The interactions between climate, vegetation and fire can strongly influence the future trajectories of vegetation in Earth system models. We evaluate the relationships between tropical climate, vegetation and fire in the global vegetation model JSBACH, using a simple fire scheme and the complex fire model SPITFIRE with the aim to identify potential for model improvement. We use two remote-sensing products (based on MODIS and Landsat) in different resolutions to assess the robustness of the obtained observed relationships. We evaluate the model using a multivariate comparison that allows us to focus on the interactions between climate, vegetation and fire and test the influence of land use change on the modelled patterns. Climate–vegetation–fire relationships are known to differ between continents; we therefore perform the analysis for each continent separately.
The observed relationships are similar in the two satellite data sets, but maximum tree cover is reached at higher precipitation values for coarser resolution. This shows that the spatial scale of models and data needs to be consistent for meaningful comparisons. The model captures the broad spatial patterns with regional differences, which are partly due to the climate forcing derived from an Earth system model. Compared to the simple fire scheme, SPITFIRE strongly improves the spatial pattern of burned area and the distribution of burned area along increasing precipitation. The correlation between precipitation and tree cover is higher in the observations than in the largely climate-driven vegetation model, with both fire models. The multivariate comparison identifies excessive tree cover in low-precipitation areas and a too-strong relationship between high fire occurrence and low tree cover for the complex fire model. We therefore suggest that drought effects on tree cover and the impact of burned area on tree cover or the adaptation of trees to fire can be improved.
The observed variation in the relationship between precipitation and maximum tree cover between continents is higher than the simulated one. Land use contributes to the intercontinental differences in fire regimes with SPITFIRE and strongly overprints the modelled multimodality of tree cover with SPITFIRE.
The multivariate model–data comparison used here has several advantages: it improves the attribution of model–data mismatches to model processes, it reduces the impact of biases in the meteorological forcing on the evaluation and it allows us to evaluate not only a specific target variable but also the interactions.
Capturing the interactions of vegetation cover and composition with the
climatic drivers and related disturbances in Earth system models is crucial
to reliably estimate changes in vegetation for a changing climate. Climate is
the main driver of global vegetation patterns, but vegetation also has
crucial impacts on the Earth system due to its influence on the surface
albedo and the water cycle
Interactions between vegetation, fire and climate are particularly important
to understand the spatial patterns in tropical vegetation, which is
characterised by strong gradients from deserts to tropical rainforests.
Remotely sensed tropical tree cover shows a bimodality between forest
(
While data analysis can
provide insights into driving factors for certain variables, process-based
models summarise the understanding of the process and allow us to perform
experiments that are impossible in reality. Dynamic global vegetation models
(DGVMs) were developed to understand ecosystem dynamics, the carbon cycle and
biosphere–atmosphere interactions
The development of remotely sensed global burned area
products facilitated the implementation and evaluation of complex fire models
within DGVMs
Here, we aim (1) to assess the robustness of observed climate–vegetation–fire relationships across the tropical continents based on two remotely sensed tree cover data sets; (2) to test a multivariate model evaluation to identify opportunities for model improvements in JSBACH, the vegetation model used within the MPI Earth system model, and (3) to test the contribution of land use change on the obtained relationships.
To investigate the climate–fire–vegetation relationships in the tropical
regions, we represent climate by the mean annual precipitation (
We define the tropical region as between
We use the JSBACH land
surface model
Both fire models interact with the vegetation model as follows: JSBACH provides fuel amounts, vegetation composition and soil moisture as inputs to the fire model. The fire model in turn reduces the carbon pools of JSBACH according to the simulated carbon combustion of vegetation fires and reduces the cover fractions of burned vegetation. In the JSBACH-standard fire scheme the burned area directly translates into a reduction of the cover fractions of the plant functional types (PFTs) (100 % of the cover fractions on burned area are removed). However, in SPITFIRE the mortality of woody vegetation depends on the fire intensity, fire residence time, the vegetation height and bark thickness.
The model's plant functional types for the tropics include C3 and C4 grass,
tropical evergreen and deciduous trees, and rain green shrubs. Shrubs and
trees compete according to their net primary productivity. Grasses and shrubs
have an advantage compared to trees in regions with disturbances due to their
lower establishment timescale (Reick et al., 2013, grasses: 1 year, shrubs:
12 years, tropical trees: 30 years). PFTs do not establish if the 5-year
running mean net primary productivity (NPP) turns negative. Trees prevail in
grid cells without disturbance and positive NPP. Land use is included
following the protocol of
JSBACH-SPITFIRE shows a reasonable agreement with remotely sensed data
products for present-day burned area and carbon emissions for simulations
with prescribed land cover
JSBACH was forced with meteorological data for the historical period
1850–2005, which was extracted from a coupled simulation with the MPI-ESM
version 1.1. For the computation of ignitions the SPITFIRE model additionally
uses a population density data set
We averaged the remote-sensing data sets over the years that were covered by
all data sets (2001–2010). Model output is only available until the year
2005. Using only the overlapping period (2001–2005) would decrease the
robustness of the mean fire regime and climate characterisation. We therefore
use different averaging periods for the model (1996–2005) and observations
(2001–2010). The presentation of the relationship between precipitation,
tree cover and burned fraction based on remote-sensing data is based on
0.25
We use two tree cover data sets based on satellite data, one based on the
MODIS (Moderate-Resolution Imaging Spectroradiometer) sensor
The maximum tree cover in the MODIS data set is 80 %. This, however,
corresponds to 100 % crown cover
The MODIS collection 5 land cover data set
The Global Fire Emissions Database (GFED,
The TRMM and Other Data Precipitation Data Set (TMPA) is based on the
version 7 TRMM Multi-satellite Precipitation Analysis algorithm
We used the shortwave radiation and temperature of the CRU-NCEP v5 data set
reanalysis
We use quantile regressions to characterise the relationship between
precipitation and maximum tree cover. The quantile regressions were computed
with the R package quantreg
We first give an overview of the geographical distribution of the used observation and model output data sets. The comparison of geographical patterns is an important assessment of model performance, it is, however, difficult to assess whether the interactions between precipitation, fire and tree cover are well captured. Moreover as the JSBACH model is usually used as a land surface model for the MPI-ESM and therefore also forced here with MPI-ESM output, biases in model forcing can cause geographical biases of vegetation and fire variables even with a perfect fire and vegetation model. To reduce the influence of biases in forcing data on the model–data comparison and allow us to more closely evaluate the interactions between model components we propose a multivariate evaluation of climate–fire–vegetation relationships. We assess the robustness of observed relationships for two tree cover data sets and two spatial resolutions and compare them to the model simulations. The last paragraph of this section addresses the influence of land use change on the simulated relationships.
The two observational satellite-based tree cover data sets are consistent and
show only small differences in their spatial pattern (Fig. 1a). The overall
clear pattern in tree cover is a transition from very high tree cover in
moist rainforest regions to low tree cover in the drier savannas to the
absence of trees in the desert regions. Both models reproduce this overall
observed pattern, but with marked local differences. Both model versions
overestimate tree cover in northern Australia to a similar extent. In the
north-eastern Amazon region, the simulations underestimate tree cover compared
to the observations. This underestimation is much smaller for
JSBACH-SPITFIRE. The simulations overestimate tree cover in southern
hemispheric Africa, which is again smaller for JSBACH-SPITFIRE.
The simulated grass cover has higher maximum values that are generally often
lower than those observed by satellite (Fig.
Generally, JSBACH-standard strongly
underestimates the total area burnt and the spatial variability
(Fig.
Precipitation
of the MPI-ESM forcing shows a dry bias in the eastern and central Amazon
region, a dry bias in Asia, and moister conditions in the western part of
southern hemispheric Africa (Fig.
Spatial distribution of modelled and observed data sets used in this
study.
Maximum tree cover shows an increase along the precipitation gradient across
all continents, with trees being absent until a certain threshold
(300–500 mm year
The Spearman rank correlation between precipitation and tree cover is very
similar for both tree cover data sets (Table 1). The statistical precipitation
thresholds for low (but higher than 0) and high tree cover differ by less
than 100 mm. The aggregation to the model resolution shows the strongest
effect on the precipitation threshold for high tree cover and shifts this
value to higher precipitation. The association between precipitation and
burned area is less sensitive to the aggregation: 80 % of the global burned
area occurs in regions with precipitation between 609 and 1518 mm at
0.25
Tree cover (TC) versus precipitation (mm year
Spearman rank correlation (
In the tropics, the observed burned area is strongly constrained by
precipitation: around 80 % of the burned area is observed in regions with
mean annual precipitation between 600 and 1500 mm year
Cumulative burned area normalised with the total burned area for increasing precipitation. For the GFEDv4 burned area the TMPA data set was used; for the model simulations the MPI-ESM precipitation was used.
Surprisingly the observations show a higher Spearman correlation between tree
cover and precipitation than the models (Table 1). The lower correlation of
the modelled relationship most likely originates from the lower precipitation
regions (
Modelled and observed tree cover (TC) versus precipitation (
Models and observations generally agree on the absence of fire for very high
tree cover (
Burned fraction is much lower in Asia and South America compared to Australia
and Africa in the observations. Both models show an underestimation of the
fire occurrence in Australia. SPITFIRE reproduces the fire regime with high
annual burned fraction in Africa. In JSBACH-standard the difference in burned
fraction between the continents is smaller than in JSBACH-SPITFIRE
(Fig.
Models and observations show differences between
continents in the relationship between precipitation and maximum tree cover
(Fig.
Modelled and observed relationship between precipitation and maximum tree cover based on a linear quantile regression (dashed line) and a local quantile regression (solid line). Different colours indicate the different continents.
The grass cover has a much higher variability in the model compared to the
MODIS data (Fig.
Modelled and observed grass cover (GC) and non-vegetated fraction
over precipitation (
The simulation with pre-industrial land use represents a state with low
influence of land use change. The comparison with the historical simulation
allows us to assess the influence of land use change after 1850. The impact of
fire on tree cover, as quantified by the Spearman rank correlation, between
burned fraction and tree cover is higher for simulation with
pre-industrial land use (Table 1). This indicates that anthropogenic land
cover change decreases the impact of fire on the vegetation distribution.
Land use change did not affect the rank correlation between precipitation and
tree cover. The precipitation range for 80 % of the burned area is only
slightly narrower for the simulation including land use change (Table 1).
Tree cover, however, is even higher for low precipitation and reaches canopy
closure for lower precipitation Table 1 and Fig.
Same as Fig. 4 for JSBACH-SPITFIRE but with pre-industrial land use.
The multivariate model–data comparison identified differences and agreements between modelled and observed interactions between fire, vegetation and climate. It goes beyond spatial comparisons by providing better guidance on which processes in the model need improvement. Here we discuss which model improvements can help to address the differences, what causes agreements in intercontinental differences and whether limitations of the observations might influence our findings.
JSBACH overestimates tree cover for low precipitation on all tropical
continents. In these dry regions no or only very low burned fractions are
observed, and SPITFIRE shows a good response to precipitation, while the
JSBACH-standard overestimates the burned area (Fig. 3). The improved
burned area pattern of SPITFIRE did not lead to an improvement in tree cover
for these dry regions. It is therefore unlikely that further improvements in
burned fraction will improve this model–data mismatch for tree cover in dry
regions. Satellite data, however, show that the intensity of fires increases in
these regions
The absence
of fire for closed canopies is captured well by JSBACH-SPITFIRE. The modelled
strong relationship between higher burned fraction and lower tree cover for
open canopies (Fig.
A more detailed representation of
vegetation structure, including a sapling state of trees that is more
sensitive to fire
For Australia underestimation of burned
area for both fire models is strong (Fig.
The rank correlation between precipitation and tree
cover is higher for the observations compared to the model outputs (Table 1).
One reason might be the lower maximum tree cover for low precipitation in the
observations, which limits the range of tree cover values in these regions. In
JSBACH-standard the correlation between tree cover and precipitation is
stronger than in JSBACH-SPITFIRE. In the JSBACH-standard model, fire is only
driven by meteorological variables and vegetation properties (which also
largely follow climatic gradients). JSBACH-SPITFIRE, however, also uses
population density and lightning data sets as input, which are potentially
inconsistent with the meteorological forcing derived from the MPI-ESM output.
Lightning strikes are strongly related to precipitation
The suggested processes are known to be important for vegetation distribution and it seems plausible that they can help to improve the vegetation distribution. How exactly these plausible modifications would change the patterns of tree cover, fire and their relation to climate likely strongly depends on the exact parameterisation and needs to be tested with stepwise model development and factorial simulations.
We find differences in the climate–vegetation–fire relationships between
continents in the satellite products as well as in the model simulations with
JSBACH-SPITFIRE and the JSBACH standard model. Differences in the
climate–vegetation–fire relationships have been described based on site-level
data sets
Our model simulations show that
global vegetation models can also have differences in
climate–vegetation–fire relationships between continents. We separated the
effect of land use change by comparing the historical simulation to a
simulation with pre-industrial land use. We find that land cover change
influences the differences in the modelled fire regime between Africa and
South America. Land cover change influences simulated fire occurrence as
cropland areas are excluded from burning and pastures have a higher fuel bulk
density in the JSBACH-SPITFIRE model. A reduction in burned area due to
increases in croplands is well supported by statistical analysis of satellite
data for Africa
Vegetation in the MPI
Earth system model including SPITFIRE is not only a function of climate but
also depends on the history of previous vegetation due to the feedback
between fire and vegetation
The comparison of the increase in maximum tree cover with increasing precipitation shows that, although the model shows some variability between continents, it misses a large part of the observed variation. Finding the correct balance of the many influencing factors, e.g. climate, fire, land use, evolutionary differences, will remain a challenge for the future.
We use two remotely sensed tree cover products, which show coherent patterns.
Although these products are derived from imagery with different spectral,
temporal and spatial characteristics (MODIS and Landsat), they cannot be
considered totally independent because both are derived using a similar
classification and regression tree method as well as reference data. The
observational tree cover data sets are limited to trees taller than 5 m and do
not include shrubs. For the model, however, we included shrubs and all trees.
Previously differences in the threshold at which maximum tree cover is reached
were attributed to different precipitation data sets and exclusion or inclusion of
shrub cover
Compared to the satellite data sets, an
African site-level data set shows lower thresholds of precipitation for the
absence of trees (ca. 100 mm year
The maximum value of a variable can decrease due to spatial
averaging. We tested this effect by not using the mean when aggregating the
satellite tree cover to the resolution of the precipitation data set but
instead using the maximum value of the underlying 0.05
Tree cover seems to be a clearly defined variable but already varies between the two satellite
data sets. The MODIS tree cover data set defines a maximum tree cover of
80 %, while the Landsat tree cover data set allows a cover of 100 %. In
the observations canopies that are not fully closed due to low foliar biomass might be
tracked as reduced tree cover. In the model, however, tree cover and
biomass are two rather independent variables, meaning that tree cover can be
high in spite of a low biomass. Biomass data sets might therefore give
additional valuable insights and pan-tropical data sets are available
The latest release of the
GFED burned area and emissions data sets includes an extension for small fires
By evaluating tree cover and
fire for a given mean annual precipitation we account for biases in the
MPI-ESM forcing of this parameter. Mean annual precipitation is a strong
driver of vegetation patterns, especially in the tropics; however other
aspects of precipitation and other climatic parameters might be biased and
influence our results. Many climate models have problems representing
extremes
The interactions between climatic parameters are, however, difficult to
disentangle, based on this simple analysis and other approaches such as
multivariate regression or random forrest approaches
This study combines two satellite data sets with model simulations using a simple and a complex fire algorithm to investigate relationships between fire, vegetation and climate. Our analysis shows that the two satellite data sets are consistent in terms of the relationship between tree cover, precipitation and fire occurrence, but the spatial scale needs to be considered, as some statistical characteristics change with the resolution.
Our analysis showed the strength of the multivariate comparison in detecting
model inconsistencies and guiding model development. It goes beyond the insights gained
by standard spatial comparisons. For JSBACH, independently of the fire model
used, we find an overestimation of tree cover for low precipitation where
typical fire occurrence is low due to limited fuel availability. The response
of burned area to precipitation was captured well for SPITFIRE, but the
simple fire scheme showed an overestimation of burned area for dry regions.
This indicates that an improvement in the fire model cannot improve the
response of vegetation to climate in dry regions but improved modelling of
drought effects on the vegetation dynamics can. Dry regions often show a
strong coupling between land and atmosphere
While fire occurrence and vegetation patterns are observed well by remote sensing, the impact of fire on vegetation is much less constrained by satellite observations limiting the possibilities of evaluating that part of fire models. The multivariate comparison revealed a too-strong impact of fire on tree cover for grid cells with very high fire occurrence, which leads to too-low tree cover. To boost the tree cover in exactly these regions with high fire occurrence, possible model modifications are an adaptation of trees to fire, by increasing bark thickness in response to high fire frequencies or a stronger negative feedback between fire occurrence and fuel load. This stronger feedback should then reduce fire intensity and consequently fire mortality.
The complex fire model SPITFIRE improves the difference in fire regimes between the continents, especially Africa and South America, compared to the simple fire model. The intercontinental variation in the relationship between precipitation and maximum tree cover is much smaller for the models compared to the observations. Known variations in vegetation are not sufficiently understood to be represented in models. However, our finding that models do show differences in the climate–fire–vegetation relationships between continents shows that further exploration of why models show differences can be helpful to better understand causes for intercontinental differences.
Overall the multivariate model evaluation highlights the potential for more targeted model improvements with respect to the interactions between climate, vegetation and fire, which are crucial for our understanding of future vegetation and climate projections.
The observational data sets are freely available. The processed data and model output as displayed in this publication and the processing scripts are available upon request to publications@mpimet.mpg.de.
Same as Fig. 4 but tree cover filtered for the presence of shrublands (using the MODIS open and closed shrubland classification). This indicates a low sensitivity of the climate–vegetation–fire relationships to shrublands.
Modelled tree cover (TC) versus precipitation (
Tree cover (TC) versus precipitation (
Additionally to the total amount of rainfall the seasonality can play a role
for vegetation or the length of dry periods. We therefore assess whether
the rainfall seasonality and the number of dry days are reasonable in our
climatic forcing here. We use the TMPA 3B42 daily data set
Relationship between annual precipitation and precipitation seasonality and number of dry days, respectively, for the ECHAM simulation used as meteorological forcing for the JSBACH simulations used here and the TMPA 3B42 daily data set. Slope indicates the slope of the regression line.
Relationship between annual burned area and fire line intensity. The expected decrease in fire line intensity for frequently burning areas due to the feedback between fire and fuel load is not found in the simulation results and might indicate that the feedback between fire occurrence, fuel load and fire intensity is too weak.
GL wrote the manuscript. GL and TM designed the study and performed the analysis. SH, DD, SK helped to refine the analysis and to develop and shape the manuscript.
The authors declare that they have no conflict of interest.
We would like to thank the DKRZ for excellent computing facilities. Donatella D'Onofrio acknowledges support from the European Union Horizon 2020 research and innovation programme under grant agreement no. 641816 (CRESCENDO). Stijn Hantson acknowledges support by the EU FP7 projects BACCHUS (grant agreement no. 603445) and LUC4C (grant agreement no. 603542). We thank Victor Brovkin for valuable discussions and comments on this manuscript and are grateful to the three anonymous reviewers for their detailed reviews. The article processing charges for this open-access publication were covered by the Max Planck Society. Edited by: Alexey V. Eliseev Reviewed by: three anonymous referees