Contents lists available atScienceDirect
Int J Appl Earth Obs Geoinformation
journal homepage:www.elsevier.com/locate/jagEffects of prediction accuracy of the proportion of vegetation cover on land
surface emissivity and temperature using the NDVI threshold method
Elnaz Neinavaz
a,⁎
, Andrew K. Skidmore
a,b, Roshanak Darvishzadeh
aaDepartment of Natural Resources, Faculty of Geo-Information Science and Earth Observation, University of Twente, Hengelosestraat 99, 7500 AE, Enschede, the Netherlands
bDepartment of Environmental Science, Macquarie University, NSW, 2109, Sydney, Australia
A R T I C L E I N F O Keywords:
Proportion of vegetation cover Thermal infrared remote sensing Land surface emissivity Land surface temperature Vegetation index Landsat-8
Artificial neural network
A B S T R A C T
Predicting land surface energy budgets requires precise information of land surface emissivity (LSE) and land surface temperature (LST). LST is one of the essential climate variables as well as an important parameter in the physics of land surface processes at local and global scales, while LSE is an indicator of the material composition. Despite the fact that there are numerous publications on methods and algorithms for computing LST and LSE using remotely sensed data, accurate prediction of these variables is still a challenging task. Among the existing approaches for calculating LSE and LST, particular attention has been paid to the normalised difference vege-tation index threshold method (NDVITHM), especially for agriculture and forest ecosystems. To apply NDVITHM,
knowledge of the proportion of vegetation cover (PV) is essential. The objective of this study is to investigate the
effect of the prediction accuracy of the PVon the estimation of LSE and LST when using NDVITHM. In August
2015, a field campaign was carried out in mixed temperate forest of the Bavarian Forest National Park, in southeastern Germany, coinciding with a Landsat-8 overpass. The PVwas measured in the field for 37 plots. Four
different vegetation indices, as well as artificial neural network approaches, were used to estimate PVand to
compute LSE and LST. The results showed that the prediction accuracy of PVimproved using an artificial neural
network (R2
CV= 0.64, RMSECV= 0.05) over classic vegetation indices (R2CV= 0.42, RMSECV= 0.06). The results
of this study also revealed that variation in the accuracy of the estimated PVaffected calculation results of the
LSE. In addition, our findings revealed that, though LST depends on LSE, other parameters should also be taken into account when predicting LST, as more accurate LSE results did not increase the prediction accuracy of LST.
1. Introduction
Vegetation biophysical and biochemical variables have been widely retrieved with differing degrees of success by means of remote sensing data from various parts of the electromagnetic spectrum (Bacour et al., 2006; Baret and Guyot, 1991; Baret et al., 1989;Boegh et al., 2002; Broge and Mortensen, 2002;Brown et al., 2000;Darvishzadeh et al., 2019,2008;Inoue et al., 2016;Neinavaz et al., 2017;Ullah et al., 2012; Verrelst et al., 2015;Zheng and Moskal, 2009). However, remote sen-sing data from the thermal infrared (TIR, 8–14 μm) spectrum have not been utilized as much for the estimation of vegetation variables, due to their coarse resolution (60m-1 km) and the difficulty in deriving and separating land surface emissivity (LSE) and land surface temperature (LST) (Dash, 2005). Despite all these challenges, TIR data have been more beneficial than visible/near-infrared (VNIR, 0.3–1.0 μm), and shortwave-infrared (SWIR, 1.0–2.5 μm) data for monitoring solar
effects day/night as well as diurnal and intra-annual temperature var-iations. In addition, TIR data have been shown to be complementary to other remote sensing data such as VNIR, and SWIR, and even to be unique in aiding the identification of surface materials and spectral characteristics for vegetation, rock types, and soil moisture (Prakash, 2000;Ullah et al., 2012). The atmosphere is mostly transparent across the TIR domain (Clerbaux et al., 2011), and TIR data can be measured day and night, allowing diurnal cycles to be observed and monitored. TIR spectral data have the advantage that they can extend observation and monitoring options for many types of natural phenomena, such as land and ocean surface temperatures, in which negligible variations may have a significant impact on climate and ecosystems (Smith et al., 2008).
LST, which is defined as the effective kinetic temperature of the Earth’s surface skin (Pinheiro et al., 2004), provides a valuable set of observations to characterise land surface states and land-atmosphere
https://doi.org/10.1016/j.jag.2019.101984
Received 26 June 2019; Received in revised form 20 September 2019; Accepted 1 October 2019
⁎Corresponding author.
E-mail addresses:e.neinavaz@utwente.nl(E. Neinavaz),a.k.skidmore@utwente.nl(A.K. Skidmore),r.darvish@utwente.nl(R. Darvishzadeh).
0303-2434/ © 2019 Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/).
exchange. LST is recognised as an essential climate variable (ECV) and is acknowledged as an essential parameter for diagnosing the behaviour of the Earth system and evaluating Earth system models (Bojinski et al., 2014) . Furthermore, LST has proved to be important in various aspects of crop management, such as stress detection, crop growth monitoring, yield forecasting, as well as in irrigation scheduling (Gonzalez-Dugo et al., 2013;Sruthi and Aslam, 2015;Tormann, 1986).
LSE plays an important role in computing LST (Gillespie et al., 1998; Sobrino et al., 2004), particularly when only a few TIR bands are available (Dash et al., 2002). The retrieval of LSE, which was defined by Sobrino et al. (2001) as a measure of the inherent efficiency of the surface in converting kinetic into radiant energy above the surface, is not an easy task. The prediction of LSE from passive satellite mea-surements may encounter difficulties due to, for instance, atmospheric absorption and emission, atmospheric contamination as well as surface reflection (Li et al., 2013b). However, knowledge of LSE values is
required for estimating some of the vital vegetation parameters in biodiversity monitoring, such as energy budget, evapotranspiration, water and energy balances (Jiménez-Muñoz et al., 2006; Seemann et al., 2008;Sobrino et al., 2002), as LSE has considerable control over the amount of thermal radiation emitted (Dash et al., 2002;Jacob et al., 2004).
Many studies have been carried out to derive LSE and LST from remotely sensed data (Becker and Li, 1995; Gillespie et al., 1998; Jiménez-Muñoz and Sobrino, 2003). A comprehensive overview on LSE and LST retrieval has been provided byLi et al. (2013a). Among the existing approaches for LSE and LST retrieval, special consideration has been given to the normalised difference vegetation index threshold method (NDVITHM) (Neinavaz et al., 2019;Oguz, 2013;Rajeshwari and Mani, 2014; Tang et al., 2015). NDVITHM was described byVan de Griend and Owe (1993)and has been further modified bySobrino and Raissouni (2000). Despite the challenges of using NDVITHM for Fig. 1. The location of the Bavarian Forest National Park, Germany, and the distribution of the sample plots.
ecosystems covered by rocks, ice, and water, NDVITHM has become widely accepted as a practical approach to retrieving the LSE and LST for agriculture and forest ecosystems (Oltra-Carrió et al., 2012;Sobrino et al., 2001;Tang et al., 2015). In order to use NDVITHM, the emissivity of bare soil and vegetation, as well as the proportion of vegetation cover (PV), must be known. PVis defined as the ratio of the vertical projection area of vegetation (containing leaves, stalks, and branches) on the ground to the total vegetation area (Deardorff, 1978). PVis one of the important vegetation biophysical variables associated with the Earth surface processes, as well as biodiversity monitoring, climate model-ling, and numerical weather prediction models (Gutman and Ignatov, 1998).
Among the existing approaches for estimation of PVfrom remote sensing data, vegetation indices, such as the normalised difference ve-getation index (NDVI), have been frequently used (Camacho-De Coca et al., 2004;Gutman and Ignatov, 1998;Jiapaer et al., 2011;Zeng et al., 2000). NDVI is generally saturated in areas with relatively high vege-tation cover (Atzberger, 2004; Baret and Guyot, 1991; Danson and Plummer, 1995;Gitelson et al., 2002). Therefore, other vegetation in-dices such as variable atmospherically resistant index (VARIgreen) Gitelson et al. (2002)are used to estimate PVusing remote sensing data (Jiménez-Muñoz et al., 2005, 2009). Jiménez-Muñoz et al. (2005)
revealed that VARIgreen could predict PVwith higher accuracy than the NDVI-approach could. Additionally, the wide dynamic range vegetation index (WDRVI), which was also developed byGitelson (2004), and a three-band gradient difference vegetation index (TGDVI) (Tang et al., 2005) may also be used for the retrieval of PV. Theoretically, the TGDVI mitigates the effect of the background to some extent and the influence of cirrus cloud, while having a high saturation point.
Furthermore, the sensitivity of the WDRVI to a high LAI value (dense vegetation cover) is at least three times greater than the sensi-tivity of the NDVI (Gitelson, 2004). An alternative approach for esti-mating PVis an artificial neural network (ANN), in which spectral re-flectance is applied as an input layer (Boyd et al., 2002). To our knowledge, the effect of prediction accuracy of the PVon the estimation of LSE has been explored over agricultural area (Jiménez-Muñoz et al., 2005), but has, to date, not been addressed mixed temperate forest. Therefore, in this study, we aim to assess the prediction accuracy of the PVcalculated using vegetation indices and an artificial neural network, to evaluate the effect of the PVprediction accuracy obtained by various methods for calculating LSE and eventually LST in mixed temperate forest.
Table 1
Summary statistics of the proportion of vegetation cover calculated using normalised difference vegetation index (PVNDVI), variable atmospherically resistant index
values (PVVARIgreen), the three-band gradient difference vegetation index (PVTGDVI), wide dynamic range vegetation index (PVWDRVI), artificial neural network (PVANN),
and in situ measurements (PVIn situ) for 37 plots in Bavarian Forest National Park.
Variables Range Minimum Maximum Mean Std. Deviation Variance Statistic Std. Error PVNDVI 0.37 0.53 0.91 0.766 0.013 0.084 0.007 PVVARIgreen 0.11 0.29 0.40 0.363 0.003 0.023 0.001 PVTGDVI 0.09 0.12 0.20 0.137 0.003 0.018 0.000 PVWDRVI 0.77 1.73 2.50 2.011 0.032 0.195 0.038 PVANN 0.43 0.40 0.83 0.614 0.015 0.095 0.009 PVIn situ 0.43 0.39 0.82 0.619 0.020 0.127 0.016
Fig. 2. Box plots are demonstrating the median, lower and upper quartile values of the proportion of vegetation cover calculated using normalised difference
vegetation index (PVNDVI), variable atmospherically resistant index values (PVVARIgreen), the three-band gradient difference vegetation index (PVTGDVI), wide dynamic
2. Materials and methods
2.1. Study area
The Bavarian Forest National Park (BFNP) is situated in south-eastern Germany along the border with the Czech Republic (49˚3′19″ N, 13˚12′9″E) (Fig. 1). The total area of the BFNP is 24,250 ha and has a temperate climate, with elevation ranging from 600 to 1453 m (Huber, 2005). There are three main forest types in the BFNP. Sub-alpine spruce, Norway spruce (Picea abies) and Mountain ash (Sorbus
aucu-paria) can be found at high altitudes, above 1100 m. On slopes and at
the lower altitudes, between 600 and 1100 m, Norway spruce, Silver fir (Abies alba), European beech (Fagus sylvatica), and Norway maple (Acer
pseudoplatanus) are visible. The valleys have Spruce, mixed with Birches
(Betula pendula, and Betula pubescens), Norway spruce, and Mountain ash. The dominant tree species in the BFNP is Norway spruce (67%), while European beech and Silver fir are found in 24.5%, and 26% of the park’s area, respectively (Heurich et al., 2010).
2.2. In situ measurement of the proportion of vegetation cover
The fieldwork was carried out in August 2015. The study area was divided into three stands, namely needle-leaf (conifer), broad-leaf as well as mixed forest. Consistent with the random sampling strategy, 37 plots were selected (plot size = 30 × 30 m). The coordinates of the center location of each plot were recorded with an accuracy of 1 m using a Leica system GPS 1200 (Leica Geosystems AG, Heerbrugg, Switzerland). Each plot was divided into four sub-plots using diagonal lines. In this study, PVwas measured for each sub-plot based on five upward-pointing digital hemispherical photography (DHP) according to the guideline proposed byZhou et al. (1998), using a Canon EOS 5D equipped with a fish-eye lens (Sigma 8 mm F3.5 EX DF), under clear sky condition. The camera was levelled on a tripod at about 1.3 m above ground level (i.e., breast height) (Whitmore et al., 1993). In order to reduce subjective thresh-holding on the blue channel for all captured
images and to classify sky and canopy pixels, the two-corner classifi-cation procedure was applied (Macfarlane, 2011). Further, CAN_EYE software was used to estimate PV. More information on the basic principle and specific features of CAN_EYE software can be found in Weiss and Baret (2017). The average of the PV’s estimated for each four sub-plots was considered to be the PVof that plot. The summary sta-tistics of the measured PVis presented inTable 1.
2.3. Satellite data and processing
Landsat-8 data were acquired on 9 August 2015. Digital numbers of the operational land imager (OLI) and thermal infrared (TIRS) images of Landsat-8 were converted to radiance value, using the coefficients provided by the United States Geological Survey (USGS, https:// landsat.usgs.gov/using-usgs-landsat-8-product). For the OLI images, radiance was converted to reflectance using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) module (Matthew et al., 2002). Since 2010, TIRS bands have been resampled by USGS to 30 m to match the OLI spectral bands’ resolution. In this study, we only considered band 10 of Landsat-8, as instability has been re-ported in the calibration of band 11 (Barsi et al. (2014).
2.4. Estimation of land surface emissivity and land surface temperature
One of the practical methods for determining LSE is the use of statistical relationships between NDVI derived from VNIR data and the vegetation and soil emissivity values (Van de Griend and Owe, 1993). This approach was later modified and called NDVITHM(Sobrino and Raissouni, 2000;Valor and Caselles, 1996). The LSE using NDVITHMis calculated according to the following equation (Sobrino and Raissouni, 2000;Sobrino et al., 2008): = < + + + × + NDVI b NDVI NDVI 0.2, 0.5, d 0.2 0.5, P (1 P ) d red v v v s v (1)
Fig. 3. Represents analysis of variance for the proportion of vegetation cover calculated using normalised difference vegetation index (PVNDVI), variable
atmo-spherically resistant index values (PVVARIgreen), artificial neural network (PVANN), wide dynamic range vegetation index (PVWDRVI), the three-band gradient difference
vegetation index (PVTGDVI), and in situ measurements (PVIn situ).
wherea andb denote channel-dependent regression coefficients; red is reflectance in the red region; v and s denote vegetation and bare soil emissivity over TIR band, respectively. In this study, the v and s were extracted from the MODIS emissivity library of the University of the California, Santa Barbara (USA) (Wan and Dozier, 1996). PVis the proportion of vegetation cover; dε is cavity effect, which is incon-sequential for flat surfaces, however, for diverse and rough surfaces can reach a value of 2% (Sobrino, 1989;Sobrino et al., 1990). dε can be computed by using the following equation:
=
d (1 s)(1 P Fv) v (2)
where F denotes shape factor, the mean value of which, assuming di-verse geometrical distributions, is 0.55 (Sobrino et al., 1990,2004).
To calculate LST using TIR data, the brightness temperature (BT) should be calculated, using the spectral radiance of TIR bands, and thermal constants (Markham, 1986):
= +
(
)
BT K In KL 1 2 1 (3)where L is spectral radiance at the top of the atmosphere, and K1and K2denote bands-specific thermal conversion constants that are avail-able at the Landsat-8 image metadata file. Eventually, the LST was computed applying the following equation, suggested byStathopoulou and Cartalis (2007): = + × ×
{
}
LST BT Ln 1 W BTp ( )i (4)where W denotes wavelength of emitted radiance, istands as LSE and
pcorresponds to 1.438×10−2mK. As LST is computed in Kelvin, it is converted to Celsius by subtracting 273.15. The estimated LSE and LST were subsequently evaluated for 37 plots.
2.4.1. Estimating the proportion of vegetation cover
As can be observed from Eq.(2), PVis an essential parameter for calculating LSE according to NDVITHM. There are several methods for estimating PVusing empirical models. In this study, PVwas estimated following the NDVI traditional method (Rouse et al., 1974) and then Fig. 4. Scatterplots of the measured versus the predicted proportion of vegetation cover calculated using artificial neural network (a), variable atmospherically
compared to the VARIgreen (Gitelson et al., 2002), WDRVI (Gitelson, 2004), and TGDVI (Tang et al., 2005) approaches, as well as to an ar-tificial neural network. PVcan be predicted using NDVI, VARIgreen, and WDRVI as follows:
= P NDVI NDVI NDVI NDVI V soil veg soil (5) = P VARIgreen VARIgreen VARIgreen VARIgreen V soil veg soil (6) = × × +
WDRVI ( RefNIR RedRed)/( RefNIR RefRed) (7.a)
= P WDRVI WDRVI WDRVI WDRVI ( ) ( ) V soil veg soil (7.b)
where vegetation indexvegand vegetation indexsoilrepresent the fully vegetated and bare soil pixels of the considered index, respectively. RefNIR and RefRed are reflectance values on near infrared, and red bands, respectively. Parameter α in Eq.(7.a), was empirically set to 0.3 (Jia et al., 2017). More details on estimating PVusing NDVI, VARIgreen, and WDRVI can be found inGutman and Ignatov (1998),Carlson and Ripley (1997), andJia et al. (2017), respectively.
The estimation of PV using TGDVI is put forward as follows, ac-cording toTang et al. (2005):
=
=
TGDVI Ref Ref Ref Ref
P TGDVI TGDVI NIR Red NIR Red Red Green Red Green V max (8)
Where RefGreenis the green band reflectance, and λNIR, λRed, and λGreen are corresponding wavelengths for RefNIR, RefRed, and RefGreen, re-spectively. TGDVImaxstands for the maximal three-band gradient dif-ference vegetation index. Finally, an ANN was used to estimate PV. In this case, the surface reflectance values from OLI bands were extracted for each plot and used for PVestimation as an input layer. For network training, one of the common training algorithms in back-propagation networks was used: the Levenberg-Marquardt algorithm (Hagan and
Menhaj, 1994). As the number of neurons in the hidden layer influences the performance of an ANN prediction, the ideal ANN size was de-termined by examining different numbers. The early stopping technique was used to avoid over-fitting.
The estimated PVobtained from the different approaches was then applied to compute the LSE according to Eq.(1)and eventually the LST. Thus, the LSE and LST values were extracted for the 37 plots. PV esti-mated using NDVI, VARIgreen, WDRVI, TGDVI, and an ANN, is referred to as PVNDVI, PVVARIgreen, PVWDRVI, PVTGDVI, and PVANN, respectively. Hereafter, the LSE and LST, which have been calculated using PVANN, PVNDVI, PVVARIgreen, PVWDRVI, PVTGDVI, and PVIn situ, respectively, are referred to as LSEPVANN, LSEPVNDVI, LSEPVVARIgreen, LSEPVWDRVI, LSEPVTGDVIand LSEPVIn situ, and as LSTPVANN, LSTPVNDVI, LSTPVVARIgreen, LSTPVWDRVI, LSTPVTGDVI, and LSTPVIn situ, respectively.
2.5. Validation
Two-way analysis of variance (ANOVA) was applied to explore whether the mean values of the estimated PV’s obtained with the var-ious methods differed significantly. For PVANN, linear regression ana-lyses were performed between the predicted and the measured PVto identify the best ANN model. The reliability of the ANN in the esti-mation of PV was evaluated using the cross-validated coefficient of determination, and cross-validated root mean squared error. In addi-tion, the cross-validation procedure was reiterated 1000 times, and results were averaged to minimize unwanted effects from the random initialization of the optimization routine. The estimated PV, LSE, and LST were evaluated using the cross-validated procedure to select the best model based on its predictive ability (Duda and Hart, 1973;Shao, 1993).
3. Results
3.1. Measured and estimated proportion of vegetation cover
The in situ measured PVranged from 0.39 to 0.82 and has a mean value of 0.61 for the 37 plots. Estimated PVANNand PVWDRVIranged Fig. 5. Box plots of the land surface emissivity using proportion of vegetation cover in which estimated using normalised difference vegetation index (LSEPvNDVI),
variable atmospherically resistant index values (LSEPvVARIgreen), the three-band gradient difference vegetation index (LSEPvTGDVI), wide dynamic range vegetation
index (LSEPvWDRVI), artificial neural network (LSEPvANN), and in situ measurements (LSEPvIn situ) approaches.
widely compared with the PVpredicted using other approaches (i.e., PVNDVI, PVVARIgreen, and PVTGDVI) (Table 1). In addition, the boxplot shows that the PVWDRVIis overestimated, while PVANNand PVNDVIare closer in range to the PvIn situ(Fig. 2). As can be seen inFig. 3, there are significant differences between PVIn situ and PV calculated using
empirical approaches (i.e., Vegetation indices); however, there is no significant difference between PVIn situand PVANN.
The relationships between the estimated PVand the measured PVare presented inFig. 4. As can be seen, PVANNis estimated with slightly greater accuracy (R2
CV= 0.64, RMSECV= 0.05) than PVVARIgreen Fig. 6. Land surface emissivity calculated over Bavaria Forest National Park using PVNDVI(a), PVVARIgreen(b), PVTGDVI(c), and PVWDRVI(d).
(R2
CV= 0.42, RMSECV= 0.06) and PVNDVI (R2CV= 0.408, RMSECV= 0.06).Fig. 4a also demonstrates that the ANN model is less insensitive to greater vegetation cover values (greater than 0.7) com-pared with other approaches. Additionally, the scatter plot inFig. 4b indicates that estimated PVVARIgreenmay be underestimated. The results also showed that PVis predicted with low accuracy when using WDRVI and TGDVI approaches (Fig. 4d & e).
3.2. Retrieval of land surface emissivity
The results indicate that the estimated ranges of LSEPVANN and LSEPVNDVIare closer to the LSEPVIn situthan those of LSEPVNDVIand LSEPVVARIgreenfor the 37 plots over BFNP (Fig. 5). The LSEPVWDRVI's estimated range is wider and set much further away from the LSEPVIn situ range. The results show that the LSE, estimated using PVNDVIand PVVARIgreen, disclosed a similar pattern over the BFNP, whereas the LSEPVTGDVI and LSEPVWDRVI displayed a differing pattern to the LSEPVNDVI, and LSTPVVARIgreen(Fig. 6). The two-way analysis of variance
reveals that estimated LSE using different methods are significantly different, except for LSEPVNDVIand LSEPVVARIgreen, which are not stati-cally significantly different from each other. The relationship between LSEPVInsitu and the estimated LSEPVANN, LSEPVNDVI, LSEPVVARIgreen, LSEPVTGDVI, and LSEPVWDRVIfor the 37 plots is presented inFig. 7. As can be seen, LSEPVANNwas predicted with reasonably high accuracy. LSEPVNDVIand LSTPVVARIgreenwere estimated with almost similar ac-curacy, although LSEPVNDVI was slightly overestimated, and LSEPVVARIgreen slightly underestimated (Fig. 7b and c). The results showed that the prediction accuracy for LSEPVTGDVIand LSEPVWDRVIwas very low.
3.3. Land surface temperature retrieval
In general, the results across BFNP show that LSTs estimated using different approaches exhibit very similar patterns despite having dif-ferent ranges (Fig. 8). The pattern and range of the LSTPVNDVI and LSTPVVARIgreen were similar, while the LSTPVWDRVI and LSTPVTGDVI Fig. 7. Scatterplots of the land surface emissivity calculated from in situ proportion of vegetation cover and the one estimated using artificial neural network (a),
normalised difference vegetation index (b), variable atmospherically resistant index (c), the three-band gradient difference vegetation index (d), and wide dynamic range vegetation index (e).
showed a different range. Additionally, the maximum value of the LST was calculated using PVWDRVI (LSTPVWDRVImax= 29.16 °C). As can be seen inFig. 8, LSTPVWDRVIand LSTPVTGDVIappear to be overestimated in comparison to LSTPVNDVIand LSTPVVARIgreen. The difference in surface temperature range between LSTPVNDVIand LSTPVVARIgreenis negligible (Fig. 9a), whereas the differences between LSTPVNDVI and both
LSTPVWDRVI and LSTPVTGDVIare greater. For instance, the difference between LSTPVNDVIand LSTPVWDRVI is 0.54 °C, which is remarkable (Fig. 9b). In addition, LSTPVANN is predicted with higher accuracy (R2
CV= 0.92, RMSECV= 0.16) (Fig. 10). The prediction accuracy of LSTPVNDVIand LSTPVVARIgreenis similar (R2CV= 0.21, RMSECV= 0.70). However, surprisingly, LSTPVWDRVI (R2
CV= 0.61, RMSECV= 0.48) and Fig. 9. The difference in land surface temperature between LSTPVNDVIand (a) LSTPVVARIgreen, (b) LSTPVTGDVI, and (c) LSTPVWDRVI.
LSTPVTGDVI (R2
CV= 0.61, RMSECV= 0.46) have been predicted with greater accuracy than LSTPVNDVIand LSTPVVARIgreen(Fig. 10). 4. Discussion
Previously,Sobrino et al. (2002)discussed how NDVITHMyielded promising results for retrieving LSE and LST. To apply NDVITHM, prior knowledge of soil and vegetation emissivities is essential (Becker and Li, 1995;Li et al., 2013b;Van de Griend and Owe, 1993). Notwith-standing the advantages of the NDVITHM, such as obtaining high-re-solution LSE and LST maps, as well as not requiring a sensor with a large number of TIR bands (Sobrino et al., 2008), some issues should still be taken into consideration. Our results reveal that in addition to information on soil and vegetation emissivity values, it is essential to estimate the PV with high accuracy in order to obtain LSE using NDVITHM. Our findings show that PVcan be predicted with different
degrees of accuracy depending on the applied method. For instance, using NDVI and VARIgreen resulted in near- similar results regarding the prediction accuracy of the PV, which is contrary to the findings of Jiménez-Muñoz et al. (2005; 2009), who suggested that VARIgreen could retrieve PVmore accurately than NDVI in agricultural areas (e.g., corn, alfalfa). In our study, PVis estimated with higher accuracy using a machine learning (ANN) approach (R2
CV= 0.64, RMSECV= 0.05) than using an empirical approach, such as VARIgreen (R2
CV= 0.42, RMSECV= 0.06) and NDVI (R2
CV= 0.40, RMSECV= 0.06). This result confirms the earlier results ofBoyd et al. (2002), who recommended the ANN approach as the preferred option for the retrieval of PV, because the machine learning technique offers a powerful means of analyzing without making any assumptions, even for complex datasets. In addi-tion, PV has been estimated with low accuracy using WDRVI, and TGDVI approaches, whereas it has been suggested previously that PV can be retrieved with reasonable accuracy over agricultural area with Fig. 10. Scatterplots of land surface temperature calculated from in situ proportion of vegetation cover and the one which is estimated using artificial neural network
(a), normalised difference vegetation index (b), variable atmospherically resistant index (c), the three-band gradient difference vegetation index (d), and wide dynamic range vegetation index (e).
both approaches (Gitelson, 2004;Tang et al., 2005). This study shows that PV might be predicted with reasonable accuracy over the agri-cultural area using vegetation indices; however, these approaches are not necessarily reliable for estimating PVover forest ecosystems.
The results of our study reveal that differences in prediction accu-racy of the PV(i.e., empirical and machine learning approaches) may lead to different results for the estimation of LSE. It should be noted that 1% uncertainty in prediction of the LSE can lead to an error of ap-proximately 0.5 K in the calculation of the LST under moderate condi-tions (Li et al., 2013a), with an obvious error of roughly 1 K expected in warmer and less humid environments (Chen et al., 2016). Therefore, for a heterogeneous ecosystem such as a forest, accurate estimation of the PVis paramount for computing LSE using NDVITHM, as LSE constitutes the primary source of error in the LST calculation (Jiménez-Muñoz and Sobrino, 2006).
The findings of this study reveal that, although the LSE is an im-portant variable for LST calculation, LSE is not the only factor, which may affect the prediction accuracy of LST. Unexpectedly, LSTPVWDRVI and LSTPVTGDVIwere estimated with higher accuracy than LSTPVNDVI and LSTPVVARIgreen, therefore suggesting that in addition to LSE and the atmospheric parameters on which LST depends, attention should be paid to other parameters, such as elevation (i.e., an abiotic variable) (Peng et al., 2017) and soil moisture (i.e., abiotic variable) (Sun and Pinker, 2004),. which have previously been suggested to influence LST. 5. Conclusion
This study aimed to explore the effect of the prediction accuracy of PVon computing LSE and LST using NDVITHM. PV, as one of the es-sential parameters for calculating LSE via NDVITHM, was estimated with reasonable accuracy using an ANN technique. The ANN-based models can be seen as a reliable approach for estimating PV by means of Landsat-8 data. Our findings have revealed that there is surprisingly no difference in the prediction accuracy of the PVbetween using NDVI and VARIgreen methods (empirical models) over the mixed temperate forest. This is in contradiction with previous studies and needs to be explored further and across different ecosystems. In addition, this study revealed that the WDRVI and TGDVI are not reliable approaches for estimating PV over mixed temperate forest. Our findings established that NDVITHMcan still be considered one of the most practical methods for estimating LST and LSE; however, accurate prediction of the PV remains an important factor as inaccuracy and uncertainty in the pre-diction of the PVcould lead to a significant error in the calculation of LSE. Additionally, in our view, the results of this study suggest that to calculate LST using NDVITHMmore parameters need to be considered, as LSE is not the only variable capable of influencing LST prediction. Declaration of Competing Interest
None.
Acknowledgments
This research received financial support from the EU Erasmus Mundus External Cooperation Window (EM8) Action 2 and was co-funded by the Natural Resources Department, Faculty of Geo-Information Science and Earth Observation, University of Twente, the Netherlands and Bavarian Forest National Park, Germany. The authors extend their appreciation for the data pool initiative for the Bohemian Forest Ecosystem and the excellent support received during fieldwork by Bavarian Forest National Park management.
References
Atzberger, C., 2004. Object-based retrieval of biophysical canopy variables using artificial neural nets and radiative transfer models. Remote Sens. Environ. 93, 53–67.
Bacour, C., Baret, F., Béal, D., Weiss, M., Pavageau, K., 2006. Neural network estimation of LAI, fAPAR, fCover and LAI× C ab, from top of canopy MERIS reflectance data: principles and validation. Remote Sens. Environ. 105, 313–325.
Baret, F., Guyot, G., 1991. Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sens. Environ. 35, 161–173.
Baret, F., Guyot, G., Major, D., 1989. TSAVI: a vegetation index which minimizes soil brightness effects on LAI and APAR estimation. In: Geoscience and Remote Sensing Symposium, 1989. IGARSS’89. 12th Canadian Symposium on Remote Sensing., 1989 International. 10-14 July, Vancouver, Canada: IEEE. pp. 1355–1358.
Barsi, J.A., Schott, J.R., Hook, S.J., Raqueno, N.G., Markham, B.L., Radocinski, R.G., 2014. Landsat-8 thermal infrared sensor (TIRS) vicarious radiometric calibration. Remote Sens. 6, 11607–11626.
Becker, F., Li, Z.L., 1995. Surface temperature and emissivity at various scales: definition, measurement and related problems. Remote. Sens. Rev. 12, 225–253.
Boegh, E., Søgaard, H., Broge, N., Hasager, C., Jensen, N., Schelde, K., Thomsen, A., 2002. Airborne multispectral data for quantifying leaf area index, nitrogen concentration, and photosynthetic efficiency in agriculture. Remote Sens. Environ. 81, 179–193.
Bojinski, S., Verstraete, M., Peterson, T.C., Richter, C., Simmons, A., Zemp, M., 2014. The concept of essential climate variables in support of climate research, applications, and policy. Bull. Amer Meteor. 95 (9), 1431–1443.
Boyd, D., Foody, G., Ripple, W., 2002. Evaluation of approaches for forest cover esti-mation in the Pacific Northwest, USA, using remote sensing. Appl. Geogr. 22, 375–392.
Broge, N.H., Mortensen, J.V., 2002. Deriving green crop area index and canopy chlor-ophyll density of winter wheat from spectral reflectance data. Remote Sens. Environ. 81, 45–57.
Brown, L., Chen, J.M., Leblanc, S.G., Cihlar, J., 2000. A shortwave infrared modification to the simple ratio for LAI retrieval in boreal forests: an image and model analysis. Remote Sens. Environ. 71, 16–25.
Camacho-De Coca, F., García-Haro, F., Gilabert, M., Meliá, J., 2004. Vegetation cover seasonal changes assessment from TM imagery in a semi-arid landscape. Int. J. Remote Sens. 25, 3451–3476.
Carlson, T.N., Ripley, D.A., 1997. On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote Sens Environ. 62 (3), 241–252.
Chen, F., Yang, S., Su, Z., Wang, K., 2016. Effect of emissivity uncertainty on surface temperature retrieval over urban areas: investigations based on spectral libraries. ISPRS J. Photogramm. Remote Sens. 114, 53–65.
Clerbaux, C., Drummond, J.R., Flaud, J.-M., Orphal, J., 2011. Using thermal infrared absorption and emission to determine trace gases. The Remote Sensing of Tropospheric Composition From Space. Springer, Berlin, Heidelberg, pp. 123–151.
Danson, F., Plummer, S., 1995. Red-edge response to forest leaf area index. Remote Sens. 16, 183–188.
Darvishzadeh, R., Skidmore, A., Abdullah, H., Cherenet, E., Ali, A., Wang, T., Nieuwenhuis, W., Heurich, M., Vrieling, A., O’Connor, B., 2019. Mapping leaf chlorophyll content from Sentinel-2 and RapidEye data in spruce stands using the invertible forest reflectance model. Int. J. Appl. Earth Obs. Geoinf. 79, 58–70.
Darvishzadeh, R., Skidmore, A., Schlerf, M., Atzberger, C., Corsi, F., Cho, M., 2008. LAI and chlorophyll estimation for a heterogeneous grassland using hyperspectral mea-surements. ISPRS J. Photogramm. Remote. Sens. 63, 409–426.
Dash, P., 2005. Land Surface Temperature and Emissivity Retrieval from Satellite Measurements. FZKA Karlsruhe.
Dash, P., Göttsche, F.-M., Olesen, F.-S., Fischer, H., 2002. Land surface temperature and emissivity estimation from passive sensor data: theory and practice-current trends. Int. J. Remote Sens. 23, 2563–2594.
Deardorff, J.W., 1978. Efficient prediction of ground surface temperature and moisture, with inclusion of a layer of vegetation. J. Geophys. Res-Oceans. 83 (C4), 1889–1903.
Duda, R.O., Hart, P.E., 1973. Pattern Classification and Scene Analysis. A Wiley-inter-science Publication. Wiley, New York 1973.
Gillespie, A., Rokugawa, S., Matsunaga, T., Cothern, J.S., Hook, S., Kahle, A.B., 1998. A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images. IEEE Trans. Geosci. Remote. Sens. 36, 1113–1126.
Gitelson, A.A., 2004. Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation. J. Plant Physiol. 161, 165–173.
Gitelson, A.A., Kaufman, Y.J., Stark, R., Rundquist, D., 2002. Novel algorithms for remote estimation of vegetation fraction. Remote Sens. Environ. 80, 76–87.
Gonzalez-Dugo, V., Zarco-Tejada, P., Nicolás, E., Nortes, P.A., Alarcón, J.J., Intrigliolo, D.S., Fereres, E.J.P.A., 2013. Using high resolution UAV thermal imagery to assess the variability in the water status of five fruit tree species within a commercial orchard. Precis Agric. 14 (6), 660–678.
Gutman, G., Ignatov, A., 1998. The derivation of the green vegetation fraction from NOAA/AVHRR data for use in numerical weather prediction models. Int. J. Remote Sens. 19, 1533–1543.
Hagan, M.T., Menhaj, M.B., 1994. Training feedforward networks with the Marquardt algorithm. IEEE Trans. Neural Netw. 5, 989–993.
Heurich, M., Beudert, B., Rall, H., Křenová, Z., 2010. National Parks As Model Regions for Interdisciplinary Long-term Ecological Research: the Bavarian Forest and Šumavá National Parks Underway to Transboundary Ecosystem Research. Long-term Ecological Research. Springer, Dordrecht, the Netherlands, pp. 327–344.
Huber, C., 2005. Long lasting nitrate leaching after bark beetle attack in the highlands of the Bavarian Forest National Park. J. Environ. Qual. 34, 1772–1779.
Inoue, Y., Guérif, M., Baret, F., Skidmore, A., Gitelson, A., Schlerf, M., Darvishzadeh, R., Olioso, A., 2016. Simple and robust methods for remote sensing of canopy chlor-ophyll content: a comparative analysis of hyperspectral data for different types of vegetation. Plant Cell Environ. 39, 2609–2623.
Jacob, F., Petitcolin, F., Schmugge, T., Vermote, E., French, A., Ogawa, K., 2004.
Comparison of land surface emissivity and radiometric temperature derived from MODIS and ASTER sensors. Remote Sens. Environ. 90, 137–152.
Jia, K., Li, Y., Liang, S., Wei, X., Mu, X., Yao, Y., 2017. Fractional vegetation cover esti-mation based on soil and vegetation lines in a corn-dominated area. Geocarto Int. 32, 531–540.
Jiapaer, G., Chen, X., Bao, A., 2011. A comparison of methods for estimating fractional vegetation cover in arid regions. Agric. For. Meteorol. 151, 1698–1710.
Jiménez-Muñoz, J., Sobrino, J., Guanter, L., Moreno, J., Plaza, A., Martínez, P., 2005. Fractional vegetation cover estimation from PROBA/CHRIS data: methods, analysis of angular effects and application to the land surface emissivity retrieval. In: Proc. 3rd Workshop CHRIS/Proba. ESRIN, Frascati, Italy.
Jiménez-Muñoz, J.C., Sobrino, J.A., Gillespie, A., Sabol, D., Gustafson, W.T., 2006. Improved land surface emissivities over agricultural areas using ASTER NDVI. Remote Sens. Environ. 103, 474–487.
Jiménez-Muñoz, J.C., Sobrino, J.A., Plaza, A., Guanter, L., Moreno, J., Martínez, P., 2009. Comparison between fractional vegetation cover retrievals from vegetation indices and spectral mixture analysis: case study of PROBA/CHRIS data over an agricultural area. Sensors 9, 768–793.
Jiménez-Muñoz, J., Sobrino, J., 2006. Error sources on the land surface temperature re-trieved from thermal infrared single channel remote sensing data. Int. J. Remote Sens. 27, 999–1014.
Jiménez-Muñoz, J.C., Sobrino, J.A., 2003. A generalized single-channel method for re-trieving land surface temperature from remote sensing data. J. Geophys. Res. Atmos. 108.
Li, Z.-L., Tang, B.-H., Wu, H., Ren, H., Yan, G., Wan, Z., Trigo, I.F., Sobrino, J.A., 2013a. Satellite-derived land surface temperature: current status and perspectives. Remote Sens. Environ. 131, 14–37.
Li, Z.-L., Wu, H., Wang, N., Qiu, S., Sobrino, J.A., Wan, Z., Tang, B.-H., Yan, G., 2013b. Land surface emissivity retrieval from satellite data. Int. J. Remote Sens. 34, 3084–3127.
Macfarlane, C., 2011. Classification method of mixed pixels does not affect canopy me-trics from digital images of forest overstorey. Agric. For. Meteorol. 151, 833–840.
Markham, B.L., 1986. Landsat MSS and TM post-calibration dynamic ranges, exoatmo-spheric reflectances and at-satellite temperatures. Landsat Tech. Notes 1, 3–8.
Matthew, M.W., Adler-Golden, S.M., Berk, A., Felde, G., Anderson, G.P., Gorodetzky, D., Paswaters, S., Shippert, M., 2002. Atmospheric correction of spectral imagery: eva-luation of the FLAASH algorithm with AVIRIS data. In: Applied Imagery Pattern Recognition Workshop, 2002. Proceedings. IEEE. pp. 157–163.
Neinavaz, E., Darvishzadeh, R., Skidmore, A.K., Abdullah, H., 2019. Integration of landsat-8 thermal and visible-short wave infrared data for improving prediction ac-curacy of forest leaf area index. Remote Sens. 11, 390.
Neinavaz, E., Skidmore, A.K., Darvishzadeh, R., Groen, T.A., 2017. Retrieving vegetation canopy water content from hyperspectral thermal measurements. Agric. For. Meteorol. 247, 365–375.
Oguz, H., 2013. Lst calculator: a program for retrieving land surface temperature from landsat TM/ETM+ imagery. Environ. Eng. Manage. J. 12, 549–555.
Oltra-Carrió, R., Sobrino, J., Franch, B., Nerry, F., 2012. Land surface emissivity retrieval from airborne sensor over urban areas. Remote Sens. Environ. 123, 298–305.
Peng, W., Zhou, J., Wen, L., Xue, S., Dong, L., 2017. Land surface temperature and its impact factors in Western Sichuan Plateau, China. Geocarto Int. 32, 919–934.
Pinheiro, A.C.T., Privette, J.L., Mahoney, R., Tucker, C.J., 2004. Directional effects in a daily AVHRR land surface temperature dataset over Africa. IEEE Trans. on Geosci. Remote Sens. 42 (9), 1941–1954.
Prakash, A., 2000. Thermal remote sensing: concepts, issues and applications. ISPRS Int. Archives. Photogramm. Remote. Sens. 33 (B1; PART 1), 239–243.
Rajeshwari, A., Mani, N., 2014. Estimation of land surface temperature of Dindigul dis-trict using Landsat 8 data. Int. J. Res. Eng. Technol. 3, 122–126.
Rouse, J., Haas, R., Schell, J., Deering, D., Harlan, J., 1974. Monitoring the Vernal Advancements and Retrogradation of Natural Vegetation. NASA/GSFC, Final Report. Greenbelt, MD, USA. pp. 1–137.
Seemann, S.W., Borbas, E.E., Knuteson, R.O., Stephenson, G.R., Huang, H.-L., 2008. Development of a global infrared land surface emissivity database for application to clear sky sounding retrievals from multispectral satellite radiance measurements. J. Appl. Meteorol. Climatol. 47, 108–123.
Shao, J., 1993. Linear model selection by cross-validation. J. Am. Stat. Assoc. 88, 486–494.
Smith, T.M., Reynolds, R.W., Peterson, T.C., Lawrimore, J., 2008. Improvements to NOAA’s historical merged land–ocean surface temperature analysis (1880–2006). J. Clim. 21, 2283–2296.
Sobrino, J., 1989. Desarrollo de un modelo teórico para implementar la medida de la temperatura realizada mediante teledetección. Aplicación a un campo de naranjos. PhD dissertation. University of Valencia, Valencia, Spain.
Sobrino, J., Caselles, V., Becker, F., 1990. Significance of the remotely sensed thermal infrared measurements obtained over a citrus orchard. ISPRS J. Photogramm. Remote. Sens. 44, 343–354.
Sobrino, J., Raissouni, N., 2000. Toward remote sensing methods for land cover dynamic monitoring: application to Morocco. Int. J. Remote Sens. 21, 353–366.
Sobrino, J., Raissouni, N., Li, Z.-L., 2001. A comparative study of land surface emissivity retrieval from NOAA data. Remote Sens. Environ. 75, 256–266.
Sobrino, J.A., Jiménez-Muñoz, J.C., Paolini, L., 2004. Land surface temperature retrieval from LANDSAT TM 5. Remote Sens. Environ. 90, 434–440.
Sobrino, J.A., Jiménez-Muñoz, J.C., Sòria, G., Romaguera, M., Guanter, L., Moreno, J., Plaza, A., Martínez, P., 2008. Land surface emissivity retrieval from different VNIR and TIR sensors. IEEE Trans. Geosci. Remote. Sens. 46, 316–327.
Sobrino, J.A., Jiménez-Muñoz, J.C., Labed-Nachbrand, J., Nerry, F., 2002. Surface emissivity retrieval from digital airborne imaging spectrometer data. J. Geophys. Res. Atmos. 107.
Sruthi, S., Aslam, M.M., 2015. Agricultural drought analysis using the NDVI and land surface temperature data; a case study of Raichur district. Aquat Procedia. 4, 1258–1264.
Stathopoulou, M., Cartalis, C., 2007. Daytime urban heat islands from landsat ETM+ and Corine land cover data: an application to major cities in Greece. Sol. Energy 81, 358–368.
Sun, D., Pinker, R.T., 2004. Case study of soil moisture effect on land surface temperature retrieval. IEEE Geosci. Remote. Sens. Lett. 1, 127–130.
Tang, B.-H., Shao, K., Li, Z.-L., Wu, H., Tang, R., 2015. An improved NDVI-based threshold method for estimating land surface emissivity using MODIS satellite data. Int. J. Remote Sens. 36, 4864–4878.
Tang, S., Zhu, Q., Wang, J., Zhou, Y., Zhao, F., 2005. Principle and application of three-band gradient difference vegetation index. Sci. China Ser. D Earth Sci. 48, 241–249.
Tormann, H., 1986. Canopy temperature as a plant water stress indicator for nectarines. South African J. Plant. Soil. 3 (3), 110–114.
Ullah, S., Skidmore, A.K., Naeem, M., Schlerf, M., 2012. An accurate retrieval of leaf water content from mid to thermal infrared spectra using continuous wavelet ana-lysis. Sci. Total Environ. 437, 145–152.
Valor, E., Caselles, V., 1996. Mapping land surface emissivity from NDVI: application to European, African, and South American areas. Remote Sens. Environ. 57, 167–184.
Van de Griend, A., Owe, M., 1993. On the relationship between thermal emissivity and the normalized difference vegetation index for natural surfaces. Int. J. Remote Sens. 14, 1119–1131.
Verrelst, J., Rivera, J.P., Veroustraete, F., Muñoz-Marí, J., Clevers, J.G., Camps-Valls, G., Moreno, J., 2015. Experimental Sentinel-2 LAI estimation using parametric, non-parametric and physical retrieval methods–a comparison. ISPRS J. Photogramm. Remote. Sens. 108, 260–272.
Wan, Z., Dozier, J., 1996. A generalized split-window algorithm for retrieving land-sur-face temperature from space. IEEE Trans. Geosci. Remote. Sens. 34, 892–905.
Weiss, M., Baret, F., 2017. CAN_EYE V6. 4.91 user manual. French National Institute for Agriculture Research. INRA, Toulouse, France.
Whitmore, T., Brown, N., Swaine, M., Kennedy, D., Goodwin-Bailey, C., Gong, W.-K., 1993. Use of hemispherical photographs in forest ecology: measurement of gap size and radiation totals in a Bornean tropical rain forest. J. Trop. Ecol. 9, 131–151.
Zeng, X., Dickinson, R.E., Walker, A., Shaikh, M., DeFries, R.S., Qi, J., 2000. Derivation and evaluation of global 1-km fractional vegetation cover data for land modeling. J. Appl. Meteorol. 39, 826–839.
Zheng, G., Moskal, L.M., 2009. Retrieving leaf area index (LAI) using remote sensing: theories, methods and sensors. Sensors 9, 2719–2745.
Zhou, Q., Robson, M., Pilesjo, P., 1998. On the ground estimation of vegetation cover in Australian rangelands. Int. J. Remote Sens. 19, 1815–1820.
