Landsat-based approaches for mapping of land degradation prevalence and soil functional properties in Ethiopia

This article appeared in a journal published by Elsevier. The attachedcopy is furnished to the author for internal non-commercial researchand education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling orlicensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of thearticle (e.g. in Word or Tex form) to their personal website orinstitutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies areencouraged to visit:

http://www.elsevier.com/authorsrights

Author's personal copy

Landsat-based approaches for mapping of land degradation prevalence and soilfunctional properties in Ethiopia

Tor-G. Vågen a,⁎, Leigh A. Winowiecki b, Assefa Abegaz c, Kiros M. Hadgu d

a World Agroforestry Centre (ICRAF), P.O. Box 30677, 00100 GPO, Nairobi, Kenyab International Center for Tropical Agriculture (CIAT), P.O. Box 823-00621, Nairobi, Kenyac Department of Geography and Environmental Studies, Addis Ababa University, P.O. Box 1176, Addis Ababa, Ethiopiad World Agroforestry Centre (ICRAF), P.O.Box 2658, Addis Ababa, Ethiopia

a b s t r a c ta r t i c l e i n f o

Article history:Received 14 November 2012Received in revised form 7 March 2013Accepted 7 March 2013Available online xxxx

Keywords:Soil organic carbonpHSoil erosionLand degradationAgricultureMappingLandsatEthiopia

Agriculture is the basis of the Ethiopian economy, accounting for the majority of its employment and exportearnings. Land degradation is, however, widespread and improved targeting of land management interven-tions is needed, taking into account the variability of soil properties that affect agricultural productivityand land degradation risk across landscapes. In the current study we demonstrate the utility of LandsatETM + imagery for landscape-level assessments of land degradation risk and soil condition through a com-bination of systematic field methodologies, infrared (IR) spectroscopy and ensemble modeling techniques.The approaches presented allow for the development of maps at spatial scales that are appropriate for mak-ing spatially explicit management recommendations. Field data and soil samples collected from 38 sites, each100 km2, were used to develop predictive models that were applied as part of a case study to an independentdataset from four sites in Ethiopia. The predictions based on Landsat reflectance were robust, with R-squaredvalues of 0.86 for pH and 0.79 for soil organic carbon (SOC), and were used to create predicted surfaces(maps) for these soil properties. Further, models were developed for the mapping of the occurrence of soilerosion and root depth restrictions within 50 cm of the soil surface (RDR50), with an accuracy of about80% for both variables. The maps generated from these models were used to assess the spatial distributionof soil pH and SOC, which are important indicators of soil condition, and land degradation risk factors inorder to target relevant management options.

© 2013 Elsevier Inc. All rights reserved.

1. Introduction

Since agriculture remains the basis of the Ethiopian economyup-to-date information on soil functional properties is needed for un-derstanding soil constraints in order to improve and maintain landproductivity. In its country strategy paper (2011–2015) for Ethiopia,the African Development Bank (ADB) reports that the agriculturalsector accounts for 42% of GDP, 80% of employment and 85% of its ex-port earnings, with a mainly smallholder dominated structure. Inorder to achieve the formulated goals for agricultural developmentof Ethiopia, establishment of sustainable agricultural strategies is apre-requisite, for which development of appropriate soil resource in-formation and management systems is one of the tools.

Despite their economic importance, agricultural and rangeland sys-tems in Ethiopia are subject to continuous and widespread disequilibri-um dynamics (Coppock, 1993) through severe land degradation due toits topography, population growth and adverse agricultural and range-land land-use practices over centuries (Berry, 2003; Bishaw, 2001;Desta & Coppock, 2004; Hurni, 1988, 1993; Taddese, 2001). Ethiopia's

geological history and diverse parent materials, coupled with its com-plex topography, are reflected in its diversity of soil types. However, na-tional soil mapswere developed between 30 and 40 years ago at a scaleof 1:2 M (FAO, 1998a) and are hence both outdated and have a coarsespatial resolution. Traditional soil maps, generally based on classifica-tion of soils into taxonomic units, are potentially useful for understand-ing the occurrence of soil types across landscapes and their propertiesmay determine the constraint envelopes for dynamic soil properties(Norfleet et al., 2003). However, such maps are static (use-invariant)and do not capture soil degradation prevalence or risk, nor do they re-flect the dynamic nature of soil functional properties. There is thereforea need for soil information that reflects use-dependent or dynamic soilproperties at landscape scale and land use/cover characteristics that aresensitive tomanagement. This information can be used to better informland management decision-making, as well as for modeling of soil loss,hydrology or crop suitability. Given the increasing pressure on the soilresource base, it has become critical to develop maps of such dynamicsoil properties at scales appropriate for targeted soil management,also in smallholder systems.

While many studies have assessed land degradation, and specifi-cally soil erosion, in Ethiopia (Daba, 2003; Descheemaeker et al.,2006; Grepperud, 1996; Hurni, 1993; Nyssen et al., 2008; Sonneveld

Remote Sensing of Environment 134 (2013) 266–275

⁎ Corresponding author. Tel.: +254 724332508; fax: +254 20 7224001.E-mail address: [email protected] (T.-G. Vågen).

0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved.http://dx.doi.org/10.1016/j.rse.2013.03.006

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r .com/ locate / rse

Author's personal copy

& Keyzer, 2003) spatially explicit approaches at fine-resolution scalesthat incorporate diverse climatic conditions, dynamic soil propertiesand topography are still needed to better assess land health andtarget soil conservation activities. Many of the assessments of soilerosion in Ethiopia apply the Universal Soil Loss Equation (USLE)(Wischmeier & Smith, 1962) or some form of revision of the USLE,such as the MUSLE (Smith et al., 1984) or the RUSLE (Renard &Ferreira, 1993). While the RUSLE incorporates mean annual rainfall,soil data, land cover, slope length and steepness, as well as land man-agement information to estimate soil loss, we present an alternativemethod that uses systematically collected field data to train modelsfor the prediction of soil erosion prevalence at 30 m resolution,using Landsat reflectance data as predictors.

We use the following definition of land health: “The capacity ofland to sustain delivery of essential ecosystem services” (Vågen etal., 2012), and propose an approach that uses dynamic soil properties,instead of soil classes, and indicators of soil degradation to model andmap soil and land health. For example, mapping the distribution ofsoil erosion prevalence across landscapes is important to better targetland management practices that reduce degradation risk or rehabili-tate already degraded areas, and can be coupled with additional infor-mation about landscape attributes and socioeconomic factors thatmay contribute to increasing degradation prevalence (Daba, 2003;Dragan et al., 2003; Piot et al., 1995). Sensitivity to soil erosion, orsoil erosion risk, is a commonly used indicator in agricultural man-agement. Similarly, root depth restrictions may strongly inhibitplant growth, particularly in cropping systems, and also influencethe hydrological characteristics of the soil, including infiltration andwater holding capacities. Land management practices that exposesoil, deform soil via traffic and/or cultivation or grazing increase soilerodibility, while in grazing areas, over-stocking may lead to soilcompaction, reducing infiltration capacity and increasing runoff dueto collapse of macro-pores.

Studies of soil and ecosystem health often rely heavily on chemicalanalysis of soil samples (Foley et al., 1998). Soil analysis using tradi-tional wet chemistry is generally costly, and since studies are oftenunable to afford all the required analyses, the result is less thanideal sampling schemes and poor levels of replication. Soil infraredspectroscopy can reduce these constraints by providing quick,non-destructive and quantitative analyses of an enormous range oforganic and other constituents of soil material, including for the dis-crimination between complex mixtures, and is now an establishedtechnology (Jong, 1994; McCarty et al., 2002; Osborne, 1986). Thistechnology has been used for the characterization of soils for morethan two decades (Ben-Dor & Banin, 1995; Chang et al., 2001), andhas been shown to have many applications for approaches to the de-velopment of soil condition indices as well (Vågen et al., 2006). In thelast decade in particular, there has been a strong increase in the use ofIR spectroscopy for soil and plant characterization, and the method isnow in routine use (Odlare et al., 2005; Shepherd & Walsh, 2002;Terhoeven-Urselmans et al., 2010; Zornoza et al., 2006).

An important component for rigorously assessing land health is asystematic sampling design that includes sufficient sampling densitiesand replication across landscapes at different spatial scales. The currentstudywas undertaken to explore the development of predictionmodelsfor moderate to high resolution (≤30 m)mapping of soil condition, fo-cusing on soil organic carbon (SOC) and soil pH. Both of these are con-sidered important indicators of soil condition (Baldock et al., 2009). Inaddition, the prediction and mapping of land degradation risk factorssuch as erosion prevalence and root depth restrictions were explored.Field measurements from 38 sites, which were surveyed using system-atic field sampling procedures referred to as the Land Degradation Sur-veillance Framework (LDSF) (Vågen et al., 2012) as part of severalprojects, were used to develop models for prediction of soil conditionand land health indicators. Earlier studies suggest that models basedon satellite imagery have significant potential in predicting soil

condition and land degradation processes (Hill & Schütt, 2000;Monastersky, 1989; Pinet et al., 2006; Vågen et al., 2006). Also, severalstudies have applied Landsat data for predicting soil taxonomic unitsand drainage classes (Alaily & Pohlmann, 1995; Kilic, 2009; Su et al.,1989), while others have applied image reflectance data in the predic-tion of soil salinity (Melendez-Pastor et al., 2012; Panah & Pouyafar,2005; Shrestha, 2006). In the current study, a combination of ap-proaches using remote sensing and soil IR spectroscopy was appliedto a case study of four study sites in Ethiopia with the aim to exploremethods for development of cost-effective mapping techniques thatcan be applied across larger areas (i.e. landscapes). We propose tobuild on these studies and develop an assessment of land degradationand soil health for diverse environments in Ethiopia at scales appropri-ate for management.

2. Study sites

Four sites were used in this case study from Ethiopia, each randomlylocatedwithin four different LandsatWRS2 path/row tiles, representingthree different Köppen–Geiger climate zones. The climate zones includ-edwere tropicalwet savanna (Dambidolo), humid subtropical highlandclimate (Kutaber) and warm semi-arid savanna (Mega andMerar) (cli-mate classification after Kottek et al., 2006) (Fig. 1). Existing data ondominant soil types indicate that Mega is dominated by Cambisolsand Leptosols; Merar by Cambisols, Leptosols and Vertisols; Kutaberby Leptosols and Dambidolo by Phaeozems (FAO, 1998b).

A cluster sampling design (Thompson, 1991)was used by first divid-ing each site into 16 tiles (2.5 × 2.5 km in size), then generating onerandom centroid location per tile, and finally generating 15 randomsampling plots, each 1000 m2, within a 564 m radius of each clustercentroid. Five of these plots were used as alternate plots and hence 10were characterized and sampled in each cluster. Thus, the data foreach site consisted of 160 stratified-random sampling plots with anarea of 1000 m2 each. Within each individual plot, four sub-plots wereestablished, each with an area of 100 m2, one in the center and threeon a radial arm with 120° angles between them. This form of stratifiedcluster sampling allows the assessment of variability of soil propertiesas different spatial scales (in our case: subplot, plot, cluster, site, be-tween sites) by applyingmodels incorporating random effects that rep-resent different groups, including spatial nested scales (Pfeffermann &Nathan, 1981; Raudenbush & Bryk, 2002). Assessing and understandingspatial variability is essential for developing accurate predictive modelsand understanding landscape dynamics.

3. Materials and methods

The LDSF approach to mapping of soil properties and land degrada-tion prevalence makes use of (i) systematic and unbiased field surveysto collect soil information and other ecological parameters, (ii) laboratoryanalyses using wet chemistry and soil IR spectroscopy, and (iii) remotesensing information. The aim of this study was to explore the potentialof applying analytical frameworks that utilize this information as wellas statistical approaches to develop predictive models and create mapsof various indicators of land health that may be used to developmanage-ment recommendations.

3.1. Field data collection

Field survey data collection in Ethiopia was conducted betweenFebruary and June 2011. The field team was led by the staff fromMekelle and Addis Ababa Universities, with assistance from the Inter-national Center for Tropical Agriculture (CIAT) drivers, all of whomwere trained and supervised by CIAT and the World AgroforestryCentre (ICRAF) scientists. Local extension agents and farmers fromeach Woreda were also involved and assisted in the field surveys.The surveys included soil sampling, vegetation measurements, root

267T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

depth restriction and soil erosion prevalence scores, which wereconducted at plot and subplot levels. Soil samples were collected attwo vertical depths (0–20 and 20–50 cm) at each subplot (n = 4)using a soil auger. Subplot soil samples were subsequently combinedto form one composite sample for 0–20 cm depth and one compositesample for 20–50 cm depth per plot, making a total of 320 soil sam-ples per site. Prevalence and typology of soil erosion (sheet, rill andgully) were recorded at each subplot, while root depth restrictionswere recorded by measuring the depth (in cm) from the soil surfaceto the auger depth in cases where the auger could not penetrate thesoil. In the current study we assessed the presence of root depth re-strictions in the upper 50 cm of the soil profile (RDR50) within eachplot, by scoring these occurrences from 0 (none) to 4 (in caseswhere all subplots had restrictions). Vegetation was also scored atthe subplot level, rating woody and herbaceous cover as: absent;b4; 4–15; 15–40; 40–65 and >65%. Tree and shrub counts wereconducted for density estimates with shrubs classified as woodyplants between 1.5 and 3 m in height, while woody vegetation tallerthan 3 mwas classified as trees. In addition, ecosystem impact factorssuch as agriculture, grazing, fire and tree cutting were scored for eachplot, based on methods adapted from Moat and Smith (2007).

Field data from 38 sites, surveyed in the period 2009 to 2012 usingthe LDSFmethodology (Vågen et al., 2010; Vågen et al., 2012),were col-lected as part of several projects, but using the same consistentmethod-ology. The countries surveyed were Ethiopia (4 sites), Kenya (3 sites),

Tanzania (7 sites), Malawi (2 sites), Mozambique (6 sites), DemocraticRepublic of Congo (2 sites), Ghana (2 sites), Burkina Faso (2 sites),South Africa (1 site), Nigeria (3 sites), Zambia (4 sites) and Zimbabwe(2 sites). These data were used to train and validate predictive modelsfor SOC, pH, soil erosion prevalence and occurrence of RDR50 basedon a total of 3378 topsoil samples for SOC and pH and observationsfrom 6237 plots for soil erosion and RDR50. All field data were collectedin near real time using an electronic data entry systemdeveloped by thefirst author, building on the CyberTracker platform (Liebenberg, 2003)for handheld PDAs, with duplicate entries on paper forms as backup.

3.2. Soil laboratory analysis

All soil samples were air-dried, crushed using a wooden rolling pinand sieved using a two millimeter mesh certified sieve. Samplesscanned on the HTS-XT Fourier transform mid-infrared (FT-MIR)spectrometer (Bruker Optics) were further ground to 0.05 μm beforeanalysis. All soil samples were characterized using MIR spectroscopyfollowing procedures outlined in Terhoeven-Urselmans et al. (2010)at ICRAF in Nairobi, Kenya. Soil pH, base cations and other soil fertilitycharacteristics were analyzed on 10% of the samples using a 1:1 H2Osolution at the Crop Nutrition Laboratory in Nairobi, Kenya. On thesame 10% of the samples, soil texture was analyzed using laser dif-fraction and soil organic carbon (SOC) was analyzed using flash

Fig. 1.Map of Ethiopia showing annual average precipitation (MAP) and the locations of the sites included in this study. The map was derived with data fromWorldClim version 1.4(Hijmans et al., 2005).

268 T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

dynamic combustion using a Thermo Scientific CN analyzer at ICRAF,after acidifying the samples with HCl at 80 °C.

3.3. Remote sensing data analysis

In the current study, we used Landsat ETM + imagery from May(Merar) and November (Kutaber, Dambidolo and Mega) 2009 tomap land degradation risk factors and soil functional properties atlandscape scale. Prior to model development, satellite digital num-bers (DNs) were converted to ground reflectance through radiomet-ric calibrations and corrections of the imagery, including for sensorgains and offsets, solar irradiance and solar zenith angles. A set ofsemi-automated scripts developed by Gumbricht (2011) were usedto generate an archive of standardized surface reflectance (SRF) im-ages following methods developed and adapted from Chavez (1996).

3.4. Model development

3.4.1. Soil IR spectral data pre-processing and calibrationA soil MIR spectral library consisting of 3378 samples was used,

which included 764 reference soil samples that had analytical data onSOC and pH using the soil analytical methods described above. The cal-ibration models (see Terhoeven-Urselmans et al., 2010) developedbased on the training dataset were applied to the full library (N =3,378) and then related to satellite image reflectance (Section 3.4.2).Prior to the development of calibration models, MIR spectra werez-score normalized and linear feature selection (decorrelation) was ap-plied for data reduction using a linear correlation filter (Yu & Liu, 2003)to retain only the information-rich parts of the spectra.

3.4.2. Statistical procedures and generation of mapsA range of statistical analytical methods and machine learning

techniques are available for multivariate calibration, multilevelmodeling of ecological variables, and clustering and classification ofsoil IR spectral libraries and remote sensing imagery. In recentyears, ensemble methods have been introduced to increase predictionaccuracy in cases where base learners are weak, such as in classifica-tion trees. In these approaches, variable selection bias can be elimi-nated through the use of split selection criteria (Strobl et al., 1993).In the current study random forest (RF) models (Breiman, 2001;Svetnik et al., 2003) and gradient boosting techniques (Friedman,1999) were applied for classification of land degradation risk factorsand for prediction of soil pH and SOC, using Landsat ETM + image

reflectance bands as predictors. These model classes have become in-creasingly important for ecological applications (Cutler et al., 2007;Prasad et al., 2006) and have been applied for mapping of soil carbonbased on terrain attributes (Grimm et al., 2008), although their appli-cation is relatively rare in soil science. In a random forest model, a dif-ferent bootstrap sample from the original data is used during theconstruction of each tree, generating a test set classification for eachcase in about one-third of the trees. At the end of the run, take j tobe the class that got most of the votes every time case n wasout-of-bag. The proportion of times that j is not equal to the trueclass of n averaged over all cases is the out-of-bag error estimate.This has been proven to be unbiased in many tests. We assessed var-iable importance for each predictor by calculating the total decreasein node impurities from splitting on the variable, averaged over alltrees (Genuer et al., 2010).

Independent training and test data were generated by randomlyselecting 64% of the original data for model training and using theremaining independent random subset of 36% for validation. Wealso assessed the influence of the predictors included in the models.This is a particularly useful aspect of random forest models in remotesensing of ecological variables to better understand the relative influ-ence of the predictors. These variable importance measures have alsobeen shown to be useful for model reduction (e.g. to build simplermodels) (Genuer et al., 2010).

For the modeling of presence/absence of soil erosion and RDR50we calculated the sensitivity, specificity, Kappa and the area underthe ROC (receiver operating characteristic) curve (or AUC) for eachmodel (Fielding & Bell, 1977). Sensitivity measures the proportionof actual positives that are correctly identified. Specificity is the in-verse and measures the proportion of correctly predicted true nega-tives. The Kappa statistic is the proportion of correctly classifiedlocations after accounting for the probability of chance agreement.

The spatial distribution of soil pH and SOC concentrations in thetop 20 cm of the soil profile were mapped for the study areas andare presented as continuous surfaces. The prediction performance ofthe RF models for SOC and pH were tested by comparing predictedversus measured values when applying the training model to the in-dependently generated test data set. To assess the uncertainty ofmaps generated for pH and SOC, we calculated the coefficient of var-iation (CV) between each regression model (tree) in the ensemble foreach image pixel (Sexton & Laake, 2009). Predicted soil pH valueswere also classified into classes: acid (3–4.5); moderately acid (4.5–6.5); neutral (6.5–7.5) and alkaline (7.5–9).

Table 1Summary of ecological site characteristics for the sites included in this study.

Selected ecological site characteristics Mega Dambidolo Kutaber Merar

Cultivation (%) 0 75 54 40Average tree density (trees ha−1) 30 64 159 4Average shrub density (shrubs ha−1) 88 48 216 140Average slope (°) 1.5 7.4 18 2.2Average herbaceous cover rating (%) 15–40 15–40 4–15 4–15Presence of soil conservationmeasures (%)

0 0 50 0

Table 2Summary statistics of selected basic topsoil properties for the 10% of soil samples selected for wet chemistry analyses (n = 64). Values are medians by depth, while the values in thebrackets are lower and upper quartiles, respectively.

Site SOC pH Sand Clay Ca Mg K Na

cmolc kg−1

Dambidolo (n = 32) 36.5 (20.8/51.5) 6.4 (6.0/6.5) 13 (8/21) 57 (45/69) 14.8 (9.9/26.3) 4.7 (3.3/6.8) 0.7 (0.5/1.0) 0.2 (0.2/0.2)Kutaber (n = 30) 18.7 (16.5/24.1) 6.5 (6.3/6.8) 14 (12/16) 56 (50/61) 30.6 (27.4/36.4) 11.0 (10.2/14.2) 0.5 (0.3/1.2) 0.2 (0.2/0.3)Mega (n = 32) 20.0 (15.5/21.2) 8.2 (8.0/8.2) 10 (8/19) 73 (63/77) 40.9 (27.5/51.5) 10.0 (8.5/11.8) 2.2 (1.2/2.6) 0.2 (0.2/0.5)Merar (n = 32) 25.2 (20.6/29.0) 8.1 (8.1/8.2) 9 (5.5/17.1) 68 (51/76) 79.0 (61.9/97.7) 4.5 (3.6/5.4) 1.1 (0.9/1.4) 0.4 (0.3/0.5)

Table 3Summary of random forest model classification performance for erosion and RDR50prevalence overall, including ranking of the importance of input Landsat ETM + bands.

AUC Sensitivity Specificity Kappa Predictor rankinga

Erosion 0.81 0.72 0.71 0.42 SLOPE,RFRED,PVI,RFGREEN,RFMID1,RFNIR,RFMID2

RDR50 0.83 0.75 0.74 0.49 RFBLUE,RFMID1,PVI,RFMID2,RFGREEN,RFRED,RFNIR

a RF = calibrated ground reflectance, PVI = perpendicular vegetation index (Rich-ardson & Wiegand, 1977).

269T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

4. Results and discussion

4.1. Ecological site characteristics

Reference soil analysis indicates that Dambidolo has the highestmedian SOC and lowest pH of the four sites, while Mega and Merarhave the highest median clay contents and also the highest pH values(Table 2). Merar has extremely high exchangeable calcium (Ca), asexpected given the presence of Vertisols in this site, while Kutaberhas similar SOC to Mega and similar pH to Dambidolo (Table 2).Land uses and land cover were highly variable between the studysites, ranging from intensively cultivated areas to uncultivated savan-na systems. For example, in Dambidolo about 75% (7500 ha) of thearea is under cultivation, while Mega is a traditional pastoralist sa-vanna, as summarized in Table 1. Herbaceous cover ratings were rel-atively low in all sites, ranging between 4–15% (in Kutaber and

Merar) and 15–40% (in Dambidolo and Mega). Average tree (woodyvegetation taller than 3 m) densities varied from 4 trees ha−1 inMerar to 159 trees ha−1 in Kutaber, with predominantly Eucalyptusspp. in the latter site. Shrub densities also showed a high level of var-iability between sites (Table 1).

4.2. Predicting and mapping soil erosion and root depth restrictions

Soil erosion prevalence (based on field survey data) was high in allsites, except Merar, which is mainly on a plateau with low annualrainfall. The most influential predictors for soil erosion were slope,red reflectance and the perpendicular vegetation index (PVI —

Richardson & Wiegand, 1977), while blue and mid-infrared bandswere most influential for predicting RDR50, as shown in Table 3. AnAUC of about 0.8 was achieved for predictions of presence/absenceof both soil erosion and RDR50. An AUC of 1 would indicate a perfect

Fig. 2. Maps of soil erosion (left column) and RDR50 prevalence (right column), expressed as probability estimates for each pixel (from top to bottom: Merar, Kutaber, Dambidolo,Mega). The points (dots) on the maps are sampling locations, while the coordinates along the side of the maps are decimal degrees (WGS84).

270 T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

classification result, while 0.5 would mean that the model perfor-mance is no better than random.

As is evident from the maps (Fig. 2), soil erosion prevalence washighest in Dambidolo, followed by Mega, while within-site variationwas highest in Merar. The Merar site was characterized by distinctlydifferent landforms and soils, with part of the site located on a plateauof predominantly Vertisols (southern and eastern section). This partof the site showed low prevalence of soil erosion. Merar also hadthe lowest occurrence of soil erosion on average, but there were high-ly eroded areas in the west and northwest of the site, on landformsconsisting of secondary limestone (Badel & Mishra, 2007) (Fig. 2).The latter areas were predominantly used as rangelands. In Kutaber,soil erosion was widespread, particularly on steep slopes, but in gen-eral not as severe as for example parts of Dambidolo or Mega. Kutaberhas steep slopes and shallow soils (Table 1) and root depth restric-tions are prevalent throughout the site. In Mega, root depth restric-tions occurred mainly on hill slopes and at the base of hills, whilethe flatter plains (savannas) did not show much evidence of restric-tions. However, erosion was relatively prevalent also in much of thesavanna systems, even throughout the flatter parts of the site(Fig. 2), probably due to overgrazing and exposure of soils. Soils inDambidolo are relatively deep and root depth-restrictions weretherefore not very prevalent in this site, despite the high incidenceof soil erosion. The two upper panels in Fig. 6 show the distributionsof erosion and RDR50 prevalence, respectively, based on the pixelvalues for the predicted surfaces in Fig. 2. These distributions clearlyshow the differences between sites and the intensity of soil erosionin Dambidolo and Mega is evident, while the much flatter distributioncurve for Merar shows the low erosion prevalence on the plateau(peak around 30%). Kutaber has an almost normal distributioncurve for soil erosion, with intensities (prevalence rates) around60% (Fig. 6). The distribution curve for root depth restrictions is sim-ilar to that of erosion in Mega, although much flatter, whileDambidolo has a distinctively different distribution for root depth re-strictions than soil erosion (Fig. 6). These findings have implicationsfor management strategies in the different sites, as discussed later.

4.3. Predicting and mapping SOC and pH

Comparisons of measured versus predicted pH and SOC usingLandsat reflectance yielded R-squared values of 0.86 and 0.79 for pHand SOC, respectively, for the independent validation data (Figs. 3and 4). These results indicate that we can predict SOC and pH wellfrom Landsat ETM + reflectance libraries. Prediction performancefor SOC was similar to that reported by Vågen et al. (2012) usinghigh resolution satellite imagery as well as Jarmer et al. (2010), andalso confirms the potential of ensemble approaches for the predictionof SOC using remote sensing data reported by Spencer et al. (2006). InFig. 5 we show the results of fitting the above model to a set ofLandsat ETM + images from 2009, covering the four sites includedin the study. These images are part of a continent-wide archive of im-agery currently stored at ICRAF for which these models are being ex-plored for mapping of soil degradation prevalence and soil functionalproperties across the African continent.

In Merar, predicted SOC values were highest in Vertisol areas, andpH was neutral to alkaline (Fig. 5). The distribution curve for predict-ed SOC in Merar (lower panels in Fig. 6) shows a peak around22 g kg−1, which represents these Vertisol areas, while the averagefor this site was around 20 g kg−1 (vertical line). A smaller peak inthe distribution is found at about 5 g kg−1, representing highly erod-ed shallow soils. These results are comparable to the results of otherstudies in this region (e.g. Badel & Mishra, 2007). Most of the cultivat-ed area in Merar was within the Vertisol areas of the upland plateau,with wheat and barley cultivation. The savanna grassland Mega siteand the steep highland site at Kutaber had lower SOC concentrationsthan Merar on average, with means of about 17 and 18 g kg−1,

respectively. Mega was the only site with significant occurrences ofalkaline soils according to our analysis, mainly occurring in depres-sions, most likely due to poor drainage. The areas with higher pHalso had higher SOC in this site, while the relationship between SOCand pH was the opposite in Dambidolo (Fig. 5). On average, SOC con-centrations were highest in Dambidolo (Fig. 6), but with a gradient ofdecreasing SOC from west to east (Fig. 5). Dambidolo was also themost humid site and had the lowest predicted pH (Fig. 6), but withhigh potential SOC storage both due to the humid climate and todeep soils (i.e. little occurrence of RDR50). A large proportion of theDambidolo site was converted to cultivation (from forest and grass-lands) during the last 7 years, which may be part of the explanationfor the relatively high SOC contents observed in this site. PredictedSOC concentrations may be classified into ranges for identifying prior-ity areas for recommendation of appropriate management interven-tions. However, thresholds for SOC are widely debated and wetherefore do not show results of these explorations here.

Fig. 3. Predicted versus measured SOC concentrations (g kg−1) for the independentvalidation dataset with Landsat reflectance data for the sites included in this study(n = 473).

Fig. 4. Predicted versus measured pH for the independent validation dataset withLandsat reflectance data for the sites included in the study (n = 473).

271T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

4.4. Implications for land management

While acknowledging that sustainable land management is highlycomplex in that it has dimensions beyond the ecological variablespresented in this paper, including social, economical and political as-pects (Bouma, 2002), we give some examples of how assessments andmapping of soil condition and land degradation risk can be useful fortargeting land management interventions. Through the developmentof maps at moderate to high resolution, land degradation hotspotsand areas with soil constraints such as low SOC, alkaline or acid soilscan be identified at scales relevant to individual farmers and Woredaextension agents (Figs. 2 and 5). Using such maps, spatially explicitmanagement recommendations can be made to target for examplesoil fertility management or soil and water conservation measures.

In Dambidolo our analysis shows higher prevalence of soil erosionin the southeastern and eastern parts of the site, as well as 35% lower

SOC concentrations in eroded versus non-eroded plots. These findingsnot only highlight the need for management options that reduce soilerosion but also manage soil carbon. These particular areas are stillpredominantly grasslands and interventions need to be appropriatefor optimizing stocking densities to reduce overgrazing and compac-tion. They are also likely to be vulnerable to severe land degradation ifconverted to agriculture. Approximately 75% of Dambidolo was culti-vated at the time of the field survey (Table 1), with conversion to ag-riculture having taken place largely since 2004, at least in part as aresult of resettlement schemes that began in the 1990s (Woube,1995). Also, soil and water conservation measures were largely ab-sent in this site. Soil conservation measures and possibly conservationagricultural strategies that incorporate crop residues and increase or-ganic matter status and stabilize soil carbon could be beneficial inter-ventions in cultivated areas of Dambidolo. In addition, 53% of thesampled plots were impacted strongly by fire, which is a commonly

Fig. 5.Maps of predicted topsoil (0–20 cm) pH (left column) and SOC (right column) for each study site (from top to bottom: Merar, Kutaber, Dambidolo, Mega). The points (dots)on the maps are sampling locations, while the coordinates along the side of the maps are decimal degrees (WGS84).

272 T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

used by farmers in the Gambela region to stimulate regrowth both ingrasslands and cropping systems (Woube, 1998).

Vertisol areas in Merar generally have higher SOC concentrationsand pH than the rest of the site, as expected given the inherent proper-ties of this soil type, and as reflected also in the maps in Fig. 5. Theseareas have low soil erosion prevalence, but our analysis indicates somepresence of root depth restrictions (soil physical constraints). Amelio-rating soil physical degradation can be difficult if caused by long-termplowing, which is likely to be the case in the wheat-producing area ofMerar. In contrast to Dambidolo, Merar is in a low rainfall region andfire was not a prominent landmanagement practice according to the in-formation collected at the time of the survey of this site.

Other studies from areas near Kutaber, show that this region hasbeen under intensive farming for decades (Tekle & Hedlund, 2000), ifnot centuries. The major soil constraints found in this site were physicaldue to steep slopes and shallow soils, and overall the site has moderatelyhigh soil erosion prevalence and high prevalence of root depth restric-tions (Fig. 2). Soil and water conservation structures in Kutaber needto be optimized to reduce erosion and increase soil depth as depthrestrictions are likely to be amajor constraint to agricultural productionwith continued soil erosion. Also, management options that increase

vegetation cover need to be considered both as part of soil conservationmeasures and to improve organic matter status.

Mega has significant areas with high erosion and the results showthat root depth restrictions are extensive. This region is predominantlypastoral, with 100% of the plots surveyed under grazing. Hence, themanagement of livestock in this area should focus on reducing over-grazing and compaction. The results of the current study show thatwhile tree and shrub densities are important ecological site descriptors,sites such as Mega do not have lower erosion prevalence in areas withhigher tree densities, on average. When herbaceous cover was takeninto consideration, the results suggested a reduction in erosion preva-lence of between 25 and 30% for areas with 40% herbaceous cover ormore. Hence, proposed rehabilitation strategies in grasslands mayfocus on restoration of herbaceous cover, including perennial grasses,and the implementation of vegetative and/or structural soil conserva-tion measures, particularly on steep slopes.

5. Conclusions

There is significant potential in the approaches applied in thisstudy, and particularly in the combination of systematic field surveys

Fig. 6. Summary of map predictions (all pixels within each site) for land degradation risk factors and soil functional properties. The density plots show distributions of predictedvalues, with means superimposed as vertical lines. The mosaic plot for pH shows predicted values categorized into classes based on acidity levels (orange = moderately acid,green = neutral and blue = alkaline) by site. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

273T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

with remote sensing data. Since the methods used were applied con-sistently across several projects, we were able to develop modelsbased on a wide range of soil and environmental conditions andapply these for site-specific prediction of land degradation risk factorsand indices of soil functional properties to four sites in Ethiopia. Theresults indicate that the ensemble modeling approaches used areable to explain most of the variability in SOC and pH as well as the oc-currence of soil erosion and root depth restrictions, when applied toan independent validation dataset. Further, the models developedcan be applied for mapping of these indicators of soil and land healthto generate spatially explicit recommendations for improved landmanagement in these landscapes. The results of the analysispresented in this paper shows that soil erosion is highly prevalentin both agricultural and pastoral land use systems, but with highlevels of spatial variability. This variability calls for modeling andmapping approaches that are spatially resolved enough for manage-ment interventions to be targeted appropriately.

Acknowledgments

This research was conducted with support from the Bill andMelinda Gates Foundation (BMGF), Wajibu MS (Kenya sites) andGhent University (DRC sites). We would like to thank the fieldteam, which sampled the Ethiopia sites: Biadglign DemissieMulawu(Mekelle University), Tesema Bekele Silewondim (Addis AbabaUniversity), Venance Kengwa (CIAT), Chris Ekise (CIAT) andZelalem Hadush (Mekelle University). We would also like to thankthe individual Woredas and farmers for allowing us access to con-duct the field surveys. We would like to thank the followingauthorities: In Dambidolo: the agriculture office of Oromiya Region,the agriculture and administrative offices of Kelem Wollega Zone,and the agriculture and administrative offices of Hawa GelanDistrict. In Kutaber: the zonal agricultural offices in Hayke andDese. In Merar: Tulli-Guled district of Somali Regional state andChinaksen district of Eastern Harerge Zone of Oromia RegionalState and Jijiga University. In Mega: the Yabello zonal administrativeand agriculture office.

References

Alaily, F., & Pohlmann, J. (1995). Soil mapping of Bir-Tarfawi region (SW-Egypt) basedon digital classification of Landsat-MSS-data. In H. P. Blume, & S. M. Berkowicz(Eds.), Arid ecosystems. Catena; advances in geoecology, 28. (pp. 89–107).

Badel, M., & Mishra, B. (2007). Characterization of some soils of Jijiga plain, Ethiopia.Journal of Food, Agriculture and Environment, 5, 416–424.

Baldock, J., Grundy, M., Griffin, E., Webb, M., Wong, M., & Broos, K. (2009). Building afoundation for soil condition assessment. CSIRO land and water science report.

Ben-Dor, E., & Banin, A. (1995). Near infrared analysis (NIRA) as a rapid method tosimultaneously evaluate several soil properties. Soil Science Society of America Journal,59, 364–372.

Berry, L. (2003). Land degradation in Ethiopia: Its extent and impact. Washington D.C:The World Bank.

Bishaw, B. (2001). Deforestation and land degradation in the Ethiopian highlands: Astrategy for physical recovery. Northeast African Studies, 8, 7–25.

Bouma, J. (2002). Land quality indicators of sustainable land management acrossscales. Agriculture, Ecosystems and Environment, 88, 129–136.

Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.Chang, C., Laird, D. A., Mausbach, M. J., & Hurburgh, C. R. (2001). Near-Infrared Reflec-

tance Spectroscopy—Principal components regression analyses of soil properties.Soil Science Society of America Journal, 65, 480–490.

Chavez, P. S. (1996). Image-based atmospheric corrections — Revisited and improved.Photogrammetric Engineering and Remote Sensing, 62, 1025–1036.

Coppock, D. L. (1993). Vegetation and pastoral dynamics in the southern Ethiopianrangelands. Implications for theory and management. In R. Behnke, I. Scoones, &C. Kerven (Eds.), Range ecology at disequilibrium: New models of natural variabilityand pastoral adaptation in African savannas. Proceedings of a meeting held 19–21November, 1990. (pp. 42–61). Woburn, United Kingdom: The Commonwealth Sec-retariat and the Overseas Development Institute.

Cutler, D. R., Edwards, T. C., Beard, K. H., Cutler, A., Hess, K. T., Gibson, J., et al. (2007).Random forests for classification in ecology. Ecology, 88, 2783–2792.

Daba, S. (2003). An investigation of the physical and socioeconomic determinants ofsoil erosion in the Hararghe Highlands, Eastern Ethiopia. Land Degradation and De-velopment, 81, 69–81.

Descheemaeker, K., Nyssen, J., Poesen, J., Raes, D., Haile, M., Muys, B., et al. (2006). Run-off on slopes with restoring vegetation: A case study from the Tigray highlands,Ethiopia. Journal of Hydrology, 331, 219–241.

Desta, S., & Coppock, D. L. (2004). Pastoralism under pressure: Tracking system changein Southern Ethiopia. Human Ecology, 32, 465–486.

Dragan, M., Feoli, E., Fernetti, M., & Zerihun, W. (2003). Application of a spatial decisionsupport system (SDSS) to reduce soil erosion in northern Ethiopia. EnvironmentalModelling & Software, 18, 861–868.

FAO (1998a). Digital soil map of the world and derived soil properties (CD-ROM). Landand water digital media series no 1.

FAO (1998b). The soil and terrain database for northeastern Africa (CD-ROM). Land andwater digital media series no 2.

Fielding, A. H., & Bell, J. F. (1977). A review of methods for the assessment ofprediction errors in conservation presence/absencemodels. Environmental Conservation,24, 38–49.

Foley, W. J., McIlwee, A., Lawler, I., Aragones, L., Woolnough, A. P., & Berding, N. (1998).Ecological applications of near infrared reflectance spectroscopy — a tool for rapid,cost-effective prediction of the composition of plant and animal tissues and aspectsof animal performance. Oecologia, 116, 293–305.

Friedman, J. H. (1999). Stochastic gradient boosting. Computational Statistics & DataAnalysis, 38, 367–378.

Genuer, R., Poggi, J., & Tuleau-malot, C. (2010). Variable selection using random forests.Pattern Recognition Letters, 31, 2225–2236.

Grepperud, S. (1996). Population pressure and land degradation: The case of Ethiopia.Journal of Environmental Economics and Management, 30, 18–33.

Grimm, R., Behrens, T., Märker, M., & Elsenbeer, H. (2008). Soil organic carbon concen-trations and stocks on Barro Colorado Island — Digital soil mapping using RandomForests analysis. Geoderma, 146, 102–113.

Gumbricht, T. (2011). Automated multispectral satellite reflectance calibration using SML.Hijmans, R., Cameron, S., Parra, J., Jones, P., & Jarvis, A. (2005). WorldClim, version 1.3.

edu/worldclim/worldclim.htmHill, J., & Schütt, B. (2000). Mapping complex patterns of erosion and stability in dry

Mediterranean ecosystems. Remote Sensing of Environment, 74, 557–569.Hurni, H. (1988). Degradation and conservation of the resources in the Ethiopian high-

lands. Mountain Research and Development, 8, 123–130.Hurni, H. (1993). Land degradation, famine, and land resource scenarios in Ethiopia. In D.

Pimentel (Ed.), World soil erosion and conservation (pp. 27–61). : CAB International.Jarmer, T., Hill, J., Lavee, H., & Sarah, P. (2010). Mapping topsoil organic carbon in

non-agricultural semi-arid and arid ecosystems of Israel. Photogrammetric Engi-neering and Remote Sensing, 75, 85–94.

Jong, S. de (1994). Applications of reflective remote sensing for land degradation studies ina Mediterranean environment. Amsterdam and Utrech: Koninklijk NederlandsAardrijkskundig Genootschap.

Kilic, S. (2009). Mapping soil drainage classes of Amik Plain using Landsat images.African Journal of Agricultural Research, 4, 847–851.

Kottek, M., Grieser, J., Beck, C., Rudolf, B., & Rubel, F. (2006). World map of the Köppen–Geiger climate classification updated. Meteorologische Zeitschrift, 15, 259–263.

Liebenberg, L. (2003). A new environmental monitoring methodology.wwwcybertrackercozaMethodologyhtml

McCarty, G. W., Reeves, J. B., III, Reeves, V. B., Follett, R. F., & Kinble, J. M. (2002).Mid-infrared and near-infrared diffuse reflectance spectroscopy for soil carbonmeasurement. Soil Science Society of America Journal, 66, 640–646.

Melendez-Pastor, I., Hernández, E. I., Navarro-Pedreño, J., & Gómez, I. (2012). Mappingsoil salinization of agricultural coastal areas in Southeast Spain. In B. Escalante(Ed.), Remote sensing applications (pp. 117–140). : InTech.

Moat, J., & Smith, P. P. (2007). CEPF Madagascar vegetation mapping [WWW document].http://www.kew.org/science-research-data/directory/projects/CEPFMadaVegMapping.htm

Monastersky, R. (1989). Spotting erosion from space. Science News, 136, 61.Norfleet, M. L., Ditzler, C. A., Puckett, W. E., Grossman, R. B., & Shaw, J. N. (2003). Soil

quality and its relationship to pedology. Soil Science, 168, 149–155.Nyssen, J., Poesen, J., Moeyersons, J., Haile, M., & Deckers, J. (2008). Dynamics of soil

erosion rates and controlling factors in the Northern Ethiopian Highlands —

towards a sediment budget. Earth, 711, 695–711.Odlare, M., Svensson, K., & Pell, M. (2005). Near infrared reflectance spectroscopy for

assessment of spatial soil variation in an agricultural field. Geoderma, 126,193–202.

Osborne, B. (1986). Near infrared spectroscopy in food analysis. Encyclopedia of analyt-ical chemistry (pp. 14).

Panah, S. K. A., & Pouyafar, A. M. (2005). Potentials and constraints of soil salinity studiesin two different conditions of Iran using Landsat TM data. 31st InternationalSymposium on Remote Sensing of Environment.

Pfeffermann, D., & Nathan, G. (1981). Regression analysis of data from a cluster sample.Journal of the American Statistical Association, 76, 681–689.

Pinet, P., Kaufmann, C., & Hill, J. (2006). Imaging spectroscopy of changing Earth'ssurface: A major step toward the quantitative monitoring of land degradationand desertification. Comptes Rendus Geoscience, 338, 1042–1048.

Piot, Y., Blaikie, P. M., Jackson, C., & Palmer-Jones, R. (1995). Rethinking research onland degradation in developing countries. World Bank discussion papers.

Prasad, A. M., Iverson, L. R., & Liaw, A. (2006). Newer classification and regression treetechniques: Bagging and random forests for ecological prediction. Ecosystems, 9,181–199.

Raudenbush, S., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data anal-ysis methods. Advanced Quantitative Techniques in the Social Sciences Series 1, XXIV, 485.

Renard, K. G., & Ferreira, V. A. (1993). RUSLE model description and database sensitiv-ity. Journal of Environmental Quality, 22, 458–466.

274 T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275

Author's personal copy

Richardson, A., & Wiegand, C. (1977). Distinguishing vegetation from soil backgroundinformation (by gray mapping of Landsat MSS data). Photogrammetric Engineeringand Remote Sensing, 43, 1541–1552.

Sexton, J., & Laake, P. (2009). Standard errors for bagged and random forest estimators.Computational Statistics & Data Analysis, 53, 801–811.

Shepherd, K. D., & Walsh, M. G. (2002). Development of reflectance libraries for char-acterization of soil properties. Soil Science Society of America Journal, 66, 988–998.

Shrestha, R. P. (2006). An investigation on the relation between the remote sensingdata and soil salinity. Asian Association on Remote Sensing 27th Asian Conferenceon Remote Sensing ACRS (pp. 83–88).

Smith, S. J., Williams, J. R., Menzel, R. G., & Coleman, G. A. (1984). Prediction of sedi-ment yield from southern plains grasslands with the modified universal soil lossequation. Journal of Range Management, 37, 295–297.

Sonneveld, B. G. J. S., & Keyzer, M. A. (2003). Land under pressure: soil conservation con-cerns and opportunities for Ethiopia. Land Degradation & Development, 14, 5–23.

Spencer, M. J., Whitfort, T., McCullagh, J., & Bui, E. (2006). Dynamic ensemble approachfor estimating organic carbon using computational intelligence. ACST06 Proceed-ings of the 2nd IASTED International Conference on Advances in Computer Scienceand Technology (pp. 186–192). : ACTA Press.

Strobl, C., Zeileis, A., Boulesteix, A. L., & Hothorn, T. (1993). Variable selection bias inclassification trees and ensemble methods. Book of abstracts (pp. 159). : Citeseer.

Su, H., Ransom, M. D., & Kanemasu, E. T. (1989). Detecting soil information on a nativeprairie Using Landsat TM and spot satellite data. Soil Science Society of America Jour-nal, 53, 1479–1483.

Svetnik, V., Liaw, A., Tong, C., Culberson, J. C., Sheridan, R. P., & Feuston, B. P. (2003).Random forest: A classification and regression tool for compound classificationand QSAR modeling. Journal of Chemical Information and Computer Science, 43,1947–1958.

Taddese, G. (2001). Land degradation: A challenge to Ethiopia. Environmental Management,27, 815–824.

Tekle, K., & Hedlund, L. (2000). Land cover changes between 1958 and 1986 in Kaludistrict, southern Wello, Ethiopia. Mountain Research and Development, 20,42–51.

Terhoeven-Urselmans, T., Vagen, T. -G., Spaargaren, O., & Shepherd, K. D. (2010).Prediction of soil fertility properties from a globally distributed soil mid-infraredspectral library. Soil Science Society of America Journal, 74, 1–8.

Thompson, S. K. (1991). Adaptive cluster sampling: Designs with primary and second-ary units. Biometrics, 47, 1103.

Vågen, T. -G., Davey, F., & Shepherd, K. D. (2012). Land health surveillance: Mappingsoil carbon in Kenyan rangelands. In P. K. R. Nair, & D. Garrity (Eds.), Agroforestry —

The future of global land use. Springer.Vågen, T. -G., Shepherd, K. D., & Walsh, M. G. (2006). Sensing landscape level change in

soil fertility following deforestation and conversion in the highlands of Madagascarusing Vis–NIR spectroscopy. Geoderma, 133, 281–294.

Vågen, T. -G., Shepherd, K. D., Walsh, M. G., Winowiecki, L. A., Tamene Desta, L., &Tondoh, J. E. (2010). AfSIS technical specifications — Soil health surveillance, Africa.Nairobi, Kenya: CIAT (the AfSIS project).

Wischmeier, W. H., & Smith, D. D. (1962). Soil-loss estimation as a tool in soil andwater management planning. International Association of Scientific Hydrology Publi-cation, 59, 148–159.

Woube, M. (1995). Ethnobotany and the economic role of selected plant species inGambela, Ethiopia. Journal of Ethiopian Studies, 28, 69–86.

Woube, M. (1998). Effect of fire on plant communities and soils in the humid tropicalsavannah of Gambela, Ethiopia. Land Degradation Development, 9, 275–282.

Yu, L., & Liu, H. (2003). Feature selection for high-dimensional data: A fast correlation-based filter solution. Twentieth International Conference on Machine Learning(ICML-2003), 8 (Washington, DC).

Zornoza, R., Guerrero, C., Mataix, J., Morales, J., Mayoral, M., & Mart, M. (2006). The useof near infrared spectroscopy (NIR) for soil quality and soil degradation assess-ment. Geophysical research abstracts (pp. 05033). : European Geoscience Union.

275T.-G. Vågen et al. / Remote Sensing of Environment 134 (2013) 266–275