With remote sensing we can readily
observe the Earth's surface, but direct observation of the sub-surface
remains a challenge. In hydrology, but also in related disciplines such as
agricultural and atmospheric sciences, knowledge of the dynamics of soil
moisture in the root zone of vegetation is essential, as this part of the
vadose zone is the core component controlling the partitioning of water into
evaporative fluxes, drainage, recharge, and runoff. In this paper, we compared
the catchment-scale soil moisture content in the root zone of vegetation,
computed by a lumped conceptual model, with the remotely sensed Normalized
Difference Infrared Index (NDII) in the Upper Ping River basin (UPRB) in
northern Thailand. The NDII is widely used to monitor the equivalent water
thickness (EWT) of leaves and canopy. Satellite data from the Moderate
Resolution Imaging Spectroradiometer (MODIS) were used to determine the NDII
over an 8-day period, covering the study area from 2001 to 2013. The results
show that NDII values decrease sharply at the end of the wet season in
October and reach lowest values near the end of the dry season in March. The
values then increase abruptly after rains have started, but vary in an
insignificant manner from the middle to the late rainy season. This paper
investigates if the NDII can be used as a proxy for moisture deficit and
hence for the amount of moisture stored in the root zone of vegetation, which
is a crucial component of hydrological models. During periods of moisture
stress, the 8-day average NDII values were found to correlate well with the
8-day average soil moisture content (
Estimating the moisture content of the soil from remote sensing is one of the major challenges in the field of hydrology (e.g. De Jeu et al., 2008; Entekhabi et al., 2010). Soil moisture is generally seen as the key hydrological state variable determining the partitioning of fluxes (into direct runoff, recharge, and evaporation) (Liang et. al., 1994), the interaction with the atmosphere (Legates et. al., 2011), and the carbon cycle (Porporato et al., 2004). The root zone of ecosystems, being the dynamic part of the unsaturated zone, is the key part of the soil related to numerous sub-surface processes (Shukla and Mintz, 1982). Several remote sensing products have been developed especially for monitoring soil moisture (e.g. SMOS, ERS, and AMSR-E) but until now correlations between remote sensing products and observed soil moisture at different depths have been modest at best (Parajka et al., 2006; Ford et al., 2014). There are a few possible explanations. One is that it is not (yet) possible to look into the soil deep enough to observe soil moisture in the root zone of vegetation (Shi et al., 1997; Entekhabi et al., 2010); the second is that soil moisture observations at certain depths are maybe not the right indicators for the amount of moisture stored in the root zone (Mahmood and Hubbard, 2007), which is rather determined by the vegetation-dependent, spatially variable, three-dimensional distribution and density of roots.
These mainstream methods to derive soil moisture from remote sensing have
concentrated on direct observation of soil moisture below the surface. The
vegetation, through the vegetation water content (VWC), perturbs this
picture. As a result, previous studies have tried to determine the VWC from a
linear relationship with the equivalent water thickness (EWT) that is
measured by the Normalized Difference Infrared Index (NDII) (e.g. Yilmaz et
al., 2008). The NDII was developed by Hardisky et al. (1983) using ratios of
different values of near infrared reflectance (NIR) and short wave infrared
reflectance (SWIR), defined by
(
Water is one of the determinant environmental variables for vegetation growth, especially in water-limited ecosystems during dry periods. From the plant physiology point of view, water absorption from the root zone is driven by osmosis. Subsequently, water transport from the roots to the leaves is driven by water potential differences, caused by diffusion of water out of stomata, called transpiration. This physiological relationship supports the correlation between root zone soil moisture content, moisture tension in the leaves, and the water content of plants.
Hence, the root zone moisture deficit is connected to the water content of the canopy/leaves, because soil moisture suction pressure and moisture content in the leaves are directly connected (Rutter and Sands, 1958). The NDII was developed to monitor leaf water content (Hardisky et al., 1983), so one would expect a direct relation between NDII and root zone moisture deficit. The deficit again is a direct function of the amount of moisture stored in the root zone.
So, if leaf water thickness and the suction pressure in the root zone are connected, then the NDII would directly reflect the moisture content of the root zone. It would only reflect the moisture content in the influence zone of roots and not beyond that. Hence, the NDII could become a powerful indicator for monitoring root zone moisture content, providing an integrated, depth-independent estimation of how much water is accessible to roots, available for vegetation. In other words, the NDII would allow us to see vegetation as a sort of natural manometer, providing us with information on how much water is available in the sub-surface for use by vegetation. It would be an integrated indicator of soil moisture in the root zone, available directly at the scale of interest.
Thus, the hypothesis is that we can monitor the moisture content in the root zone from the observed moisture state of the vegetation by means of the NDII.
In this paper, we tested whether there exists a direct and functional relationship between a remote sensing product (the NDII) and the amount of moisture stored in the root zone, as simulated by a semi-distributed conceptual hydrological model, in which the root zone moisture content is a key state variable in the short- and long-term dynamics of the rainfall–runoff signal. Because the NDII is an indicator for water stress, the index is only expected to show a strong link with the moisture content of the root zone when there is a soil moisture deficit. Without water stress occurring within the leaves, particularly during wet periods, NDII would possibly not reflect variation in root zone soil moisture content (Korres et al., 2015).
The analysis was done using data from eight sub-basins of the Upper Ping River basin (UPRB), a tropical seasonal evergreen catchment in northern Thailand. This catchment is adequate for the purpose because it has eight well-gauged sub-basins with clearly different aridity characteristics and strong seasonality, providing a good testing ground for the comparison.
The UPRB and the locations of the rain-gauge and runoff stations. The numbers indicate the 14 sub-basins of the UPRB.
The remotely sensed NDII values have been compared to the root zone storage as modelled by a semi-distributed conceptual model (semi-distributed meaning that for each sub-catchment a separate conceptual model has been used). The different sub-catchments demonstrate a variety of climatic properties that allow a more rigorous test than a fully lumped model could provide. In this way, a compromise has been found between the complexity and data requirements of a fully distributed model and the simplicity of a completely lumped model. One could argue that a fully distributed conceptual model would have been a better tool to assess the spatial and temporal pattern obtained by the NDII. This is correct, but this would have required the availability of more detailed spatially distributed forcing data (particularly rainfall), which were not available. Moreover, if a semi-distributed lumped model, potentially less accurate than a distributed model, provides a good correlation with NDVI, then this would be a tougher text than with a fully distributed model.
The UPRB is situated between latitude
17
Data from 65 non-automatic rain-gauge stations covering the period from 2001 to 2013 were used. A total of 42 stations are located within the UPRB while 23 stations are situated in its surroundings. These rain gauges are owned and operated by the Thai Meteorological Department and the Royal Irrigation Department. Quality control of the rainfall data was performed by comparing them to adjacent rainfall data. For each sub-basin, daily spatially averaged rainfall, by inverse distance squared, has been used as the forcing data of the hydrological model.
Daily runoff data from 1995 to 2011 at eight stations located in the UPRB were adequate to be used for FLEX calibration. These eight stations are operated by the Royal Irrigation Department in Thailand. The locations of these eight stations and the associated sub-basins are shown in Fig. 1. These eight stations control the runoff of the eight sub-basins on which the eight lumped conceptual models were calibrated. Runoff data at these stations are not affected by large reservoirs and have been checked for their reliability by comparing them with rainfall data covering their catchment areas at the same periods. Catchment characteristics and available data periods for model calibration of the selected eight sub-basins are summarized in Table 1.
The satellite data used for calculating the NDII is the MODIS level 3 surface
reflectance product (MOD09A1), which is available at 500 m resolution in an
8-day composite of the gridded level 2 surface reflectance products. Each
product pixel contains the best possible L2G observation during an 8-day
period selected on the basis of high observation coverage, low view angle,
absence of clouds or cloud shadow, and aerosol loading. MOD09 (MODIS Surface
Reflectance) is a seven-band product, which provides an estimate of the
surface spectral reflectance for each band as it would have been measured at
ground level without atmospheric scattering or absorption. This product has
been corrected for the effects of atmospheric gases and aerosols (Vermote et
al., 2011). The available MODIS data covering the UPRB from 2001 to 2013 were
downloaded from
Catchment characteristics and data period for selected eight sub-basins in the UPRB.
Water balance and constitutive equations used in FLEX
Parameter ranges of the FLEX model.
Estimates of vegetation water content (the amount of water in stems and
leaves) are of interest to assess the vegetation water status in agriculture
and forestry and have been used for drought assessment (Cheng et al., 2006;
Gao, 1996; Gao and Goetz, 1995; Ustin et al., 2004; Peñuelas et al.,
1993). Evidence from physically based radiative transfer models and
laboratory studies suggests that changes in water content in plant tissues
have a large effect on the leaf reflectance in several regions of the
0.7–2.5
On the basis of this idea, Hardisky et al. (1983) derived the NDII:
The 8-day NDII values, as collected from MODIS, were averaged over each
sub-basin to allow comparison to the 8-day average
We did not use field observations of soil moisture. One could argue that field observations should be used to link NDII to moisture stress. However, besides not being available, it is doubtful if point observations at fixed depth would provide a correct measure for the moisture content in the root zone. It is more likely that vegetation distributes its roots and adjusts its root density to the specific local conditions and that the root density and distribution is not homogeneous in space and depth.
Model structure of the FLEX.
FLEX (Fig. 2) is a conceptual hydrological model with an HBV-like model
structure developed in a flexible modelling framework (Fenicia et al., 2011;
Gao et al., 2014a, b). The model structure comprises four
conceptual reservoirs: the interception reservoir
The interception reservoir uses the water balance equation, Eq. (2),
presented in Table 2. The interception evaporation
The moisture content in the root zone is simulated by a reservoir
(Eq. 5) that partitions effective rainfall into infiltration and runoff
Average NDII values during the wet season, the dry season, and the whole year from 2001 to 2013, and their order of moisture content (range is 1–13; lower values indicate less NDII) for the entire UPRB.
In Eq. (8),
The linear response reservoirs, representing linear relationships between
storages and releases, are applied to conceptualize the discharge from the
fast runoff reservoir, and slow response reservoir. Eq. (12) presents the
water balance of the fast reservoir in which
A multi-objective calibration strategy has been adopted in this study to
allow for the model to effectively reproduce different aspects of the
hydrological response, i.e. high flow, low flow, and the flow duration curve.
The model was therefore calibrated to three Kling–Gupta (K–G) efficiencies (Gupta
et al., 2009): (1) the K–G efficiency of flows (
The MOSCEM-UA (Multi-Objective Shuffled Complex Evolution
Metropolis-University of Arizona) algorithm (Vrugt et al., 2003) was used as
the calibration algorithm to find the Pareto-optimal solutions defined by the
mentioned three objective functions. This algorithm requires three parameters
including the maximum number of iterations, the number of complexes, and the
number of random samples that is used to initialize each complex. To ensure
fair comparison, the parameters of MOSCEM-UA were set based on the number of
model parameters. Therefore, the number of complexes is equal to the number
of free parameters
Monthly average NDII values for the UPRB in 2004. The green colour
indicates an NDII between 0.15 and 0.30, yellow between 0 and 0.15, orange
between
Monthly average NDII values for six sub-basins compared to the basin average in the UPRB. Note that three wettest and three driest basins are presented in this graph.
Seasonal signals exist both in the NDII and
To demonstrate the spatial and seasonal behaviour of the NDII over the UPRB, the 8-day NDII values were aggregated to monthly values for 2001–2013. Figure 3 shows examples of monthly average NDII values for the UPRB in 2004, which is the year with the lowest annual average NDII value. The figure shows that NDII values are higher during the wet season (May–October) and lower during the dry season (November–April). The lower amounts of rainfall between November and April cause a continuous reduction of NDII values. On the other hand, higher amounts of rainfall between May and October result in increasing NDII values. However, NDII values appear to vary little between July and October.
The average NDII values during the wet season, the dry season, and the whole
year within the 13 years are presented in Table 4. The table also shows the
order of the NDII values from the highest (number 1) to the lowest
(number 13). It can be seen that the annual average NDII value for the whole
basin is approximately 0.165, while the average values during the wet and dry
season are about 0.211 and 0.118, respectively. The highest mean annual value
(NDII
Examples of flow duration curves and simulated hydrographs using FLEX at runoff stations P.20 and P.21.
Monthly average NDII values between 2001 and 2013 and the order of basin moisture content for each of 14 sub-basins within the UPRB.
FLEX parameters calibrated at eight runoff stations located in the UPRB.
FLEX model performance at eight runoff stations.
Exponential relationships between the average NDII values and
simulated root zone moisture storage (
Note:
The 8-day NDII values were also computed for each of the 14 tributaries within the UPRB from 2001 to 2013. Table 5 shows the monthly averaged NDII values between 2001 and 2013 and the ranking order for each of the 14 tributaries. The results suggest that the Nam Mae Taeng, Nam Mae Rim, and Upper Mae Chaem, which have higher mean annual NDII values, have a higher moisture content than other tributaries, while Nam Mae Haad, Nam Mae Li, and Ping River sections 2 and 3, with lower mean annual NDII values, have lower moisture content than other tributaries. Monthly average NDII values for these six tributaries are presented in Fig. 4. It can be seen that during the dry season, NDII values of the three tributaries with the lowest values are a lot lower than those of the three with the highest NDII values. However, NDII values for these two groups are not significantly different during the wet season. The figure also reveals that NDII values tend to continuously increase from relatively low values in March to higher values in June. The values slightly fluctuate during the wet season before sharply falling once again when the rainy season ends, and reach their minimum values in February.
Calibration of FLEX was done on the eight sub-catchments that have runoff
stations. The results are summarized in Table 6. The performance of the model
was quite good, as demonstrated in Table 7. In Fig. 5, the flow duration
curves of runoff stations P.20 and P.21 are presented as examples of model
performance. Table 7 shows the average Kling–Gupta efficiencies values for
The 8-day NDII values were compared to the 8-day average root zone moisture
storage values of the FLEX model. It appears that during moisture stress
periods, the relationship can be well described by an exponential function
for each of the eight sub-catchments. Table 8 presents the coefficients of the
exponential relationships as well as the coefficients of determination
(
Scatter plots between the average NDII and the average root zone
moisture storage (
Scaled time series, seasonality, and deseasonalized (dry season)
time series of the 8-day averaged NDII values compared to the
8-day averaged simulated root zone moisture storage (
Examples of deseasonalized and scaled time series of NDII and root zone
storage (
If the soil moisture in the root zone is above a certain threshold value,
then the leaves are not under stress. In the UPRB, this situation occurs
typically during the middle and late rainy season. The NDII then does not
vary significantly while the root zone moisture storage may still vary,
albeit above the threshold where moisture stress occurs. This causes a lower
correlation between NDII and root zone storage during wet periods.
Interestingly, even during the wet season dry spells can occur. We can see in
Fig. 6, that during such a dry spell, the NDII and
We can see that the
The deseasonalized results of dry periods in sub-catchments P.20 and P.21 are
shown in Fig. 7. We found these variations of deseasonalized NDII and
In bare soil, remote sensors can only detect soil moisture within a few
centimetres below the surface (
Normally, the moisture content of the surface layer is linked to the total amount of moisture in the root zone. Knowing the surface soil moisture, the root zone soil moisture can be estimated by an exponential decay filter (Albergel et al., 2008; Ford et al., 2014) or by models (Reichle, 2008). However, the surface soil moisture is only weakly related to root zone soil moisture (Mahmood and Hubbard, 2007); it only works if there is connectivity between the surface and deeper layers, and when a certain state of equilibrium has been reached (when the short-term dynamics after a rainfall event has levelled out). It is also observed that the presence of vegetation prevents the observation of soil moisture and further deteriorates the results (Jackson and Schmugge, 1991). Avoiding the influence of vegetation in observing soil moisture (e.g. by SMOS or SMAP) is seen as a challenge by some in the remote sensing community (Kerr et al., 2001; Entekhabi et al., 2010). Several algorithms have been proposed to filter out the vegetation impact (Jackson and Schmugge, 1991), also based on NDII (e.g. Yilmaz et al., 2008). But is vegetation a troublemaker, or does it offer an excellent opportunity to directly gauge the state of the soil moisture?
In this study, we found that vegetation, rather than becoming a problem, could become
key to sensing the storage dynamics of moisture in the root zone. The water
content in the leaves is connected to the suction pressure in the root zone
(Rutter and Sands, 1958). If the suction pressure is above a certain
threshold, then this connection is direct and very sensitive. We found a
highly significant correlation between NDII and
In natural catchments, it is not possible to prove a hypothesis by using a calibrated model. There are too many factors contributing to the uncertainty of results: the processes are too heterogeneous, the observations are not without error, the climatic drivers are too uncertain and heterogeneous, and finally, there is substantial model uncertainty, both in the semi-distributed hydrological model and in the remote sensing model used to determine the 8-day NDII product. In this case, we have selected a lumped conceptual model, which is good at mimicking the main runoff processes, but which lacks the detail of distributed models. Distributed models, however, require detailed and spatially explicit information (which is missing) and are generally over-parameterized, turning them highly unreliable in data-scarce environments. On top of this, there is considerable doubt if they provide the right answers for the right reasons.
This paper is not a modelling study but a test of the hypothesis whether the
observed NDII correlates with the modelled root zone storage. We have seen in
Fig. 6 that the correlation is strong during periods of moisture stress, but
that when the root zone is near saturation the correlation is weak. But we
also saw that even in the wet season, during short dry spells, the
correlation is strong. Even when the seasonality is removed, the patterns
between NDII and
Simulation of root zone soil moisture is crucial in hydrological modelling (Houser et al., 1998; Western and Blöschl, 1999). Using estimates of soil moisture states could increase model performance and realism, but moreover, it would be powerful information to facilitate prediction in ungauged basins (Hrachowitz et al., 2013). However, until now, it has not been practical (e.g. Parajka et al., 2006; Entekhabi et al., 2010). Assimilating soil moisture in hydrological models, either from top-soil observation by remote sensing, or from the deeper soil column by models (Reichle, 2008), is still a challenge. Several studies showed how difficult it is to assimilate soil moisture data to improve daily runoff simulation (Parajka et al., 2006; Matgen et al., 2012).
There are several reasons why we have not compared our results with soil moisture observations in the field. Firstly, observations of soil moisture are not widely available. Moreover, it is not straightforward to link classical soil moisture observations to the actual moisture available in the root zone. Most observations are conducted at fixed depths and at certain locations within a highly heterogeneous environment. Without knowing the details of the root distribution, both horizontally and vertically, it is hard, if not impossible, to estimate the water volume accessible to plants through their root systems. We should realize that it is difficult to observe root zone soil moisture even at a local scale. But measuring root zone soil moisture at a catchment scale is even more challenging. State-of-the-art remote sensing techniques can observe spatially distributed soil moisture, but what they can see is only the near-surface layers if not blocked by vegetation. The top layer moisture may or may not be correlated with the root zone storage, amongst others, depending on the vegetation type, but it is definitely not the same.
By observing the moisture content of the leaves, the NDII represents the soil
moisture content of the entire root zone, which is precisely the information
that hydrological models require as this is the component that controls the
occurrence and magnitude of storage deficits and thereby the moisture
dynamics of a system. This study clearly shows the temporal correlation
between
We should, of course, be aware of regional limitations. The proxy only appears to work for periods of moisture stress. This study considered a tropical seasonal evergreen ecosystem, where periods of moisture stress regularly occur. In ecosystems which shed their leaves or go dormant, other conditions may apply. We need further investigations into the usefulness of this approach in catchments with different climates. In addition, the phenology of the ecosystem is of importance, which should be taken into consideration in follow-up research. Finally, a comparison with distributed or semi-distributed models would be a further test of the value of the NDII as proxy for the root zone moisture content.
The NDII was used to investigate drought for the UPRB from 2001 to 2013. Monthly average NDII values appear to be spatially distributed over the UPRB, in agreement with seasonal variability and landscape characteristics. NDII values appear to be lower during the dry season and higher during the wet season as a result of seasonal differences between precipitation and evaporation. The NDII appears to correlate well with the moisture content in the root zone, offering a potential proxy variable for calibration of hydrological models in ungauged basins.
To illustrate the importance of NDII as a proxy for root zone moisture
content in hydrological models, we applied the FLEX model to assess the root
zone soil moisture storage (
The potential of using the NDII to constrain model parameters (such as the
power of the beta function
The data set can be found at:
We gratefully acknowledge Kasetsart University Research and Development Institute for financially supporting this research. We also appreciate Royal Irrigation Department and Thai Meteorological Department for providing the rainfall data. Finally, we sincerely thank the MODIS Land Discipline Group for creating and sharing the MODIS Land data used in this study.Edited by: B. van den Hurk Reviewed by: R. Teuling, J. Parajka, and one anonymous referee