarly, higher I/P occurs in December–February over mid-latitude regions in the Southern Hemisphere and in the tropics north of the Equator. In mid- to high-latitude regions, characterised by high seasonal variations in vegetation cover (Fig. S7), the lower c results in both lower I and lower I/P during the dormant season (Fig. 5d).

To investigate differences in I for different ecosystems, Fig. 7 illustrates the quantile range and kernel density for different plant functional types. Model estimates are presented both per squared metre of land surface as well as per squared metre of canopy cover, and the field data from past campaigns are shown as well. The highest I is found in evergreen broadleaf forests, with mean pixel-based estimates of 362.96 mmyr-1 (per squared metre of land surface), at least 3 times larger than that for other ecosystems. This large difference relates to the high I values in tropical rainforests (Fig. 4a). Evergreen broadleaf forests is followed by needleleaf forests, deciduous broadleaf forests and mixed forests, showing similar mean I values of approximately 101.74–111.18 mmyr-1. Lower I values are found in sparsely vegetated land-use types, as expected: savannas, grasslands and croplands, and shrublands. When expressed in percentage of P, differences between plant functional types are lower. No large contrasts are found between needleleaf forests, mixed forests and evergreen broadleaf forests (all around 16.58 %–17.56 %). On the other hand, values in deciduous broadleaf forests are lower (11.91 %), approaching those in savanna ecosystems (11.27 %). The lowest I/P is found in shrublands (4.86 %), followed by grasslands and croplands (5.75 %). These pixel-based I/P values agree well with the estimates reported by Miralles et al. (2010) for needleleaf forests (16.1 %) and deciduous broadleaf forests (12.7 %) but are higher than that for evergreen broadleaf forests (10.4 %). Wang-Erlandsson et al. (2014) also arrived at a comparable I/P estimate, with 18 % in evergreen broadleaf forests, 17 % in deciduous broadleaf forests, 18 %–20 % in needleleaf forests, 9 % in savannas, and 9 %–13 % in shrublands, grasslands and croplands, but their estimated I (in mmyr-1) was generally slightly higher.

Similar differences among the different land-use types are found when I is expressed per squared metre of canopy cover, but magnitudes are larger (Fig. 7). This canopy-level interception is also overall comparable to previous studies. For instance, Miralles et al. (2010) found a higher canopy-level I/P in forests; however, their reported I/P per square metre of forest of 21.8 % in needleleaf forests agrees well with our study. The estimated annual I and I/P per land cover type are further compared to the reported values in field campaigns. Notice that the global estimated I is lower than the in situ measurements, except in evergreen broadleaf forests. In terms of I/P, the estimates in forests overall agree well with the field data, which indicates the average forcing precipitation might be lower than the observed precipitation from forest experiments. Both measured I and I/P over short vegetated regions are much higher than the global estimates, which is consistent with the findings in field validations. In fact, the higher observed interception from short vegetation and deciduous broadleaf forests seems reasonable, as most of observations are taken in the growing season or the leafed period (Fathizadeh et al., 2018), while our estimates are the average of both the growing season and the dormant season.

The global multiyear (1980–2019) mean annual I estimated by the global vD–B model is 73.81 mmyr-1, accounting for 10.32 % of P. This value is within the range of other global estimates, e.g. the 64.06 mmyr-1 and 7.91 % of P reported in the Community Land Model (CML) version 5 (Lawrence et al., 2019) and the 115 mmyr-1 and 13 % of P found by Wang-Erlandsson et al. (2014) based on a hydrological land-surface model. In addition, this I/P is comparable to that of 10.08 % reported by Zheng and Jia (2020), whereas the magnitude of I is much higher than their finding (57.06 mmyr-1). This large difference suggests that forcing rainfall can bring large uncertainties, which has also been found in CML5 when driven by different precipitation datasets (Lawrence et al., 2019).

The spatial patterns are also compared with two global interception products: PML v2 and GLEAM v3.5a (Fig. 8). PML v2 is based on the same vD–B model but with different parameterisations (Y. Zhang et al., 2016; Zhang et al., 2019). Overall, annual I is in good agreement with PML v2 estimates, with a high correlation coefficient of 0.91, but higher globally with a mean difference of 21.84 mmyr-1, especially in tropical regions. GLEAM v3.5a used the version of the model proposed by Valente et al. (1997) and used the same precipitation forcing as in this study; hence both I and I/P are compared here bearing this dependency in mind. In general, our interception estimates are slightly higher than GLEAM v3.5a, with a mean difference of 7.89 mmyr-1 for I and 1.71 % for I/P. In terms of spatial discrepancies, GLEAM v3.5a estimates are higher over Amazon forests and boreal forests, while they are lower in Africa, southeastern Asia and Australia. Differences in spatial patterns between both datasets (r= 0.82 and 0.67 for I and I/P, respectively) largely come from the fact that only forest interception is estimated in GLEAM v3.5a; moreover, the phenological dynamics are not explicitly considered in GLEAM v3.5a. In addition, we validate the results of PML v2 and GLEAM v3.5a against in situ data and compare the validation results to those of our new model formulation – see Fig. 9. Compared to PML v2 and GLEAM v3.5a, both estimated I and I/P in this study have the highest correlation coefficients and lowest mean bias errors with field observations. In evergreen broadleaf forests, similar validation results are found for estimated I, while PML v2 shows the highest correlation coefficient for I/P (Fig. S8 in the Supplement). However, PML v2 significantly underestimates both I and I/P in evergreen broadleaf forests, especially for large events. Different from the constant of E/R in PML and the empirical relationship between R and lightning frequency in GLEAM v3.5a (Miralles et al., 2010), here the use of 3 h temporal resolution MSWEP precipitation enables a more realistic estimation of monthly averaged R (Fig. S9 in the Supplement), which may be partly responsible for the higher model accuracy.