Mapping land degradation risk: Potential of the non-evaporativefraction using Aster and MODIS data
M. García, S. Contreras, F. Domingo & J. PuigdefábregasEstación Experimental de Zonas Áridas (CSIC), Almería, Spain
ABSTRACT: Land degradation is associated with decreases in resources retention by ecosystems.In water-limited environments this loss of functionality has an impact in the water use efficiencywhich should be reflected in the partition of surface energy fluxes through the actual evapotranspi-ration or latent heat. Remote-sensing indicators of land degradation based on energy ratios, such asthe non-evaporative fraction, could help to monitor land condition. In this study, we test a simpleoperational model for calculating energy fluxes in a semiarid mountainous region at two differentspatial resolutions (90 m and 1 km), using Aster and MODIS data on 18-07-2004. Results showthat Aster and MODIS results are comparable within reported instrumental errors. However, thelost of detail is remarkable. If the processes of land degradation related with changes in the surfaceenergy balance are explicit at 1 km, which needs further elucidation, MODIS is an adequate toolto perform regional assessments by means of its high temporal and spatial coverage. Comparisonswith field data show low net radiation errors and large errors for sensible heat but within the rangesobtained by other authors. The spatial patterns for the ratio non-evaporative fraction (NEF) pro-posed as an indicator of land condition are coherent with the surface type. Using NEF and NDVI,a reliable identification of disturbed sites with high risk of degradation and non-degraded sites isaccomplished. Non-degraded sites are better identified using NDVI rescaled only, while degradedsites are better classified using only NEF. Therefore, despite the fact that vegetation cover is aclear symptom of land degradation, indicators not directly related with vegetation cover, basedon the surface energy balance, such as the NEF (non-evaporative fraction), can reveal importantinformation about ecosystem functioning. Because of the limited verification data and dates, theseresults are preliminary and need further testing.
1 INTRODUCTION
Currently, there is a lack of standard and operational procedures for monitoring land degradationover large regions (Puigdefábregas & Mendizábal 2003). Land degradation processes cause dis-turbances in structure and functioning of landscapes leading to decreases in its resources retentioncapacity (Ludwig & Tongway 1997). A theoretical continuum of functionality can be establishedfrom landscapes that effectively trap, store, concentrate and utilize resources to those character-ized by a severe degradation stage in which all the resources are lost (Ludwig & Tongway 2000).Desertification has been associated with higher spatio-temporal heterogeneity of water and otherresources (Schlesinger et al. 1990). To monitor landscape condition and measure the magnitudeof land degradation processes, simple indicators based on structural properties of the ecosys-tem or more complex indicators associated with water (Sharma 1998, Boer & Puigdefábregas2003, 2005) and energy fluxes can be used (Wang & Takahashi 1998). The indicators based onthe water or energy balance are linked through the latent heat flux (λE) or evapotranspiration.Calculation of surface energy ratios requires spatially disaggregated estimates of surface energybalance components at temporal scales compatible with the temporal scales of land degradationprocesses.
The law of conservation of energy states that the available energy reaching a surface is dissipatedmainly as latent heat (λE) and sensible heat (H):
Rn − G = λE + H (1)
being Rn = net radiation; G = soil heat flux; and Rn−G = available energy. To select an appro-priate energy balance indicator, it is important to consider the low value of latent heat fluxes andthe evaporative fraction (EF) during several months in semi-arid areas. For this reason, we explorethe non-evaporative fraction, NEF, as an indicator for land degradation as in (2).
NEF = 1 − EF = 1 − λE
λE + H= 1 − λE
Rn − G= H
Rn − G(2)
The NEF should present a wider range of variability than EF and a higher signal-to-noise ratio.For instance, in the study region, latent heat is within error level of models during several days(Domingo et al. 2001).
The algorithms to calculate these energy components use information in the solar and thermalrange, being remote sensing the only data source providing radiometric temperature and vegetationcover observations over large extents. This is crucial as these variables explain most of the partitionof the available energy into sensible and latent heat (Kustas & Norman, 1996).
The most vulnerable areas to land degradation are located in arid regions (Safriel et al. 2003where the development of an operative system would require data such as MODIS (ModerateResolution Imaging Spectrometer), available at high temporal resolution but with 1 km pixel size.This questions the validity of models originally designed for agricultural areas or almost idealconditions when these models are applied over arid regions and heterogeneous sites with sparsevegetation covers (Chehbouni et al. 1997, Wassenaar et al. 2002).
In this study, we test two simple operational models (Jackson et al. 1977, Seguin & Itier 1983,Carlson et al. 1995, Roerink et al. 2000) for calculating energy fluxes in a semi-arid region attwo different spatial resolutions (90 m and 1 km). We expect that the increase in bare soil due todecreases in vegetation cover taking place as a consequence of land degradation should increasesurface temperature and albedo causing increases in H, and decreases in Rn similarly to resultspresented by Dolman et al. (1977) in the Sahel. Feedback effects, such as those occuring betweenalbedo and surface temperature might counterbalance some of these impacts (Phillips 1993). Thiswork aims to elucidate some of these aspects. The specific objectives are:
1. Evaluate two simple daily energy balance algorithms in a mountainous semi-arid region inSE Spain characterized by its high land cover heterogeneity and fragmentation.
2. Compare the land surface energy fluxes estimates result from the application of two spatialresolution data: MODIS with 1 km. and Aster (Advanced Spaceborne Thermal Emission andReflection Radiometer) with 90 m.
3. Map land degradation risk in a semiarid mountainous region, Sierra Gador, using the non-evaporative fraction (NEF) as an indicator for land degradation.
Aster is currently the only sensor collecting multispectral thermal infrared data at high spatialresolution being very appropriate for testing of models and direct ground comparisons (Frenchet al. 2005). On the other hand, both sensors, Aster and MODIS, are on board the Terra platform,allowing analyses of scale. Recent work has been done to compare both sensors showing goodagreement (<1 K) (Jacob et al. 2004). It is desirable to extend this type of comparisons to othervariables and regions.
2 STUDY SITE AND DATA
The study region (Fig. 1) located in South East Iberian Peninsula (Almería, Spain) comprises3600 km2 (36.95◦ N, 2.58◦W). This region is characterized by its heterogeneity and the abrupt
Figure 1. Location of the study site in South East Spain, Almeria province. The left panel shows MODISTrue Color Composite corresponding to the study site and 18-07-2004. The right panel shows the study sitewith a false colour composite for ASTER 18-04-07 image with image relief. Location of the eddy covariancesystem is shown by the arrow (see colour plate page 393).
relief changes with altitudinal gradients ranging from sea level up to 2800 m (a.s.l.) in SierraNevada mountain. Precipitation and temperature regimes present wide contrasts driven by theorography (Lopez-Bermúdez et al. 2005). Annual precipitation is the lowest in the Tabernas desert,with less than 200 mm, while in the mountains can be enough to sustain forest growth, rangingbetween 400 mm up to 700 mm.
In the center of the study site, Sierra Gador mountain-range covers 552 km2. During the 18th and19th centuries, this range area was subjected to an intense and widespread deforestation for shipconstruction and mining purposes. The original vegetation composed of oaks (Quercus ilex L. andQuercus faginea Lam.), olive trees (Olea europaea L.) poplars (Populus L. spp.) and strawberry-trees (Arbutus unedo Lam.) was extensively cut down and nowadays is dominated by a sparseshrubland dominated by Genista cinerea Vill. and mixed with rock outcrops, bare soil or grasslandvegetation mainly dominated by Festuca scariosa Lag. Around 73% of Sierra de Gádor presentsthis pattern with vegetation cover lower than the 50%. The shrubland with a small cover of densepine (Pinus L. spp.) woodland represents the second natural land cover type (12% of the area). Onlya 1.5% of the land is covered by dense woodlands of reforestation pines. Agriculture covers the9% of Sierra de Gádor and is dominated by a mixture of unirrigated and irrigated lands (almondsand olive trees) (Contreras 2006).
The rest of the image includes part of the Natural Park of Sierra Nevada including pine forests,oak relicts, and shrublands. In the northeast is the natural desert of Tabernas with a complextopography comprising an area of badlands. Along the Andarax ephemeral river which flows byAlmeria city, there is a mosaic of citrus orchards and vinegrapes. One of the most outstandingfeatures of the scene is the large plastic greenhouse area covering more than 330 km2. This uniquecombination of land covers and uses allows using this area as a pilot site where to test a sim-ple model for calculating surface energy fluxes that could be extended to regional and globalscales.
For this study, we have used Aster and MODIS data acquired on July-18th-2004 at 11.00 UTC.The Aster products used were 2AST07 which is surface reflectance at 15 m (VNIR) and 30 m(SWIR), and 2AST08 kinetic temperature at 90 m. Two MODIS products were used. To decreaseerrors due to cloud coverage and bidirectional reflectance, we took the 8 day surface reflectanceat 500 m (MOD09). During this 8 day period, albedo and vegetation indices can be assumed tobe constant. Finally, daily land surface temperature product (MOD11) at 1 km was used. In the caseof MODIS, surface temperature errors range from 1–3 K with no incidences for ASTER, wherethe reported absolute precision is 1–4 K.
A digital elevation model (DEM) from USGS (United States Geological Survey) available at30 m resolution and a digital orthophoto (from the Andalusian Regional Government) at 0.5 mwere used at different stages of the study.
Instrumental field data are acquired continuously at Llano de los Juanes experimental field sitesince 2003. They were used to compare with model results. The eddy covariance technique wasused to measure latent and sensible heat flux using a three dimensional sonic anemometer CSTAT3and a krypton hygrometer KH20, both from Campbell Scientific Inc., USA. Llano de los Juanes isa representative flat area of ∼2 km2 located at 1600 m in the well developed karstic high plain ofSierra de Gador. The vegetation cover is 50%–60% and consists mainly of patchy dwarf perennialshrubs (30%–35%) dominated by Genista pumilla, Thymus serpylloides Bory and Hormathopyllaspinosa L. and grasses (20%–25%) dominated by Festuca scariosa Lag. and Brachypodium retusumPers. (Li et al. in press). Mean NDVI measured in Llano de los Juanes with a Dycam camera in Juneof 2004 was 0.296. Net radiation (NR-LITE; Kipp & Zonen, Delft, Netherlands), relative humidity(thermohygrometer HMP 35 C, Campbell Scientific, Logan, UT, USA) sonic and soil temperature(SBIB sensors) are also continuously measured at the site. Besides air temperature measurements atLlano de los Juanes field site, air temperature was measured from other 11 meteorological stationsat the time of the satellite overpass (11.00 UTC). At Sierra Gádor mean ± standard deviation valuesderived from Aster data for July-18th-2004 for albedo and NDVI were respectively: 0.19 ± 0.030and 0.29 ± 0.08. At Llano de los Juanes experimental field site, mean values were: 0.186 (albedo),0.269 (NDVI), very similar to the Dycam measurement from one month ago, for what it is expectedthat vegetation cover at the time of satellite overpass will be also very similar to measured valuesin June (around 50%–60%).
3 ESTIMATION OF THE NON-EVAPORATIVE FRACTION USING THE H/RN RATIO
The NEF (non-evaporative fraction) that could be linked to land degradation processes, was esti-mated with Aster data using the ratio of sensible heat versus net radiation, H/Rn (Seguin & Itier1983, Carlson et al. 1995). At a daily scale, soil heat flux (G) is negligible in relation to the othercomponents of the balance (Kustas & Norman 1995) in this case:
NEF = 1 − EF = 1 − λE
Rn − G= H
Rn − G= H
Rn(3)
3.1 Sensible heat flux (H)
H can be estimated by a model of turbulent transport from the surface to the lower atmospherebased on surface layer similarity of mean profiles of temperature and wind speed using a resistanceform:
H = ρ · CpTs − Tair
rh(4)
where: Ts is land surface temperature; Tair is air temperature, both at the time of image acquisition;rh is the atmospheric resistance to the transfer of H; ρ and Cp are air density and specific heat atconstant pressure respectively. B can be defined as an exchange coefficient to sensible heat transfer(Jackson et al. 1997, Seguin & Itier 1983) and is calculated as:
B = ρ · Cp
rh(5)
Being rh a turbulent exchange coefficient dependent on wind velocity, aerodynamic roughnesslength, roughness length for heat transfer and Monin-Obukov length (Brutsaert 1982).
As having estimates of these variables at large scales is difficult, more operationalparameterizations have been proposed. Seguin & Itier (1983) found a global mean value for B= 0.25 for unstable conditions (Ts − Tair > 0) consistent with analytical calculations. In addition,we test the approach from Carlson et al. (1995). They shown that the main factor affecting con-ductance to heat transfer is vegetation cover and established a linear relation between the exchangecoefficient to sensible heat transfer B and fractional cover. At this date, NDVI from bare soil atthe study site was: 0.16 ± 0.012 (mean ± standard deviation) and from complete vegetation cover:0.68 ± 0.20. Mean values from bare soil and complete vegetation were taken to calculate B.
3.1.1 Air temperature (Tair)Air temperature is used in the estimation of sensible heat flux, H, and also of net radiation, Rn.To avoid relying on meteorological information, air temperature was estimated from the imagesusing the triangle NDVI-Ts proposed by Carlson et al. (1995). The apex of the NDVI-Ts space (highNDVI and low temperature) should correspond to pixels with high NDVI located at the wet edgeof the triangle, and can be assumed to be at the air temperature (Idso & Jackson 1969). The apexis selected in a supervised manner. To get Ts at the apex, minimum surface temperature areas arelocated in the scene, then those with highest NDVIs corresponding to forest patches are selected,and the average Ts for that selected region is calculated. Due to the high altitudinal gradients atthe study area, it is necessary to apply a correction to air temperature considering as a referencealtitude that of the pixels at the region forming the apex. Afterwards, positive corrections for altitudeare made for pixels below the base altitude and vice-versa for pixels above considering a lapserate of 6.5 ◦C each 1000 m. This is better than approaches considering a unique air temperature forthe whole area assuming constant meteorological conditions at the blending height (Carlson et al.1995; Czajkowski et al. 2000) and also worked better than the Bastiaanssen et al. (1998) dry andwet pixel approach, probably more suited for flat areas, evaluated in preliminary tests (results notshown).
3.2 Net radiation (Rn)
Rn is calculated as the balance between incoming (↓) and outgoing fluxes (↑) of shortwave (Rs)and longwave (Lw) radiation. By agreement, incoming fluxes are positive and outgoing negative.This can be expressed as the sum of shortwave (Rns) and longwave net radiation (Lnw)
Rn = Rs ↑ + Rs ↓ + Lw ↓ + Lw ↑= Rns + Lnw (6)
3.2.1 Shortwave Net radiationThe shortwave net radiation using remote sensing is calculated as:
Rns = Rs ↓ (1 − α) (7)
where α is the broadband surface albedo estimated according to Liang (2000) for 6 Aster andMODIS bands. Rs↓ (incoming solar radiation) was estimated using a solar radiation model fromFu & Rich (2002).
3.2.2 Longwave net radiationThe longwave energy components are related to surface and atmospheric temperatures through theStephan- Boltzmann law. The longwave net radiation is calculated as in (8):
Lnw = −εsσTs∧4 + Lw ↓ (8)
Where broadband emissivity for the surface, εs, was estimated based on a logarithmic relationshipwith NDVI (van de Griend & Owe, 1993) and radiometric surface temperature, Ts, was directlyobtained from Aster and MODIS LST (Land Surface Temperature) products calculated with the
TES (Temperature Emissivity Separation) algorithm for Aster and the day/night LST algorithm forMODIS. An empirical function is used for the incoming longwave radiation Lw↓ (Idso & Jackson,1969).
4 ESTIMATION OF THE NON-EVAPORATIVE FRACTION FROM S-SEBI
The non-evaporative fraction, NEF, was derived from S-SEBI (Simplified Surface Energy BalanceIndex) model (Roerink et al. 2000). S-SEBI estimates directly the evaporative fraction, EF, on apixel basis considering the relationship between surface temperature and albedo. It assumes thatthe atmospheric conditions remain relatively constant across the study region and that enough wetand dry pixels are present in the scene. Taking into account these assumptions, we have:
EF = Tobs − TLE
TH − TLE(9)
Where Tobs is the observed temperature, TLE is the temperature at the lower boundary functionor evaporation controlled domain and TH is the temperature at the upper boundary function orradiation controlled domain. Both boundary functions, the lower and the upper, are calculated byquantile regression (Koenker & Hallock, 2001) from the albedo vs. surface temperature scatter-plot for the study region and using the 5% and the 95% quantiles respectively. When calculatingS-SEBI, Roerink et al. (2000) used Valor and Caselles (1996) emissivity which provides a the-oretical explanation for van de Griend & Owe (1993) NDVI-emissivity relationship. It requiresan a-priori knowledge of vegetation and soil emissivities. However, when this information is notavailable, errors in emissivity estimates are very similar between both methods (Valor & Caselles1996). For this reason and also to get comparable results in in terms of model performance withprevious NEF estimations using H/Rn ratio, the more operational van de Griend & Owe (1993)emissivity has been used.
5 MAPPING THE RISK OF LAND DEGRADATION IN SIERRA GADOR
A prior step when evaluating land condition is the choice of a reference state corresponding tooptimum or non-degraded status. In this case, we assumed that in Sierra Gádor, there is enoughvariability as to find degraded and non-degraded sites acting as reference levels.
Two variables were employed as inputs for classification: NEF and NDVI. They were previouslyrescaled according to the aridity index to make comparisons across different climatic regions.
The aridity index was calculated in the study area as the ratio between the long-term annualaverage values of potential evapotranspiration and precipitation (Contreras 2006).
The Hargreaves-Samani equation (Hargreaves & Samani 1982) (see equation 9 below), was usedto estimate spatially-distributed potential evapotranspiration (PET). This method is appropriate forsemiarid environments (Vanderlinden et al. 2004, Gavilán et al. 2006) and when meteorologicalinformation is scarce as in Sierra de Gádor:
PET = aRa(Tavg + b)(Tmax − Tmin)0.5 (10)
Ra is the solar radiation in equivalent evaporated water-depth (mm month−1), Tavg themonthly average temperature (◦C), Tmax and Tmin are monthly maximum and minimum averagetemperatures (◦C) respectively, and a y b are regionally-calibrated parameters.
Monthly precipitation and temperature maps were obtained by interpolating data from 35 and16 meteorological stations respectively by using multiple regression with altitude, longitude anddistance to the sea. R2 ranged between 0.6 to 0.87 for precipitation and between 0.6–0.97 fortemperature.
Spatially-distributed monthly radiation values were estimated using POTRAD5 irradiance model(van Dam 2000) developed on PCRaster (van Deursen & Wesseling 1992). Monthly reference
evapotranspiration from Penman-Montheith of 6 meteorological stations at the study site wereused for calibration. The values obtained for the empirical coefficients a and b were 0.00317 and36.45 respectively (R2 = 0.97) (Contreras 2006).
For the two variables selected for classification: NDVI, and NEF, there is a range of variationfrom optimum or reference state (low NEF or high NDVI) to degraded state (low NDVI and highNEF) for each aridity class.
To find the boundary functions corresponding to reference levels of NEF and NDVI, quantileregression for the 5% and 95% was applied in a similar way as with S-SEBI.
The rescaled NEF is:
NEFresc = NEFobs − NEF5%
NEF95% − NEF5%(11)
where:
NEFobs = observed value of NEF in the pixel.NEF5% = value of the NEF lower boundary function for the aridity level corresponding to thatpixel.NEF95% = value of NEF upper boundary function for the aridity level corresponding to thatpixel.
Classification thresholds for NEFresc and NDVIresc were established based on the histograms(mode ± standard deviation).
Classification performance was evaluated using ground truth sites from field visits and photoin-terpretation with areal photographs. Non-degraded validation sites include oak relicts with threedensity levels and a dense reforested pine area. Ground truth areas with risk of land degradationincluded disturbed sites: a strong burnt scar from 2002, an active limestone quarry, an intensivelymining area, and almond orchards plowed for weeds. Therefore, disturbed sites can be sites withrisk of degradation that are related, in this case study, with decreases in vegetation cover andincrease in albedo.
6 RESULTS
6.1 Comparison of air temperature with field data
Eleven meteorological stations were used to evaluate air temperature (Tair) estimations at the timeof the Terra satellite overpass (11.00 UTC). First, the temperature of the apex according to (8) wasselected (Fig. 2). Afterwards, corrections for altitude improved the overall error to less than 2◦C(Table 1).
The overall adjustment is good, but Tair estimates are subjected to local errors. One concernis that altitude is not the only factor affecting Tair . Nevertheless, using this approach presentsthe advantage of relief from using ancillary data. Also any systematic error in Ts estimationwill propagate in the Tair; therefore, these errors should cancel out when calculating differencesTs − Tair for estimating sensible heat flux. Difference in reference altitude between Aster andMODIS is due to the fact that MODIS covers an additional area of Sierra Nevada of high altitudebut not included in the Aster scene. In any case, Tair results between both sensors are similar interms of MAE. In both cases, the region of interest contributing to calculate apex temperature weredense pine forests located in the Sierra Nevada mountains.
6.2 Model performance at the Llano de los Juanes field site
At Llano de los Juanes field site, Rn and surface energy fluxes were measured and compared withAster estimates (Table 2). The B values modeled at Llano de los Juanes are reasonable and similarto the fixed value of 0.25 proposed by Seguin & Itier (1983) for dryland and irrigated crops in
Figure 2. NEF calculated by three methods for different surface types. Surface types have been grouped inthree categories: sites with risk of degradation or disturbed sites. Non-degraded sites with dense vegetationand others, including miscellaneous covers. NEF Carlson corresponds to the ratio H/Rn calculated accordingto Carlson et al. (1995) approach. NEF S-SEBI corresponds to 1-EF calculated by Roerink et al. (2000) andNEF Seguin corresponds to the ratio H/Rn using a fixed value of 0.25 for the exchange coefficient for sensibleheat flux, provided by Seguin & Itier (1983) for unstable conditions.
Table 1. Air temperature estimates at the study site. ‘‘MAE’’ orMean Absolute Error is the absolute average difference of the resid-uals between estimated and measured air temperature at the 11 stations.MAE after adjustment corresponds to the temperature after consideringa correction lapse rate of 6.5 ◦C every 1000 m. Reference altitude is thealtitude of the pixels for which the apex temperature has been extracted.
ASTER MODIS
R2 (Tairstations and altitude) 0.61MAE before adjustment (◦C) 4.31 2.45MAE after adjustment (◦C) 1.96 1.91T apex (◦C) 24.0 28.5Reference altitude (m) 1800 800
Table 2. Comparisons between measured and estimated values at the Gádor Mountain range field site. B isthe exchange coefficient for heat transfer, Rn is net radiation.
Measured Aster Error diff Error %
Rn (W m−2) 179.72 182.44 2.72 1.5B (W m−2◦C−1) 0.37* 0.250 −0.06 −24.0
*Estimated empirically from field data at Llano de los Juanes field site.
France and much lower than the fixed value of 0.64 of Jackson et al. (1977) in irrigated wheat fieldsin Arizona. Errors are very low for Rn (around 1.5% compared to field data) and therefore withinreported net radiometer errors (±10% and directional error of<25 W m−2). Measured B is higherthan estimated B for this day which can be due to the influence of other factors not considered inthe model such as wind speed.
In Llano de los Juanes, Aster results underestimate H when compared with eddy covariancemeasurements. Although reported errors for Ts are within acceptable quality levels, it would benecessary to evaluate its influence in H and Rn estimates in future studies. Error in H propagates
into the NEF = H/Rn. NEF values calculated from Seguin and Carlson methods provide similarresults than the measured value while NEF value from S-SEBI provides a higher error (Table 3).We also have to be aware that the eddy covariance technique is subjected to uncertainty levels of20%–30% (Baldocchi et al. 2001). Moreover, in semi-arid areas with sparse vegetation cover, errorin energy fluxes tend to be on the higher side of this range around 30% (Were 2005).
In general, the range of errors reported by other authors in H flux is very variable. Seguin et al.(1999) consider around 50 W m−2 an acceptable error for H. In the literature, errors in the best casesare around 22 W m−2 (Kustas & Norman 1996) and can reach up to 50% even with sophisticatedmodels when the parameterizations are not good. Using a fixed value for kB−1 in agricultural areasproduces errors as high as 150 W m−2 or 43–64 W m−2 (Seguin et al. 1999). Humes et al. (2000)obtained a RMSE value of 43.35 W m−2 for sensible heat while Laymon & Quattrochi (2000) inthe Great Basin desert got errors in H around 40% similarly to the worse results in this study.
6.3 Analysis of the NEF (non-evaporative fraction) spatial patterns
The assessment of model performance for mapping land degradation risk cannot be made solelyon the basis of an eddy covariance value. To evaluate the spatial coherence of results, relative NEFvalues among different surface types were compared for each method (Fig. 2). Several sites insideand in the surroundings of Sierra Gador were selected by photo-interpretation and field knowledge.
In general, NEF values from Seguin and S-SEBI methods give similar results in terms of trendsand also absolute values. Carlson method overestimates NEF for vegetation corresponding to non-degraded sites and underestimates NEF for degraded sites (e.g. oaks and pines vs quarry or miningzone). NEF from Seguin method provides the most consistent results according to surface types.The three methods yield similar results at Llano de los Juanes field site. Sites where we cannot besure about results are greenhouses and the Andarax ephemeral river. S-SEBI does not model wellfree water.
In order to spot sites with high risk of degradation, either NEFSeguin or NEFS-SEBI should work,with the exception of greenhouse areas. Rescaling of NEF values according to climate variationsshould be performed, so that more arid areas are not classified as degraded sites due to higher NEFor lower NDVI values.
The relationship between NDVI, as a proxy for vegetation cover, and NEF (Figure 3) showsthat for the selected surface types (disturbed and non-degraded sites in Sierra Gádor), high NDVIvalues are associated with low NEFs. In the case of NEFSeguin and NEFS-SEBI the NDVI-NEF trendis linear and low NDVI values are associated with high NEF, which is logic, considering also thatwe are in summer.
It seems that there is an albedo feedback effect which decreases net radiation at low vegetationcovers (Bastiaanssen et al. 1998, Roerink et al. 2000) apparent by a negative relationship between
Figure 3. Non-evaporative fraction (NEF) versus NDVI calculated according to Carlson, Seguin, and S-SEBImethods in Sierra Gador. Grey dots include all pixels within Sierra Gador. Black markers represent validationsites only within Sierra Gador with risk of degradation (burnt, limestone quarry, almond, mines from Fig. 2)and non-degraded sites (pines gador, dense oaks, oaks, sparse oaks). Seguin and S-SEBI best fit with thoseselected sites within Gador corresponds to linear regressions while Carlson fits a second order polynomialdegree function.
Table 3. Comparisons between measured (with eddy covariance technique) and estimated (with three differentmethods) values of sensible heat, H, and non-evaporative fraction, NEF, in Llano de los Juanes. Seguin methodassumes a fixed value for the exchange coefficient for heat transfer, B, while Carlson method uses a B valuerelated with NDVI. S-SEBI is the simplified SEBI method developed by Roerink et al. (2000). N/A (notapplicable to this method).
Variable Eddy Seguin Carlson S-SEBI
H (Wm−2) Mean 149.59 106.70 107.58 N/AError diff 0 −42.89 −42.01 N/AError % 0 −28.67 −28.08 N/A
NEF and NDVI values lower than 0.25, thus corresponding to bare soil, for the upper NEF valuesof the cloud in Figure 3.
It has been described in a semi-arid region in Spain (Castilla la Mancha) that pixels at the wetedge (evapotranspiring at potential rates) present a positive relationship between Ts or Ts–Tair withalbedo and negative with NDVI with no inflection points. However, pixels at the dry edge (with noavailable water for evapotranspiration) present a positive relationship between Ts or Ts–Tair withalbedo up to an inflection point. Similarly occurs with the relationship between NDVI and Ts orTs–Tair . The albedo for which a feedback effect on temperature occurs corresponds to the NDVIinflection point as well (Garcia et al. in press). For this reason, we can assume that in this casealbedo is decreasing surface temperature and NEF values at low NDVI levels.
In any case, this effect on surface temperature should not be as strong as to decrease NEF to thelevel of oaks or orchards as it is the case in NEFCarlson. In this case, for bare soil (NDVI < 0.2),there is a strong decrease in NEF that does not seem realistic. It was shown previously in Figure 2that this model underestimates surface roughness and sensible heat for bare soil. If this model isused as input for a land degradation classification, sites where a vegetation cover loss has occurredwill be confused with high vegetated areas.
When considering all the pixels at the study site, a significant scattering between NDVI and NEFappears (Fig. 3). This reveals that although NDVI and NEF are correlated to some extent, theyprovide different information. There are many factors involved in surface resistance to sensible heattransfer (Brutsaert 1982), and Carlson et al. (1995) simplification in this semi-arid conditions doesnot work. The error commited by estimating surface resistance as a function of NDVI is worse thannot estimating it at all. There is uncertainty in selection of NDVI maximum and minimum values.The errors are amplified for bare soil conditions. Not considering surface resistance variationsacross the landscape, will be similar to consider a ‘‘potential’’ or ‘‘reference’’ sensible heat flux,approach that might still be valid for mapping land degradation sites. Future efforts will be devotedto a precise estimation of surface resistance by considering vegetation height, wind speed, and othervariables, which will allow to evaluate the impact of not considering it on the land degradationindicator.
Considering this and previous results, the model finally selected for use at a regional scale isbased on calculating NEF values as H/Rn and according to Seguin et al. (1989).
6.4 Comparison between MODIS and ASTER performance
Comparisons between Aster and MODIS were performed by resampling the Aster pixel size of90 m using a cubic convolution filter into MODIS 1 km pixel size. Figure 4 shows results for outputvariables using Aster and MODIS. The main structure of the surface fluxes explained by the NEFis maintained from MODIS to Aster scene, but level of detail is not comparable. For this reason,features with a spatial resolution lower than 1 km are not resolved (e.g. the mosaic of irrigatedcitrus and bare soil by the Andarax ephemeral river). Therefore, if the processes of land degradationrelated with changes in the surface energy balance are explicit at 1 km grains, which needs further
Figure 4. Comparison of spatial patterns nd NEF (H/Rn Seguin) for ASTER (90 m) and MODIS (1 km) overthe study site on 18-07-2004. Sierra Gador is outlined in black.
elucidation, MODIS can be consider an adequate tool to perform regional assessments by meansof its high temporal and spatial coverage.
Results in Table 4 show that for the set of input variables, MODIS values tend to be slightlylower than Aster with the exception of albedo. The RMSE value in Ts is 2.4◦C. These differencesare translated to the output variables causing underestimations of MODIS with respect to Aster.All the variables present R2 greater than 0.76, the lowest corresponds to NDVI, which means thatthe variables aggregate linearly but there is a remaining ∼20% of the variance responding to non-linear aggregation effects combined with sensor differences in sensor performance and correctionalgorithms employed.
Table 4. Input and output variables for MODIS and Aster: Ts (Land Surface Temperature), H Seguin (sensibleheat flux calculated according to Seguin), Rn (net radiation), NEF Seguin (non-evaporative fraction calculatedaccording to Seguin). RMSE ( Root Mean Square Error) between both variables, R2 (Pearson coefficientafter aggregation of Aster to 1 km pixel), Slope (slope of regression between Aster -independent and Modis-dependent).
OUTPUT Rn (W m−2) 9.08 0.80 0.86H Seguin (W m−2) 18.24 0.78 0.71NEF Seguin 0.096 0.83 0.74
Figure 5. Quantile regression for the upper (95%) and lower (5%) quantiles between (a) aridity index andNEF and (b) aridity index and NDVI in Sierra Gador.
In any case, both RMSE values and mean differences are within reported net radiometer precisionfor Rn. RMSE results for H are comparable to some closure errors in energy balance using eddycovariance technique (Baldocchi et al. 2001). This means that despite some differences betweenAster and MODIS sensor performance, combined with the impact of change in spatial resolution, thetwo sensors provide similar results, for spatial features of 1 km or larger, comparable to repeatabilityof instrumental data.
6.5 Case study: mapping the risk of land degradation in Sierra Gador
The selected NEF model (H/Rn Seguin) and the NDVI were rescaled by aridity levels, usinga polynomic function for the lower boundary (5% quantile) and a linear function for the upperboundary (95% quantile) (Fig. 5).
The classification method was based on histogram thresholding corresponding to the mode ±standard deviation (Fig. 6). In this way, values at the extremes corresponding to near to optimumstatus and very far from optimum are selected.
Three classification approaches were compared (a) using NEF and NDVI rescaled as inputvariables, (b) using only NEF rescaled and (c) only NDVI rescaled. Output maps were comparedwith ground truth sites (Table 5). They show that classes associated with degraded sites correspondto (a) ↓NDVI-↑NEF, (b) ↑NEF alone, and (c) ↓NDVI alone. The mapping was better, in termsof correct identification of pixels using only (b) ↑NEF for degraded sites and (c) ↑NDVI fornon-degraded sites (Table 5). Missclassification was negligible in all cases. The three aproachesidentified all the hot spots selected as validation sites but some of the methods did not coverthe complete ground truth sites resulting in non-classification. It was prefered a conservativeclassification than an over-classification, to decrease the number of false alarms.
Figure 6. Histograms for NEF and NDVI after being rescaled according to the aridity index in SierraGador. The thresholds used for classification were: a) for NDVI rescaled, 0.35 ± 0.31; and b) for NEFrescaled, 0.60 ± 0.25 corresponding to the mode ± standard deviation.
Table 5. Confusion matrix for three classificaction methods: using NEF rescaled and NDVI rescaled byaridity levels, using only NEF rescaled and only NDVI rescaled. Non-degraded sites, n = 172; degraded site,n = 317.
Ground Truth (%)
Method Class Non degraded High risk of degradation Total
Figure 7 shows the final map of classification of areas with high risk of degradation in SierraGador mapped using ↑NEF, and non-degraded areas mapped using ↑NDVI.
Currently, only heavily disturbed sites are identified. Ongoing research is focused on the assess-ment intermediate classes of land degradation (e.g. shrublands replacing oak forests) and also toinclude more tests sites expanding the region of analysis. This issue is particularly complicated as
Figure 7. Classification map of areas with high risk of degradation (red) and low risk of degradation (green)in Sierra Gador (Almeria, Spain). Validation sites for non-degraded sites are composed of oak relicts withthree density levels and pine forests (blue polygons), and disturbed sites considered at high risk of degradationare mapped as black polygons (burnt scar from 2002, limestone quarry, almond orchards, and abandondedmines) (see colour plate page 396).
intermediate levels of degradation can present a reversible trend that a single or few images cannot capture and are even hard to identify and define in field validations.
7 CONCLUSIONS
A reliable identification of disturbed sites presenting high risk of degradation and non-degradedsites is accomplished using the non-evaporative fraction (NEF) and the NDVI. Non-degraded sitesare better identified using NDVI rescaled only, while degraded sites are better classified usingonly NEF. Both indicators, related with loss of vegetation cover and decrease of evapotranspirationlevels, are correlated to some extent. However, they provide different information as shown bythe significant scattering between them and by the fact that classification using both variablesproduces poorer results than classification using each variable separately. Therefore, despite thefact that vegetation cover is a clear symptom of land degradation, other indicators not directlyrelated with vegetation cover, based on the surface energy balance, such as the NEF, can revealimportant information about ecosystem functioning.
The spatial patterns of the non-evaporative fraction are coherent with the type of surface usingthe S-SEBI model and especially using the H/Rn ratio when a constant value of the exchangecoefficient to sensible heat based on Seguin & Itier (1989) is considered. Comparison with fielddata using eddy covariance technique at Llano de los Juanes experimental field site shows verylow net radiation errors (within net radiometer precision). Sensible heat provides larger errors butwithin the ranges obtained by other authors, within the threshold of 50 Wm−2 proposed by Seguinet al. (1999) for the three methods. Air temperature can be extracted from the images using theNDVI-Ts space relationship (Carlson et al. 1995) corrected by altitude with an acceptable overallerror (<2◦C).
Comparison between MODIS and Aster spatial patterns of surface energy balance componentsin summer reveals that the main spatial structure is maintained from Aster to MODIS. However,
the loss of detail is remarkable and features smaller than 1 km are not resolved. Regarding RMSEvalues between Aster and MODIS, net radiation values are within reported net radiometer pre-cision and sensible heat flux are comparable to closure errors in energy balance using eddycovariance technique. Therefore, if the processes of land degradation related with changes inthe surface energy balance are explicit at grains of 1 km, which needs further elucidation, MODISis an adequate tool to perform regional assessments by means of its high temporal and spatialcoverage.
Establishing a robust and simple land degradation index needs additional analyses includingseveral dates to fully understand the interaction between water and energy at different surfaces.Efforts should be devoted to scale daily results to temporal scales compatible to the time scale ofland degradation processes. Along this line, a remaining question is the spatial and temporal scalesin which alterations of energy balance fluxes related with land degradation are explicit.
ACKNOWLEDGMENTS
This research has been developed within the EU project DeSurvey: ‘‘A Surveillance System forAssessing and Monitoring of Desertification’’ (FP6-00.950). The authors wish to thank Dr. delBarrio, M. San Juan and Dr. Lázaro for help during this work. Computer support from R. Ordialesand S. Vidal is greatly appreciated. An anonymous reviewer is acknowledged for very usefulcomments that contributed to improve this paper.
REFERENCES
Brutsaert, W. 1982. Evaporation into the atmosphere. Theory, history, and applications. Dordrecht: Holland,D.Reidel Publishing Company.
Baldocchi, D.D., Falge, E., Gu, L., Olson, R., Hollinger, D., Running, S., Anthoni, P., Bernhofer, Ch., Davis,K., Fuentes, J., Goldstein, A., Katul, G., Law, B., Lee, X., Malhi, Y., Meyers, T., Munger, J.W., Oechel, W.,Pilegaard, K., Schmid, H.P., Valentini, R., Verma, S., Vesala, T., Wilson K. & Wofsy, S. 2001. FLUXNET:A new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor andenergy flux densities. Bulletin of the American Meteorology Society 82: 2415–2434.
Bastiaanssen, W.G.M., Menenti, M., Feddes, R.A. & Holtslag, A.A.M. 1998. A remote sensing surface energybalance algorithm for land (SEBAL) 1. Formulation. Journal of Hydrology 212–213(1–4): 198–212.
Boer, M.M. & Puigdefábregas, J. 2003. Predicting potential vegetation index values as a reference for theassessment and monitoring of dryland condition. International Journal of Remote Sensing 24: 1135–1141.
Boer, M.M. & Puigdefábregas, J. 2005. Assessment of dryland condition using spatial anomalies of vegetationindex values. International Journal of Remote Sensing 26: 4045–4065.
Carlson, T.N., Capehart, W.J. & Gillies, R.R. 1995. A new look at the simplified method for remote sensingof daily evapotranspiration. Remote Sensing of Environment 54: 161–167.
Chehbouni, A., Loseen, D., Njoku, E.G., Lhomme, J.P., Monteny, B. & Kerr,Y.H. 1997. Estimation of sensibleheat flux over sparsely vegetated surfaces. Journal of Hydrology 89(1–4): 855–868.
Contreras, S. 2006. Spatial distribuition of the annual water balance in semi-arid mountainous regions: appli-cation to Sierra Gádor (Almería, SE Spain) Ph.D. Thesis (in Spanish). Almería: Servicio PublicacionesUniversidad de Almería.
Czajkowski, K.P., Goward, S.N., Mulhern, T., Goetz, S.J., Walz, A., Shirey, D., Stadler, S., Prince, S.D. &Dubayah, R.O. 2000. Estimating environmental variables using thermal remote sensing.In Quattrochi, D.A. & Luvall, J.C. (eds), Thermal remote sensing in land surface processes: 11–32. BocaRaton, Florida: CRC Press.
Dolman, A.J., Gash, J.H.C., Goutorbe, J.P., Kerr, Y., Lebel, T., Prince, S.D. & Stricker, J.N.M. 1997. Therole of the land surface in Sahelian climate: HAPEX-Sahel results and future research needs. Journal ofHydrology 89(1–4): 1067–1079.
Domingo, F., Villagarcía, L., Boer, M., Alados-Arboledas, L. & Puigdefábregas, J. 2001. Evaluating the long-term water balance of arid zone stream bed vegetation using evapotranspiration modelling and hillsloperunoff measurements. Journal of Hydrology 243: 17–30.
French, A.N., Jacob, F., Anderson, M.C., Kustas, W.P., Timmermans, W., Gieske, A., Su, B., Su, H., Mccabe,M.F., Li, F., Prueger, J.H. & Brusnell, N. 2005. Surface energy fluxes with the advanced spaceborne thermalemission and reflection radiometer (ASTER) at the Iowa 2002 SMACEX site (USA). Remote Sensing ofEnvironment 99(1–2): 55–65.
Fu, P. & Rich, M. 2002. A geometric solar radiation model with applications in agriculture and forestry.Computers and Electronics in Agriculture 37: 25–35.
García, M., Palacios-Orueta, A., Puigdefábregas, J., Contreras, S., Del Barrio, G., Fernandez, F.J. & Moreno,M. in press. Estimating evapotranspiration from TVDI: Towars a land degradation index for regiona analysis.In Proceeding of the 2nd Symposium on Quantitative Advances in Remote Sensing: 664–669, Valencia(Spain) Sept. 24–29.
Gavilán, P., Lorite, I.J., Tornero, S. & Berengena, J. 2006. Regional calibration of Hargreaves equation forestimating reference ET in a semiarid environment. Agricultural Water Management 81(3): 257–281
Humes, K., Hardy, R., Kustas, W., Prueger, J. & Starks, P. 2000. High spatial resolution mapping of surfaceenergy balance components with remotely sensed data. In Quattrochi, D.A. & Luvall, J.C. (eds), ThermalRemote Sensing in Land Surface Processes: 110–132. Boca Raton, Florida: CRC Press.
Idso, S.B. & Jackson, R.D. 1969. Thermal radiation from the atmosphere. Journal of Geophysical Research74: 5397–5403.
Jackson, R.D., Reginato, R.J. & Idso, S.B. 1977. Wheat canopy temperature: a practical tool for evaluatingwater requirements. Water Resources Research 13(3): 651–656.
Jacob, F., Petitcolin, F., Schmugge, T., Vermote, E., French, A. & Ogawa, K. 2004. Comparison of landsurface emissivity and radiometric temperature derived from MODIS and ASTER sensors. Remote Sensingof Environment 90(2): 137–152.
Koenker, R. & Hallock, F.F. 2001. Quantile regression. Journal of Economic Perspectives 15(4): 143–156.Kustas, W.P. & Norman, J.M. 1996. Use of remote sensing for evapotranspiration monitoring over land surfaces.
Hydrological Sciences Journal 41(4): 495–516.Laymon, C.A. & Quattrochi, D.A. 2000. Estimating spatially-distributed surface fluxes in a semi-arid great
basin desert using Landsat TM thermal data. In Quattrochi, D.A. & Luvall, J.C. (eds), Thermal RemoteSensing in Land Surface Processes: 110–132. Boca Raton, Florida: CRC Press.
Li, X.-Y., Contreras, S. & Solé-Benet, A. Spatial distribution of rock fragments in dolines: a case study in asemiarid Mediterranean mountain-range (Sierra Gador, SE Spain). CATENA, in press.
Liang, S. 2000. Narrowband to broadband conversions of land surface albedo in algorithms. Remote Sensingof Environment 76: 213–238.
López-Bermúdez, F., Boix-Fayos, C., Solé-Benet, A., Albaladejo, J., Barberá, G.C., del Barrio, G.,Castillo, V., García, J., Lázaro, R., Martínez-Mena, M.D., Mosch, W., Navarro-Cano, J.A., Puigde-fábregas, J. & Sanjuán, M. 2005. Landscapes and desertification in South-East Spain: overview andfield sites. Field Trip Guide A-5. 6th International conference on geomorphology. Sociedad Españolade Geomorfología. 40p.
Ludwig, J.A. & Tongway, D.J. 1997. A landscape approach to rangeland ecology. In Ludwig, J.A., Tong-way, D.J., Freudenberger, D.O., Noble, J.C. & Hodgkinson, K.C. (eds), Landscape ecology, function andmanagement: principles from Australia’s rangelands: 1–12. Melbourne: CSIRO Publishing.
Ludwig, J.A. & Tongway, D.J. 2000. Viewing rangelands as landscape systems. In Arnalds, O. & Archer, S.(eds), Rangeland desertification: 39-52. Dordrecht, Netherlands: Kluwer Academic Publishers.
Phillips, J.D. 1993. Biophysical feedbacks and the risks of desertification. Annals of the Association ofAmerican Geographers 83(4): 630–640.
Puigdefábregas, J. & Mendizábal, T. 2004. Prospects for desertification impacts in Southern Europe.In A. Marquina (ed), Environmental challenges in the Mediterranean 2000–2050: 155–172. Boston, KluwerAcademic Publisher.
Roerink, G.J., Su, Z. & Menenti, M. 2000. S-SEBI: a simple remote sensing algorithm to estimate the surfaceenergy balance. Physics and Chemistry of the Earth (Part B) 25: 147–157.
Safriel, U., Adeel, Z., Niemeijer, D., Puigdefábregas, J., White, R., Lal, R., Winslow, M., Ziedler, J., Pronce, S.,Archer, E. & King, C. 2003. Dryland Systems. In ‘‘Ecosystems & human well-being. Current state andtrends’’, Millenium Ecosystem Assessment Series, Vol 1, Cap 22, pp. 623–662, World Resources Institute,Washington, DC.
Seguin, B. & Itier, B. 1983. Using midday surface temperature to estimate daily evaporation from satellitethermal IR data. International Journal of Remote Sensing 4: 37–383.
Seguin, B., Becker, F., Phulpin, T., Gu, X.F., Guyot, G., Kerr, Y., King, C., Lagouarde, J.P., Ottle, C.,Stoll, M.P., Tabbagh, A. & Vidal, A. 1999. IRSUTE: a minisatellite project for land surface heat fluxestimation form field to regional scale. Remote Sensing of Environment 68: 357–369.
Sharma, K.D. 1998. The hydrological indicators of desertification. Journal of Arid Environments 39: 121–132.Valor, E. & Caselles, V. 1996. Mapping Land Surface Emissivity from NDVI: Application to European,
African, and South American Areas. Remote Sensing of Environment 57(3): 167–184.Van Dam, O. 2000. Modelling incoming potential radiation on a land surface with PCRaster: POTRAD5.MOD
manual. Utrech Centre for Environment and Landscape Dynamics, Utrecht University.Van Deursen, W.P.A. & Wesseling, C.G. 1992. The PC-Raster Package. Dept. of Physical Geography, Utrecht
University. PCRaster webpage: http://pcraster.geo.uu.nl/index.html.Van de Griend, A.A. & Owe, M. 1993. On the relationship between thermal emissivity and the normalized
difference vegetation index for natural surfaces. International Journal of Remote Sensing 14(6): 1119–1131.Vanderlinden, K., Giráldez, J.V. & van Meirvenne, M. 2004 Assesing reference evapotranspiration by the
Hargreaves method in southern Spain. Journal of Irrigation and Drainage Engineering 130(3): 184–191.Wassenaar, T., Olioso, A., Hasager, C., Jacob, F. & Chehbouni, A. 2002. Estimation of evapotranspiration
on heterogeneous pixels. In Sobrino, J.A. (ed), Recent advances in quantitative remote sensing, Proc. 1stintern. symp.: 319–328. Valencia, September 2002.
Were, A. 2005. Spatial aggregation of evapotranspiration in semi-arid climate. Ph.D. Thesis (in Spanish).Universidad de Almeria, Spain.
Wang, Q.X. & Takahashi, H. 1998. Regional hydrological effects of grassland degradation in the Loess Plateauof China. Hydrological Processes 12: 2279–2288.
