The moisture content of vegetation canopies controls various ecosystem processes such as plant productivity, transpiration, mortality, and flammability. Leaf moisture content (here defined as the ratio of leaf water mass to leaf dry biomass, or live-fuel moisture content, LFMC) is a vegetation property that is frequently used to estimate flammability and the danger of fire occurrence and spread, and is widely measured at field sites around the globe. LFMC can be retrieved from satellite observations in the visible and infrared domain of the electromagnetic spectrum, which is however hampered by frequent cloud cover or low sun elevation angles. As an alternative, vegetation water content can be estimated from satellite observations in the microwave domain. For example, studies at local and regional scales have demonstrated the link between LFMC and vegetation optical depth (VOD) from passive microwave satellite observations. VOD describes the attenuation of microwaves in the vegetation layer. However, neither were the relations between VOD and LFMC investigated at large or global scales nor has VOD been used to estimate LFMC. Here we aim to estimate LFMC from VOD at large scales, i.e. at coarse spatial resolution, globally, and at daily time steps over past decadal timescales. Therefore, our objectives are: (1) to investigate the relation between VOD from different frequencies and LFMC derived from optical sensors and a global database of LFMC site measurements; (2) to test different model structures to estimate LFMC from VOD; and (3) to apply the best-performing model to estimate LFMC at global scales. Our results show that VOD is medium to highly correlated with LFMC in areas with medium to high coverage of short vegetation (grasslands, croplands, shrublands). Forested areas show on average weak correlations, but the variability in correlations is high. A logistic regression model that uses VOD and additionally leaf area index as predictor to account for canopy biomass reaches the highest performance in estimating LFMC. Applying this model to global VOD and LAI observations allows estimating LFMC globally over decadal time series at daily temporal sampling. The derived estimates of LFMC can be used to assess large-scale patterns and temporal changes in vegetation water status, drought conditions, and fire dynamics.
Changes in water availability and the occurrence and severity of droughts
affect various processes in land ecosystems and vegetation (Konings et al., 2021a; Sippel et al., 2018). For example, soil moisture and atmospheric water
demand affect plant water uptake, the water potential, and water content of
vegetation, stomatal conductance, and transpiration (Jarvis, 1976; Bonan, 2015). The regulation of plant water content and stomatal conductance controls the exchange of water, carbon, and energy between the ecosystem and atmosphere. Hence, soil moisture and atmospheric water demand are strong controls on plant productivity, growth, and mortality (W. Li et al., 2021;
McDowell, 2011; DeSoto et al., 2020). Furthermore, the water content of living or dead vegetation material controls the occurrence and intensity of
disturbances such as fires. A low water content of the fuel is associated
with higher flammability or a higher risk of fire occurrence and spread
(Chuvieco et al., 2010). Hence, fuel moisture content (FMC) is a key variable to estimate daily to long-term changes in fire danger (Stocks et al., 1998; Jolly et al., 2015). FMC is defined as the ratio of the mass of water to the dry biomass of a material and can be directly measured from determining the fresh and dry mass
In order to complement site observations of LFMC, satellite observations can be used to estimate LFMC over large areas (Yebra et al., 2013). Thereby, especially satellite observations in the visible and infrared domain of the electromagnetic spectrum have been used to estimate LFMC. For example, spectral information from the short-wave and near-infrared bands from Landsat are correlated with LFMC (Chuvieco et al., 2002; Bowyer and Danson, 2004). This is because leaf water content has a strong effect on the absorption of near and shortwave infrared radiation. Hence, LFMC can be computed by using empirical models or visible-infrared leaf and canopy radiative transfer models by estimating equivalent water thickness (EWT, i.e. leaf water column per unit area) and the leaf dry matter content (Danson and Bowyer, 2004; Riano et al., 2005). Medium and coarse resolution visible-infrared satellite instruments are most commonly used to estimate LFMC, as they provide a frequent temporal coverage (García et al., 2008; Yebra et al., 2008). For example, observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) are the main input for recently developed algorithms to estimate LFMC at continental or global scales (Yebra et al., 2018; Quan et al., 2021; Zhu et al., 2021). Despite the direct biophysical relations between surface reflectance and LFMC and their implementation in visible-infrared radiative transfer models, the occurrence of cloud cover, smoke, or low sun elevation angles hinders the retrieval of LFMC time series with high temporal frequency from visible-infrared satellite sensors.
Microwaves can largely penetrate clouds and smoke, are independent of the
illumination by the sun, and hence provide an alternative to derive information about the land surface. Microwave observations from either
active radar instruments or from passive microwave radiometers are sensitive
to the moisture content of soil and vegetation (Ulaby et al., 1979; Jackson et al., 1982) and hence also to FMC (Konings et al., 2019).
For example, early studies have shown that fuel moisture conditions are
related to the radar backscatter of C-band (5.3 GHz frequency
Similar relations between the microwave signal and vegetation water content
are valid for observations from passive microwave instruments that measure
naturally emitted microwaves from the Earth surface. Passive microwaves are
emitted by the soil and vegetation and are then attenuated in the vegetation
layer (Jackson et al., 1982; Mo et al., 1982). Passive microwave instruments are commonly used to estimate surface soil moisture (Njoku and Entekhabi, 1996; Njoku et al., 2003; Wigneron et al., 1998, 2021; Dorigo et al., 2017). Recently, surface soil moisture datasets from passive microwave sensors have also been used as a proxy to estimate LFMC (Jia et al., 2019; Lu and Wei, 2021). However, passive microwaves are also directly related to LFMC. The attenuation of the passive microwave signal in the vegetation layer is commonly described by the opacity or optical thickness of the vegetation (VOD, vegetation optical depth) (Jackson and Schmugge, 1991; Frappart et al., 2020). VOD is proportional to vegetation water content (VWC, i.e. mass of water per unit area) and hence to the dry biomass (
Despite this direct theoretical relationship between LFMC and VOD, which has
been established from field observations, there is currently no study that
verified this relationship at large (i.e. continental to global) scale. This
implies that no method exists that would allow to estimate LFMC from
VOD at large scales. However, the use of novel VOD datasets with almost
daily temporal coverage and data available partly since 1987 (Moesinger et al., 2020; Wang et al., 2021) offers the opportunity to estimate LFMC globally with high temporal resolution and over decadal timescales. In comparison to visible-infrared satellite observations, the main disadvantage
of using VOD to estimate LFMC is the coarser spatial resolution of passive
microwave data (usually
Here, we aim to estimate LFMC from VOD at large scales, i.e. globally at coarse spatial resolution and at decadal timescales. Therefore, we will use VOD from short wavelengths from the VOD Climate Archive (VODCA) dataset, which provides consistent time series of Ku-VOD, X-VOD, and C-VOD harmonized from VOD retrievals from different passive microwave satellites (Moesinger et al., 2020). We first investigate the relation between VOD and LFMC by comparing VOD with an LFMC dataset from MODIS (Yebra et al., 2018) and with the Globe-LFMC database of site observations (Yebra et al., 2019). In the second step, we develop different model structures to compute LFMC from VOD, and we calibrate each model against site-level observations from the Globe-LFMC database. Finally, we apply the best-performing model globally to estimate and analyse LFMC at large scales.
An overview of the properties of all used datasets is provided in Table 1.
Properties of the used datasets.
VOD was taken from the VODCA dataset (Moesinger et al., 2020). VODCA provides VOD at
The VODCA dataset has a daily temporal sampling, but observations are not
available for each day dependent on time period and latitude. The VODCA dataset was masked for artefacts because of radio frequency interference (RFI), for land surface temperature
Spatial patterns, seasonal dynamics, and long-term trends in the VODCA dataset have been intensively compared with datasets of leaf area index, gross primary production, and vegetation cover, and show that VODCA reflects common patterns of large-scale vegetation changes (Moesinger et al., 2020; X. Li et al., 2021; Wild et al., 2022).
LFMC was taken from two datasets, namely from the Globe-LFMC database of site (Yebra et al., 2019) and from LFMC data retrieved from MODIS satellite observations by applying the methodology of Yebra et al. (2018) (in the following MODIS-LFMC). The Globe-LFMC measurements are the primary dataset for the comparison with VOD and to develop and calibrate the models to estimate LFMC from VOD. However, as there is severe scale mismatch between site measurements of LFMC and the coarse spatial resolution of VOD
(
Globe-LFMC provides LFMC field measurements from 1383 sites in 11 countries, mainly in the USA (963 sites), China (229 sites), Spain (76 sites), and Australia (42 sites). However, each site has a different temporal coverage and sampling frequency of measurements, and different plant species are sampled (Yebra et al., 2019). For example, all sites in China have only one LFMC measurement, while the site “Reader Ranch” in California has 1291 measurements. At most sites, only one plant species is sampled throughout the time, but at other sites several plant species are sampled. LFMC values vary at many sites between species. In order to simplify the comparison of LFMC measurements from different species with VOD, we grouped each species according to their genus into a typical growth form. We considered the following growth forms: broad-leaved trees (TreeB), needle-leaved trees (TreeN), shrubs, grass (i.e. herbaceous graminoid), and forbs (i.e. herbaceous non-graminoid). As some plant genera can grow as tree or shrub, we decided for one or the other based on the land cover type (forest or shrubland) at the site and based on the site photos that come with the Globe-LFMC database.
MODIS-LFMC is based on an inversion (using look-up tables) of different
radiative transfer models for grass/shrubs (PROSAILH) and trees (PROGeoSAIL) to estimate LFMC from surface reflectance observations of MODIS (Yebra et al., 2018). The method and dataset were initially developed for Australia. Additionally, we used retrievals for Europe based on the same method. The dataset provides LFMC at 500 m spatial resolution for the period 2000 to 2019 at 4-daily (Australia) and 8-daily (Europe) time steps. The dataset comes in tiles based on the original sinusoidal grid of MODIS observations. We first merged all tiles within Europe or Australia and then reprojected the data to a longitude/latitude projection (WGS84) using nearest neighbour resampling in Geospatial Data Abstraction Library software. We then aggregated the dataset to
As the relationship between VOD and LFMC also depends on leaf or canopy biomass, we additionally used leaf area index (LAI) retrievals from MODIS as
a proxy for total leaf biomass. The MOD15A2H collection 6 product provides
LAI globally on 500 m spatial resolution and 8-daily time steps (Myneni et al., 2015). We only used retrievals that were flagged as good quality in the dataset. Like the MODIS-LFMC data, the LAI data were projected to geographical coordinates and then aggregated to
The cover of trees and short vegetation was used to stratify the comparison
between LFMC and VOD with land cover information and to account for land
cover in the models to calculate LFMC from VOD. For this purpose, we used
the dataset from Song et al. (2018), which provides the percentage of tree cover, short vegetation, and bare ground within grid cells of 1 km
We additionally used information about ocean and inland water cover from the
ESA CCI land cover map (version 2.0.7) (Li et al., 2016). We aggregated the
land cover information from the original spatial resolution to the
fractional coverage of different plant functional types at
Global maps of mean annual temperature and annual precipitation from the Worldclim 2.5 dataset (Fick and Hijmans, 2017) and the CGIAR SRTM digital elevation model (Jarvis et al., 2008) were used to stratify analyses with ancillary information (Fig. A4).
Time series of the 12-monthly Standardized Precipitation Index (SPI-12) and the US Drought Severity and Coverage Index (DSCI) were used in a case study to compare the large-scale estimates of LFMC with drought conditions in the western United States and in California. SPI-12 data were taken from the Global Drought Observatory (Global Drought Observatory – JRC European Commission, 2022) and DSCI data from the US Drought Monitor (2022).
We combined the VOD and LFMC data in four different data combinations for our comparisons. Each combination of VOD and LFMC data had a different purpose and used a different masking or temporal sampling of data. The four combinations enable to: (1) compare satellite retrievals from MODIS-LFMC with the different bands of VOD; (2) compare site measurements of LFMC from Globe-LFMC with VOD; (3) calibrate and test models to estimate LFMC from VOD using LFMC site measurements; and (4) apply the best performing model to global VOD data to estimate LFMC globally.
The first data combination (D1) uses MODIS-LFMC for Australia and Europe to compare the temporal dynamic of LFMC with Ku-, X-, and C-VOD. In order to make a comparison of VOD and LFMC time series per grid cell and to assess the differences between the different VOD wavelengths, we only used observations from dates that occur in all four datasets (i.e. MODIS-LFMC, Ku-VOD, X-VOD, and C-VOD). These are for Australia 1390 time steps between 22 June 2002 and 28 July 2017, and for Europe 900 time steps between 26 June 2002 and 31 July 2017. We then computed the Spearman rank correlation between LFMC and the different VOD bands per grid cell and stratified the result for tree and short vegetation cover.
The second data combination (D2) is used to compare site measurements from
Globe-LFMC with VOD. For this comparison, we used VOD data from the same days when LFMC measurements were available. As each site has a different temporal sampling of LFMC, the number of joint pairs of LFMC and VOD observations is on average per site 80, 72, and 42 for Ku-, X-, and C-VOD, respectively, whereby the differences between the number of observations for each band are caused by the longer temporal coverage of Ku-VOD than for X- or C-VOD. Single sites have up to 827 pairs of Ku-VOD/LFMC observations. For this comparison, we only matched LFMC with the dates of each individual VOD band, but did not match additionally the dates of the three VOD bands because this would decrease the availability of LFMC/VOD pairs further. We then computed the Spearman rank correlation between VOD and all LFMC measurements for each site (regardless of the sampled plant species) and also for each individual species at a site. We calculated the correlation for all sites/site-species with at least 10 pairs of LFMC/VOD observations. Based on this criterion, correlations were computed for 910 sites. We then assessed how a difference in the land cover distribution at the site and at the corresponding
The third data combination (D3) used Globe-LFMC site measurements to calibrate and evaluate models to estimate LFMC from VOD. We found from data
combination D2 that the correlation between site measurements of LFMC and
VOD decreases if the land cover distribution at
The fourth data combination (D4) uses daily-sampled VODCA Ku-VOD and monthly-averaged MODIS LAI to estimate daily LFMC globally with the best
performing model for the overlapping period of both datasets (1 February 2000 to 31 July 2017). We applied the model to all grid cells at
We developed and tested four different models to estimate daily LFMC from daily values of VOD. All models were developed in this study either by assuming non-linear regressions between LFMC and VOD or by adopting known relations between LFMC, VOD, VWC, and dry biomass from previous studies (Jackson and Schmugge, 1991; Sawada et al., 2016; Frappart et al., 2020). Specifically, in models A and B we assume a positive relationship between VOD and LFMC and use logistic regression (S-shaped curve) to estimate LFMC from VOD. We use logistic regression because LFMC cannot be smaller than 0 % and LFMC values higher than 200 % are rare (the 95th percentile of LFMC is 193 %, the maximum is 549 % in the Globe-LFMC database). In models C and D, we adopt the relationships between LFMC, VWC, and dry biomass (Eq. 1) and between VOD and VWC (Eq. 2) to calculate LFMC. The four models are described with more detail in the following paragraphs. Each model has up to four model parameters. Prior ranges and values of those model parameters were manually selected in order to always obtain a positive relationship between VOD and LFMC and to obtain typical LFMC values (Table 2).
Overview about prior parameter values and results after site-level calibration for the four models.
In Model A, we assume that LFMC is directly proportional to VOD by
using a logistic regression as follows:
In Model B, we additionally assume that LFMC depends on seasonal
changes in canopy structure. Therefore, we additionally include monthly-averaged LAI as predictor. We assume that LFMC can be expressed
based on a weighted combination
For Model C, we directly made use of the VOD–LFMC relationship presented in Eq. (2) (Jackson and Schmugge, 1991; Konings et al., 2019) and compute LFMC by solving this equation for LFMC:
We developed Model D by using the basic definition of FMC (Eq. 1) and compute LFMC as the ratio between VWC and dry biomass:
The parameters of the models A to D were calibrated separately for each species at each site from the data combination D3. For the calibration, we
used a genetic optimization algorithm together with a cost function that is
sensitive to the statistical distribution of LFMC and the temporal correlation. We initially tested several common model performance measures
as cost functions, like the root mean squared error (RMSE), modelling
efficiency, and the Kling–Gupta efficiency (KGE) (Gupta et al., 2009), to calibrate the model parameters, but we found that based on those cost functions the variance of the observed LFMC was underestimated in most cases. As an alternative, we developed a cost function that aims to fit the observed variance by minimizing the differences in the percentiles of the statistical
distribution of LFMC. The used cost function
The cost function was minimized for each model and for each plant species at
each site by using the Genetic Optimization using Derivatives (GENOUD)
algorithm (R package rgenoud, version 5.8-3) (Mebane and Sekhon, 2011). GENOUD is a global optimization algorithm that additionally uses a local search algorithm. We used GENOUD with 100 parameter sets per generation and computed it for 10 generations. The local search algorithm was used only after the third generation to avoid a too fast convergence of the algorithm to a local minimum. Prior ranges of each parameter (Table 2) were provided as search domains to the optimization algorithm. As a measure for parameter uncertainty, we then selected the best-performing parameter sets from the optimization results, which have a cost
Additionally to the used cost function, we used the Pearson correlation
coefficient
Please note that we did not perform for the site-level calibrations any evaluation with independent test data. We used at each site all available pairs of LFMC/VOD observations for model calibration, because a split of the available LFMC observations would further reduce the available data and sites for model calibration as many sites have few observations. However, we built a random forest model to predict the parameters of the best-performing model in space and we applied spatial cross-validation to evaluate the performance of the predicted LFMC (Sect. 2.8).
The calibration of model parameters was performed for each species at each
site, and allows evaluating and comparing the performance of the four models
at site level. However, in order to apply the best-performing model globally, the model parameters need to be estimated for each grid cell of the global
In order to train and evaluate the set of three RF models, we applied a 20-fold spatial cross-validation procedure. Therefore, we spatially clustered all Globe-LFMC sites from the data collection D4 based on their coordinates using a
During this spatial training and cross-validation procedure, we also attempted to propagate the uncertainty of the optimized model parameters. Therefore, we randomly sampled from each Globe-LFMC site in the training set five out of the best-performing parameter sets from the site-level calibration. Hence, in each of the 20 folds, a different combination of best-performing parameter sets was used to train the set of RF models.
The spatial training and cross-validation resulted in 20 sets of RF models
that allow estimating Model B parameters for any grid cell based on percentage tree cover. Each of the 20 sets of RF models varies based on the
spatial distribution of the used Globe-LFMC sites in training and based on
the uncertainty of the best-performing model parameters after site-level
calibration. To estimate LFMC globally, we applied all 20 sets of RF models
to all global vegetated grid cells. Therefore, we excluded grid cells with
less than 5 % vegetation cover (tree cover
As described in the previous section, we used RF to estimate parameters of models A–D in space and then apply those models to estimate LFMC. As an alternative, RF could be used directly to estimate LFMC globally, which would not require any assumptions about the type of relationships like in models A–D and allows a higher flexibility in including predictor variables. In order to assess the performance of the spatially-applied models A–D against a more flexible global RF model, we trained a global RF model directly against LFMC measurements from all sites within the 20 spatial folds and by using the same set of predictors that we used for Model B (i.e. daily Ku-VOD, monthly LAI, and tree cover). The global training of the RF was performed with the same spatial cross-validation procedure like for the other models, i.e. with the same set of 20 folds of spatially-clustered LFMC sites.
The comparison of VOD and LFMC time series from ground measurements and
MODIS retrievals shows widespread positive temporal correlations (Fig. 1). Across the 910 Globe-LFMC sites with
Temporal correlations between Ku-VOD and LFMC. Correlations of
Ku-VOD with Globe-LFMC sites are plotted as point symbols and with MODIS-LFMC as coloured back ground raster (in
The spatial pattern of temporal correlations between LFMC and Ku-VOD indicate spatial clusters with medium to high positive correlation and clusters with low or negative correlation (Fig. 1). In the USA, sites with low correlation (
All global spatio-temporal pairs of VOD and LFMC site measurements together show a weak positive correlation but a large bi-variate scatter (Fig. 2a–c). This scatter between globally distributed VOD and LFMC indicates that a unique global VOD–LFMC relation does not exist, or that such a relationship is modified by other surface and land cover properties or by the scale mismatch between VOD grid cells and LFMC site measurements.
Global scatterplots and correlation of LFMC from the Globe-LFMC
database against Ku-, X-, and C-VOD. The red lines are smoothing spline fits
between the values at the
The medium to high positive correlations between VOD and LFMC in the majority of sites support earlier studies that identified a relationship between VOD and VWC or LFMC (Jackson and Schmugge, 1991; Konings et al., 2019). Despite the strong similarity in correlations between LFMC and the different VOD bands, contaminations by residual effects of RFI could explain the slightly lower correlation for C-VOD. The VODCA dataset uses the RFI flagging in LPRM version 6.0, which is based on the method proposed by de Nijs et al. (2015). Main contamination areas in AMSR2 in both the C1- (6.9 GHz) and C2- (7.3 GHz) bands include North America and Europe (de Nijs et al., 2015), where the majority of Globe-LFMC sites is also located. We observed that some residual RFI can still be observed in these areas, which was not covered by the masking used in VODCA (Fig. A1). As the D2 data combination uses VOD from the same days as Globe-LFMC without any smoothing, it is likely that the lower correlation between C-VOD and LFMC is affected by residual RFI.
The site measurements of LFMC and the
Although this analysis allowed to quantify how the correlation between site measurements of LFMC and coarse-resolution grid cells of VOD are affected by the land cover differences between both scales, it does not allow to resolve this scale mismatch. Only local measurements of passive microwave emissions and derived estimates of VOD in conjunction with LFMC samples allow to factor out the scale mismatch for the analysis of relations between LFMC and VOD. However, such measurements are rare (Momen et al., 2017). Our results demonstrate the need to better understand the effect of the local to regional heterogeneity in land cover on coarse-scale VOD estimates in order to make better use of VOD in estimating LFMC.
Furthermore, we investigated if the correlations between VOD and LFMC are
associated to vegetation type (Fig. 3). The comparison of VOD with Globe-LFMC shows higher correlations at higher short vegetation cover within the VOD grid cell. For example, the median correlation between LFMC and Ku-VOD is 0.30 if the short vegetation cover is
Statistical distributions of the temporal correlation between VOD
and LFMC stratified by vegetation type.
The dependency on vegetation composition becomes clearer when we compute the
correlation between VOD and LFMC separately for each sampled plant species at each site and then grouped the plant species in growth forms (Fig. 3b). We find the highest correlation for forbs (median
The results are in agreement with earlier studies that established the relation between VOD and VWC or LFMC based on observations from crops and
grasses (Jackson and Schmugge, 1991; Konings et al., 2019). The more homogenous canopies of short vegetation than of forest canopies might cause the generally higher correlations between VOD and LFMC at many herbaceous sites than at forest sites. However, based on the additional high 90th percentiles of correlations at some tree-dominated sites, we assume that coarse-resolution VOD data are also sensitive to LFMC at forest sites but that the relationship is in many cases masked by the mismatch between land cover at the local site and the coarse VOD grid cell. This assumption is also supported by the findings of Holtzman et al. (2021) who report a correlation of
The higher correlation for Ku- and X-VOD with LFMC than for C-VOD might be
confounded by an effect of rain on the atmospheric transmissivity of those
wavelengths. Although microwaves are generally assumed largely independent
of atmospheric conditions, thick water clouds and rain reduce the transmissivity of the atmosphere especially for shorter wavelength microwaves. For example, the atmospheric transmissivity is between 60 % and
80 % in the case of water clouds and between 20 % and 70 % in the case of rain for Ku-band (Ulaby et al., 1981, p. 2–3). However, effects of rain on the retrievals of Ku- and X-VOD in the VODCA product are not known. Overall, the quality of the Ku-band VOD is comparable to X- and C-VOD
(Moesinger et al., 2020): Ku-VOD correlates higher (global average
Based on the finding that Ku-VOD shows slightly higher correlations with LFMC than X- or C-VOD and given the longer temporal overlap of Ku-VOD with Globe-LFMC observations, we used Ku-VOD as input to four different models to
estimate LFMC. We separately calibrated each model at each of 216 Globe-LFMC
sites that were selected based on the criteria for the data collection D3
(Sect. 2.5). An example of the calibration of Model B for one species at one site is shown in Fig. 4. The example demonstrates a very good fit between observed and estimated LFMC (correlation
Example of the fit of Model B using daily Ku-VOD and monthly LAI for
Across all sites and vegetation types, the estimated LFMC from Model B shows
a better fit against the observed LFMC than the estimates from the other three models (Fig. 5). Model B achieves correlations of 0.64
Performance of the models A–D using daily Ku-VOD (and monthly LAI in models C and D) after calibrating each model at each site. Shown is the root mean squared error (RMSE) and correlation coefficient between estimated and measured LFMC. Small dots are results from different parameter sets at each site, and big dots and bars are the median and range from the 5th–95th percentiles across all sites, respectively.
By investigating the model performance for different vegetation growth forms, we generally found that Model B performed best and that the ranking of model performance for the other three models is similar for all vegetation types (Fig. 5). We found the highest correlation between estimated and observed LFMC for shrubs (0.73
The logistic regression Model B based on daily VOD and mean monthly LAI
outperforms other model structures. The improved performance of Model B (using VOD and monthly LAI) over Model A (only using VOD) demonstrates that the dynamics in LAI need to be considered in order to provide good
estimates of LFMC. The parameter
Model C adapted the relationship between VOD and LFMC as proposed by Jackson and Schmugge (1991) and Konings et al. (2019) (i.e. Eq. 2). Hence, this model used VOD to account for VWC and used LAI to account for canopy biomass. Model C resulted on average in low correlations and high errors between estimated and observed LFMC. While this model could not be fitted successfully at some sites, it also reached good performances at others. These results suggest that the relationship between VOD and LFMC as denoted in Eq. (2) is valid for some sites but it might not be valid for all sites or is overly sensitive to scale mismatches in the local measurements and the coarse-scale VOD and LAI data.
Model D adapted the relationship between VWC and LAI as suggested by Sawada et al. (2016) and estimated canopy biomass based on VOD. As this model resulted in better performance than Model C, it indicates that VOD is indeed a valuable predictor for canopy biomass and VWC can be indeed estimated from LAI at many sites. Model D achieved on average higher correlations between estimated and observed LFMC than Model A (only using VOD), which shows that LAI is required to predict temporal dynamics in LFMC. However, Model D had higher errors than Model A, which indicates that using VOD only as predictor for canopy biomass is not sufficient but that the VOD information in Model A provides information of absolute values (and hence reduces errors) of LFMC.
Models A, C, and D with lower performance have a low flexibility in how they use VOD and LAI to estimate LFMC: Model A only uses VOD; Model C uses VOD to account for VWC and LAI to account for biomass; and Model D uses LAI to account for VWC and VOD to account for biomass. On the other hand, Model B allowed to combine daily VOD and monthly LAI in a flexible way to estimate LFMC and reached highest performance. These results demonstrate that flexible model structures are needed in order to estimate LFMC from VOD and LAI. This finding is supported by several studies that identified that the relative contributions of changes in biomass and vegetation water content (or leaf water potential) depends on land cover type (Momen et al., 2017; Zhang et al., 2019; Konings et al., 2021b).
In a next step, we investigated the applicability of the four models in space, which requires an estimate of the model parameters for each of the
0.25
As expected, the performance of all models slightly decreased in cross-validation in comparison to the site-level calibration results
(Fig. 6). However, the ranking in model performance remained the same with Model B showing the best performance. For example, Model B had correlations of 0.58
Performance of the models A–D using Ku-VOD after calibrating each model for each species at each site (cal at site) and after using sites as test data in spatial cross-validation after the application of random forest to predict model parameters (spatial-cv). The global RF model (shown in orange) was directly trained against LFMC measurements from multiple sites. Shown is the root mean squared error (RMSE) and correlation coefficient between estimated and measured LFMC. Dots and bars are the median and range from the 5th–95th percentiles across all sites, respectively.
The global RF model achieved comparable performances like the other models,
with correlations of 0.50
These results demonstrate that especially Model B can be applied in space and results in a comparable performance in estimated LFMC between site-level calibration and spatial cross-validation.
We then performed a more detailed evaluation of the cross-validation results
of the best-performing Model B by investigating the Kling–Gupta efficiency (KGE) and its components for each site and each vegetation growth form (Fig. 7). We found the highest KGE in cross-validation for shrubs (median KGE
Performance of Model B using Ku-VOD in spatial cross-validation at
each site grouped by the sampled vegetation growth form of LFMC measurements.
The ranking in performance of the four models in spatial cross-validation resembles the ranking of the performance in site-level calibrations. On the one hand, the large variability in performance at site-level calibration and the strong decrease in performance in spatial cross-validation for Model C demonstrates that this model cannot be successfully applied and transferred to estimate LFMC globally. On the other hand, the results demonstrate that Model B can be successfully used to estimate the spatial–temporal dynamics of LFMC, whereby the parameters of model B can be estimated from observed tree cover using random forest. Medium to high performance of the estimated LFMC can be expected for herbaceous vegetation, shrublands, and for most broad-leaved trees. On average, a low performance and underestimation of the observed variance can be expected for needle-leaved trees, but this is not the case for all sites with needle-leaved trees.
The application of Model B in estimating LFMC results in performances (i.e. median RMSE
The lower performance of Model B for needle-leaved trees in spatial cross-validation than at site-level calibration indicates that the calibrated parameters from each site cannot be well estimated in space. We assume that this is caused by the spatial representativeness of the used LFMC sites with needle-leaved trees. All of the used sites with needle-leaved trees are located in the western US and most of the sites are located in regions with low tree cover. Only a few sites are located in regions with higher tree cover and those sites are distributed across different spatial clusters for cross-validation. Hence, needle-leaved trees are included in 11 out of 20 spatial clusters and six of the spatial clusters include less than three sites with needle-leaved trees. This implies that in such cases, the training of model parameters is mostly based on sites without needle-leaved trees and from other regions, which will result in a low performance for needle-leaved forests. Those results suggest that still all vegetation types should be considered in spatial cross-validation in order to obtain realistic results for under-represented vegetation types.
Overall, our estimates of LFMC from coarse-resolution VOD and LAI data reach medium to high performances for most vegetation types that are comparable with other studies that use more data with higher spatial resolution or data from optical satellite systems for which the physical relations between LFMC and surface reflectance are established for several years (Yebra et al., 2013).
Example of global patterns of LFMC and associated uncertainties as estimated with Model B for 4 selected days in 2003, representing typical days during the northern seasons and the wet and dry seasons in Africa. Grey areas (missing data) is because of missing vegetation cover or gaps in the LAI or VOD data.
Finally, we applied model B and the associated RF-based parameters to global
data of Ku-VOD, LAI, and tree cover to estimate LFMC globally at
Hovmöller diagram of monthly LFMC as estimated from Model B using daily Ku-VOD and monthly LAI.
The large similarity of the global seasonal changes in LFMC with similar changes found in other vegetation properties such as LAI or gross primary
productivity might seem astonishing at first view because LFMC represents a
relative property of moisture content and not an absolute property of vegetation cover or biomass (like LAI). However, seasonal changes in leaf
cover are highly correlated with LFMC, especially in short vegetation regions. For example, MODIS LAI has an across-site median temporal correlation with measurements of the Globe-LFMC dataset between
Areas without estimates of LFMC (grey areas in Figs. 8 and 9) occur because of several reasons. (1) Missing data in deserts and ice-covered regions are because the model was not applied to grid cells with less than 5 % vegetation cover. (2) Missing data in northern latitudes in winter months are either because of months without LAI observations because of low solar zenith angles, snow or cloud cover, or because Ku-VOD observations were not available over frozen soils. (3) Other days with missing observations in some regions are because of missing coverage of passive microwave sensors or were masked in the VODCA dataset because of RFI.
We also compared the estimated LFMC from Model B with MODIS-LFMC for Australia and Europe to assess the similarity of both datasets. However, as the VOD-based LFMC uses monthly LAI from MODIS as input, which is derived from the same spectral bands like MODIS-LFMC, both LFMC datasets are not independent of each other and a high correlation can be expected. Indeed both the VOD-based LFMC and MODIS-LFMC are highly correlated (Fig. A3). The spatial patterns of correlation between the VOD-based LFMC and MODIS-LFMC show similar regions where Ku-VOD already had high and low correlations with MODIS-LFMC, respectively (Fig. 1). The correlation between VOD-based LFMC and MODIS-LFMC is higher than between Ku-VOD and MODIS-LFMC in many regions, which is likely due to the additional use of MODIS-LAI in Model B. The low correlation in parts of northern Europe and the Alps was already present in the correlation between MODIS-LFMC and Ku-VOD. The low correlation between VOD-based LFMC and MODIS-LFMC in northern Europe can be additionally caused by the low performance of the estimates in needle-leaved forests, which are widespread in those regions. However, the very high correlation between VOD-based LFMC and MODIS-LFMC demonstrates in many regions and in most fire-prone regions a good comparability of the two datasets.
The uncertainty estimates of the global LFMC estimates are on average low
(global mean relative uncertainty
The analysis of the estimated global patterns of LFMC needs to be compared
with the number of observations that support the global estimates. The majority of pairs of Ku-VOD and Globe-LFMC observations come from the western US and from sites in the Mediterranean, western Africa, and southern
Australia (Fig. A4a). The available Globe-LFMC observations cover mean annual temperatures between
We additionally estimated the number of supporting observations in space as
a function of mean annual temperature, tree cover, maximum Ku-VOD, and maximum LAI (Fig. A4b). Therefore, a random forest regression was fitted to the number of observations per site in the Globe-LFMC database and by using mean annual temperature, tree cover, mean annual maximum Ku-VOD, and mean annual maximum LAI as predictors. The fitted random forest model was then applied to each
The aim of this study was to investigate the VOD–LFMC relationship and to develop and test model approaches to estimate LFMC globally. We also generated a daily LFMC dataset for past conditions, whereby the daily information originates from the Ku-VOD data. Although the presented LFMC dataset has a much coarser spatial resolution than MODIS-LFMC datasets (Yebra et al., 2018; Quan et al., 2021; Zhu et al., 2021), the advantages are the daily coverage, because VOD is cloud- and illumination-independent, and the long timespan of VOD data (e.g. Ku-VOD starting in 1987), which potentially allow to produce long-term estimates of LFMC in future studies. Hence, the described methodology to estimate LFMC from VOD can complement LFMC retrievals from optical sensors by providing a higher temporal frequency and potentially a longer temporal coverage.
We envision several applications of the global Ku-VOD-based estimates of leaf moisture content (expressed as LFMC), but also want to raise attention to the limitations of the dataset in other applications. The VOD-based LFMC estimates are suitable to investigate large-scale patterns of vegetation responses to drought, to assess fire danger and to estimate fire emissions, or to benchmark global ecohydrological and fire-enabled vegetation models.
Several remotely sensed vegetation properties such as spectral vegetation indices, LAI, sun-induced fluorescence, or derived variables of plant productivity are frequently used to monitor drought effects on vegetation (e.g. Jiao et al., 2021; Crocetti et al., 2020) or to investigate the effects of water availability on vegetation growth. The VOD-based LFMC estimates can complement such analyses by providing information on large-scale changes in leaf moisture content.
As a case study, we compared the VOD-based LFMC with drought conditions in North America and specifically in California by using the 12-monthly Standardized Precipitation Index (SPI-12) and the US Drought Severity and Coverage Index (DSCI) (Fig. 10). August 2014 was one of the most severe drought months in the western US. The VOD-based LFMC estimates show widespread patterns of very low LFMC over the western US during this month (Fig. 10a). This corresponds to a lack in precipitation as indicated by the negative SPI-12 (Fig. 10b). Also, large regions in northern Canada show precipitation deficit with low SPI-12 in northern Canada, which also corresponds to patterns of low LFMC.
Comparison of LFMC as estimated from model B with drought conditions in North America and California.
To investigate multi-year drought events, we also compared LFMC, SPI-12, and
DSCI time series averaged for the state of California (Fig. 10c). Both SPI-12 and the DSCI show the multi-year drought between 2013 and 2016. The LFMC time series is dominated by the strong seasonal signal. Therefore, we decomposed the LFMC time series for California into a seasonal, trend, and remainder component using the seasonal decomposition of time series by Loess (STL) method (Cleveland et al., 1990). The LFMC trend shows a long period of low values between 2013 and 2016, which corresponds to the drought period. Likewise, the wet period between 2005 and 2007 with higher precipitation (i.e. high SPI-12) and no drought conditions (i.e. DSCI close to 0) corresponds to high LFMC values. The LFMC trend component is medium correlated with SPI-12 (
Generally, the main application of LFMC data is the assessment of fire risks (Chuvieco et al., 2010). The high temporal frequency and long period of the VOD-based LFMC dataset allow investigating short-term to long-term changes in fuel moisture and hence fire risk at a large scale. Previously, VOD datasets have been used as proxies for fuel conditions in global empirical models of burned area (Forkel et al., 2017; Kuhn-Régnier et al., 2021) and helped to explain how trends in climate conditions and vegetation affect large-scale trends in burned area (Forkel et al., 2019). However, the interpretation of VOD effects on the prediction of burned area was hampered in those studies by the unclear role of VOD as a proxy for fuel loads (biomass) or fuel moisture content. The VOD-based LFMC estimates overcome this problem by translating VOD into LFMC. Besides in empirical models for large-scale burned area, the VOD-based LFMC estimates can be used to investigate changes in fire radiative energy or fire emissions, which both depend on fuel moisture content. Further investigations could assess the predictive performance of the VOD-based LFMC data within large-scale empirical modelling studies to predict burned area or other properties of fire dynamics.
However, the coarse spatial resolution of the VOD-based LFMC data
(
Furthermore, the VOD-based LFMC estimates can contribute to the evaluation and improvement of moisture simulations in global ecohydrological and fire-enabled vegetation models such as from the fire-model inter-comparison project (FireMIP) (Rabin et al., 2017). FireMIP models simulate live and dead fuel moisture either based on fire danger indices (e.g. the Nesterov index, Thonicke et al., 2010) or based on empirical functions with soil moisture or relative humidity (Rabin et al., 2017). FireMIP models have been intensively evaluated for simulations of burned area, fire emissions, LAI, plant productivity, and biomass (Hantson et al., 2020), and the simulated fuel moisture has a strong effect on simulations of burned area and fire emissions (Li et al., 2019). However, fuel moisture has not yet been evaluated in those models. Hence, we propose that the VOD-based LFMC estimates or other global products (Quan et al., 2021) can be used in benchmarking activities of global fire-enabled vegetation models.
Finally, we propose several further developments of the VOD-based LFMC
datasets:
The calibration and evaluation of the applied models used only 163 sites out of 1384 sites in the Globe-LFMC database according to the selection criteria described in Sect. 2.5. This is mainly caused by the joint availability of pairs of LFMC/VOD observations. Additionally, our selection criteria also prevented us from using the measurements from all 229 sites in China, where each site has only one measurement. Future developments can apply different approaches to make use of more observations in model training. While sites with single measurements cannot be used to calibrate models at site level, they could be still used in training the spatial random forest model to estimate model parameters. At other sites, a filling of short temporal gaps in VOD time series could increase the availability of LFMC/VOD pairs and would increase the number of sites that can be used for model calibration. An estimation of LFMC for different vegetation types within a VOD grid cell can be explored as the site-level model calibration was performed for different vegetation growth forms reported in the Globe-LFMC database. LAI data at their original spatial resolution within the VOD-based models could be used to provide LFMC estimates at higher spatial resolution. One advantage of our methodology is the long timespan of VOD data (e.g. Ku-VOD starting in 1987), which potentially allows to produce long-term estimates of LFMC in future studies. Hence, the temporal coverage of the LFMC estimates can be extended back to 1987 by using longer LAI time series than provided by MODIS. Such an extension would also allow the use of older LFMC field data in model calibration. Such long time series of LFMC can facilitate climatological studies on the variability and LFMC and the potential effects on fire. The prediction of fire risks requires the availability of satellite products shortly after the observation. Our methodology could be applied to estimate LFMC in near-real time, however, this requires the availability of near-real time VOD products.
This study assessed the relationship between short-wavelength VOD from
passive microwave satellite observations and leaf moisture content (expressed as LFMC) globally, and successfully developed and applied a method to estimate LFMC from VOD globally at We investigated the relationship between VOD and LFMC. VOD and LFMC are in the majority of sites or grid cells positively correlated, whereby Ku-band VOD has slightly higher correlations than X- or C-VOD. The correlation between VOD and LFMC is on average higher for short vegetation types such as forbs, grasses, and shrubs than for trees, but also several forest sites show high correlations. Broad-leaved forests show higher correlations than needle-leaved forest. These results confirm earlier studies about the VOD–LFMC relation and demonstrate additionally that coarse-scale VOD is sensitive to LFMC at forest sites if the land cover distribution locally is similar to the coarse grid cell. We tested different model structures to estimate LFMC from VOD. A logistic regression model that uses daily Ku-VOD and monthly LAI as predictors for LFMC outperformed alternative model structures in site-level calibration and spatial cross-validation. The comparison of model structures demonstrates that LAI is needed in addition to VOD as a proxy for either canopy biomass or vegetation water content to reach acceptable model performances. We applied spatial cross-validation to assess the transferability of model structures in space and applied the best-performing model to estimate LFMC globally. The obtained model performances are comparable with results from previous studies that estimated LFMC based on visible/near-infrared satellite observations. Medium to high performance of the VOD-based LFMC estimates can be expected for herbaceous vegetation, shrublands, and for most broad-leaved trees in many fire-prone regions, such as in western Canada, the western US and Mexico, in southern South America, in the Mediterranean, central Asia, parts of China, southern and eastern Africa, and southern and eastern Australia. Large variability in performance and high uncertainties can be expected in needle-leaved forests, whereby especially estimates in boreal forest have low observational support.
We propose to use VOD-based estimates of LFMC to investigate effects of
drought and climate variability on vegetation leaf moisture at large scale,
for large-scale assessments and empirical modelling of fire dynamics, or to
benchmark global fire-enabled vegetation models.
Example time series of C-, X-, and Ku-VOD over a grid cell in Spain (5.88
Across-site correlations between the site-optimized parameters
Pearson correlation between the VOD-based LFMC from Model B and MODIS-LFMC for Australia and Europe for the time period February 2000 to July 2017.
Distribution of the number of joint observations of Ku-VOD and
Globe-LFMC measurements.
Statistical distributions of the temporal correlation between Ku-VOD or LAI and measurements from the Globe-LFMC database stratified by the percentage cover of short vegetation.
The global daily LFMC results of model B for the period
February 2000 to July 2017 are available at Zenodo:
MF and WD developed the concept. MF, LS and MY processed and curated data. MF, LS and RMZ performed the analysis. MF developed the methodology. MF and RMZ developed the visualizations and created the figures. MF wrote the original draft which has been reviewed and edited by all co-authors.
Matthias Forkel is guest editor of the special issue “Microwave remote sensing for improved understanding of vegetation–water interactions”. The peer-review process was guided by an independent editor, and the authors have also no other competing interests to declare.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the special issue “Microwave remote sensing for improved understanding of vegetation–water interactions (BG/HESS inter-journal SI)”. It is a result of the EGU General Assembly 2020, 3–8 May 2020.
We thank Niels Andela and Sandy Harrison for comments on early developments of this work. Matthias Forkel and Luisa Schmidt acknowledge funding from TU Dresden and from the H2020 project FirEUrisk (101003890).
This research has been supported by the H2020 Environment (grant no. 101003890). This open access publication was financed by the Saxon State and University Library Dresden (SLUB Dresden).
This paper was edited by Julia K. Green and reviewed by Seung Hee Kim, Lei Fan, and one anonymous referee.