• Keine Ergebnisse gefunden

Chapter 1 Synopsis

1.5 Data and m ethods

Study 1Applications of remotesensingfor quantifyingandmapping ecosystemservices

In this study,literature was systematicallyreviewed to assess the applications ofremote sensing in quantifying and mapping the supplies and demands of ecosystem services.

The definition of ecosystem services used in this study is the Millennium Ecosystem Assessment (2005)that defines ecosystem services as “benefitsthatecosystems provide to support human well-being”. Ecosystem services were defined based on the TEEB classification (TEEB 2010). The review was limited to remote sensing applications in quantifying and mapping of selected provisioning and regulatory ecosystem services.

The other ecosystem services were excluded from this review due to lack of literature dealing with suchecosystem services. Articlespublished from year1990 to 2011were collected from peer reviewed journals using key words from the ISI Web of Science (www.webofknowledge.com) andGoogle Scholar (www.googlescholar.com) as primary searchengines. Thepublications were screenedwith respect to theecosystem services considered. This review focuses particularly on literature that used remotesensing for quantifying and mapping ecosystem services.

Study2Prosopis juliflora invasion andits impacts onEcosystemservices

Detecting invasive plant species in drylands using remote sensing starts with understandingof the characteristics ofthespecies andits seasonal variations interms of aspects such as greenness. P. juliflora has distinct features that differentiates it from other species in the Baadu area. Unlike other vegetation in the area, it remains green throughout the year which makes it easily detectable specially during dry seasons.

Figure 8 shows the seasonal changes of MODIS NDVI values comparing P. juliflora dominated pixels with dry uplandvegetation.

gaussian fit gaussian fit

a

b

c

d

Figure8a)Map showingprocessing window forpixelslocated indry uplands with sparseshrubs and grassesb)Monthly variationsof

The NDVI values from the P. juliflora dominated pixels remained high over all months ranging from about 0.60 in the dry seasons to above 0.8 duringwet seasons. Whereas, the NDVI values in the dry upland vegetation ranges from 0.10 during dry periods to 0.50duringwetseasons.With high resolution images,P.julifloracanbe identifiedfrom other wetland vegetations such as croplands and grasslands especially during the dry periods.Basedon the aforementionedpreliminaryassessmentofthecharacteristics ofP.

juliflora, images from dry seasons were selected for mapping invasion of the species.

Classification

To extractP.juliflora invadedlayers from the LandsatETM+(30m) andASTER (15m) satellite images, maximum likelihood supervised classification provided by Envi 5.0 software was used. Figure 9 illustrates the difference in spatialresolution between the satellite images.

Figure 9 Example of differenceintheresolutionof thea)Landsat ETM+ and b)ASTERsatellite imageszoomed nearalakeareainBaadu

Training areas representing different land cover classes were defined using data from field observations and Google earth images by digitizing polygon features that

Maximum likelihoodclassifier assumes normal distributionforeachbandandcalculates the probabilitythat an individual pixel belongs to a given class (Paola andSchowengerdt 1995;Perumal and Bhaskaran2010; Tuia etal. 2011). The term 'maximum likelihood' thus refers to using themaximumprobability as a guideline to assign a pixel to a class.

Pixels with probability below the set threshold will be left unclassified. In supervised classification, pixels are clustered into classes based on user-defined training areas (Richards 1999). The training areas (Region Of Interests, ROs) can be defined as multiple irregular polygons, vectors, and/or individual pixels. The accuracy of classification depends on separability between the ROIs (Oskouei and Busch 2012;

Zhang et al. 2012). Hence, points within each ROI should be homogenous and tightly clustering together to avoid overlap between classes.

Assessingtheimpacts of P. julifloraonecosystemservices

The impact of the invasive species, P.juliflora on ecosystem services was analyzed by calculating the area of important land categories (wetlands, agricultural lands & dry

Figure10 a)Scatterplot ofX(Red)vsY (NIR) bands of ASTER image showing separability of different land cover classes and b) its corresponding highlighted display of theimagec)profile plot of the Region of Interest for the classes.

lands) that is invaded by the species. Ecosystem services supplied by the above land categories were identified based on the Millennium Ecosystem Assessment, 2005 ecosystem services classification scheme in order to discuss potential loss of the services due to the invasion. For comparison, ecosystem services that can be supplied by P.

juliflora itself were also identified to discuss potential gains in terms of ecosystem services supplies due to introduction of the invasive species in the area.In spite ofthese, the beneficiaries of ecosystem services that are affected by the invasion of P. juliflora were identifiedanddiscussed.In the end, the pros and cons ofP.juliflorainvasion were assessed andsummarized basedon the impacts on supplies of ecosystem services and the beneficiaries affected.

Study3Undercovercroplandinsideforests Random Forest classification

For this study,level 3ARapidEye images were used to derive LULCclasses forthestudy site. The images were corrected for atmospheric and topographic errors using ATCOR 2/3software.Figure 11shows anexample ofcomparison between the original level 3A product and the image corrected for atmospheric and topographic errors.

each node of trees (Breiman 2001; Genuer et al. 2010). The final classification is thus the result of multiple decision trees (Figure 12).

Breiman (2001) expressed the RF classification as:

*h(X, Θk), k = 1, ...} ...eq. 1 where h(X, Θk) stands for the kth classifier, the *Θk } are independent identically distributed random vectors generated for the kth tree grown using the training set. X is an input vector for which a class is voted by each tree. The classification process involves random selection of input variables (mtry)ateach node of the trees (ntree) to calculate the best split within this subset(Genuer etal.2010; Gislason etal.2006;Rodriguez-Galiano etal.2012; Zhu etal.

2012).InFigure 12theends oftree1, tree2,...treenresult indecision fork1, k2,...knwhich are later used in voting class k.

Since its introduction by Breiman (2001), it become highly popular and has been a widely used statistical method for classification (Biau et al 2012; Genuer et al. 2010).

The RF method was preferably used for classification due to its multiple advantages over other classification approaches. For instance, Pal (2005) compared RF classifier with Support VectorMachines (SVMs)andfoundthat R Frequires less numberofuser definedparameters while it provides a comparable accuracywithin similartrainingtime with SVMs. The random selection of subsets of input variables minimizes correlation between classifiers (De’ath 2002;Rodriguez-Galiano etal. 2012). Gislason etal.(2006) stated that RF is able to handle large datasets since it is not sensitive to noise or overtraining. Besides, it provides estimates of relative importance of variables used in classification including the interaction between them (Rodriguez-Galiano et al. 2012;

Zhu etal.2012).Moreover, RF provides anoption forinternallyestimatingclassification error (Breiman 2001; Rodriguez-Galiano et al. 2012).

Validation ofthesatellite image classification

The results of the satellite image classification were validated using three sets of data: high resolution Google earth images, reference LULC classes recorded at the centre of sample plots andGPS photos taken in North, East, West andSouth (NEWS) directions from the centre point. The GPS photos were converted to points using QGIS 2.0.1 software andLULCclasses were identified on the photos.The GPSphotos were merged with the sample points to validate theresults of the image classification. The details of the steps used in the validation are provided in chapter 3 and 4.

BoostedRegression Trees

To identify influential variables forcropland area in the studysite, BoostedRegression Trees (BRTs), a method for fitting statistical models was used (Leathwick et al. 2006;

De'ath 2007;Elith etal.2008).BRTs are combinations ofalgorithms ofregression trees and boosting. Regression trees are models that use recursive binary splits to relate a response to their predictors while boosting is an adaptive method that improves predictive performance bycombining multiplesimple models (Elith et al. 2008).Thus, Boosted Regression Trees can be considered as an additive regression model that undergoes forward stagewise fitting without changing existing trees when the model enlarges (De'ath 2007). An example of BRTs decision tree structure is provided in Figure 13.