**5.2 Object oriented approaches (Bock et al. 2005; Díaz Varela et al. 2008)**

Bock et al. (2005) have proposed an object-oriented approach for EUNIS habitat mapping using remote sensing data at multiple scales with good results. The approach performs well when applied to high resolution satellite data (Landsat 30m) for the production of habitat maps at regional level with coarse thematic resolution; also it performs extremely well when applied to very high spatial resolution data (Quickbird 0,7m) for the production of local scale maps with fine thematic resolution.

The use of a multi-scale segmentation (implemented in the software package eCognition, Definiens, http://www.definiens.com) allow for the accurate classification of habitat types, which occur at different scales: for example, large-scale woodland habitat can be detected at coarser segmentation levels, while small-scale habitats such as woodland corridors can be detected at finer segmentation levels.

The main advantages of the object-oriented approaches to habitats mapping are: i) the ability to integrate ancillary data into the classification processes, related to shape, texture, context, etc.; ii) the option of developing knowledge-based rules in the classification process. Both questions make especially possible the accurate identification of habitats with similar spectral properties. Some results that show the advantage of these issues are (Bock et al. 2005): i) the effective separation of different grassland types like calcareous and mesotrophic grassland habitats to a high degree of accuracy through the use of geological data; ii) the use of multi-temporal remote sensing data to distinguish among arable lands, manage grasslands and semi-natural habitats.

#### **5.3 The use of binary classifications by decision trees (DT) (Boyd et al. 2006; Foody et al. 2007; Franklin et al. 2002; Franklin et al. 2001)**

Some studies have shown binary classifications as one of the more appropriate methods to identify habitats in the territory. Binary classifications can be implemented by nonparametric Decision Trees (DT) algorithms. Some of these studies have focused on the mapping of one specific thematic class (Boyd et al. 2006; Foody et al. 2007) hypothesizing that non-parametric algorithm would be more suitable to habitats of conservation interest because of the scarce spatial distribution usually associated to them (the size of the training sample will be smaller). Other studies have combined that kind of techniques (binary classifications by DT) in hybrid approaches (Franklin et al. 2001). The hybrid approaches assumes that parametric algorithms like standard maximum likelihood (ML) are the best

Assessing Loss of Biodiversity in Europe

*Terras do Miño* 

Through Remote Sensing: The Necessity of New Methodologies 33

Fig. 2. Flowchart of the methodological steps for the model applied to Biosphere Reserve

band was not included in the classification processes.

habitats should be favored by those variables.

Early spring (March 26, 2002), late spring (May 26, 2001) and summer (August 17, 2002) images were selected for this model to account for the seasonal trends in vegetation communities. Images were geometrically registered using ground control points (GCP), first order transformations and nearest neighbor interpolation. The August 2002 image was georeferenced to 1/25,000 digital maps, produced by the National Geographical Institute of Spain and used as a reference for geometrical correction of the other images. Atmospheric correction was based on the dark-object technique proposed by Chavez(Chavez 1996). Correction for effects of ground slope and topographic orientation was computed using the Lambertian cosine method initially proposed by Teillet et al. (Teillet et al. 1982) and later modified by Civco (Civco 1989). To model illumination conditions a Digital Elevation Model (DEM) was generated using contour lines from 1/5000 digital cartography. The thermal

It was hypothesized that derived and ancillary variables would provide critical information for landscape classification and enable the identification of complex habitats. For example, topographic features and vicinity to fluvial corridors have an important influence on the distribution of natural and semi-natural habitats; therefore, the discrimination of this type of

The input dataset for classification processes included satellite derived variables (reflectance, vegetation indices, texture measures and spectral mixture analysis) along with continuous ancillary data. Reflectance bands were included using principal components transform (Mather 2004) in order to reduce the dimensionality of the dataset and optimize the number of training samples. Vegetation indices were: NDVI (Normalized Difference Vegetation Index) (Rouse et al. 1974) and NDII (Normalized Difference Infrared Index; using Landsat TM Band 5) (Hunt and Rock 1989). Texture measures (homogeneity using Band 3 of each Landsat ETM+ image) were calculated using the co-ocurrence matrix as designed by Haralick et al. (Haralick et al. 1973). The co-ocurrence matrix was computed from a window of 3x3 pixels, which was considered an optimum size for measuring neighbor conditions. Linear spectral mixture analysis (SMA) (Mather 2004) generated

option with spectrally different habitats while applies non-parametric algorithms to other complex habitats, in a so-called Integrated Decision Tree Approach (IDTA). The IDTA (Franklin et al. 2001) consist on a process with a simple set of classification decision steps, readily understood and repeatable. The approach allows mixing unsupervised, supervised and stratification decision rules such that requirements for training data were minimized.

The general advantages of this kind of approaches are: i) those linked to the use of nonparametric algorithms (Tso and Mather 2009), for example less restrictions with the size of the training sample; ii) the use of key input variables combined with key algorithms defined following specific characteristics of individual habitats; iii) The use of a type of geospatial input data (nominal, ordinal, interval or ratio data like forest inventory maps, biophysical and derived maps) that are difficult to incorporate into a statistical classifier.

#### **5.4 New model based on the identification of ecological-units and on the selection of habitat-key variables (Advances from Martinez et al., 2010)**

The methodology proposed in this model is summarized in Figure 2. The approach includes two main steps: the adaptation of an international classification scheme and the generation of *ecological unit* maps. The concept of *ecological unit* goes farther than land cover notion: through ecological expert knowledge it is possible identify in the study area ecological units directly related to Annex I habitats (Table5). Each ecological unit is linked to a distinctive set of characteristic habitats through ecological expert knowledge. Consequently, the system allows identifying and assessing the habitats listed in Annex I of Habitats Directive 92/43/EC through the identification of land covers by remote sensing.

Ecologically significant units of analysis were defined (Table 5), based on the *EUNIS* pan-European classification proposed by the EEA and the CORINE Biotopes System of Classification (www.eea.europa.eu, last accessed, May 2011). By using these systems, the approach can be applicable to other European regions and it will produce cross-comparable results.

The generation of *ecological unit* mapping is based on the selection of key input variables (spectral, derived and ancillary variables) as a function of the main characteristics of the target habitats and on the use of a standard maximum likelihood classification (MLC) algorithm (Swain and Davis 1978). The target habitats are those defined in the Annex I of the Habitats Directive for the Atlantic Biogeographical Region. Because of the *not direct* correspondence between spectral classes and habitat types, we propose the combination of ecological expert knowledge to find the relationship between both of them (step 1) along with the selection of suitable input variables (step 2) in order to achieve the best possible classification of habitats.

The study area was the Biosphere Reserve of *Terras do Miño* in the Northwest of Iberian Peninsula (Figure 3).The classification process was undertaken by the use of multi-temporal Landsat ETM+ images and ancillary data.

Input variables were rectified to the Universal Transverse Mercator Projection (UTM 29T) using the European Datum 1950 (ED50) and resampled to 30m grid size. Training samples were taken by fieldwork. They were located with a global positioning system (GPS) differential receiver. Training sites were selected as to be large and representative enough to characterize each target class and provide efficient and unbiased estimators using stratified sampling. At least 50 additional points per ecological unit were surveyed on the field for results assessment. These data were not used in the training process.

option with spectrally different habitats while applies non-parametric algorithms to other complex habitats, in a so-called Integrated Decision Tree Approach (IDTA). The IDTA (Franklin et al. 2001) consist on a process with a simple set of classification decision steps, readily understood and repeatable. The approach allows mixing unsupervised, supervised and stratification decision rules such that requirements for training data were minimized. The general advantages of this kind of approaches are: i) those linked to the use of nonparametric algorithms (Tso and Mather 2009), for example less restrictions with the size of the training sample; ii) the use of key input variables combined with key algorithms defined following specific characteristics of individual habitats; iii) The use of a type of geospatial input data (nominal, ordinal, interval or ratio data like forest inventory maps, biophysical

**5.4 New model based on the identification of ecological-units and on the selection of** 

The methodology proposed in this model is summarized in Figure 2. The approach includes two main steps: the adaptation of an international classification scheme and the generation of *ecological unit* maps. The concept of *ecological unit* goes farther than land cover notion: through ecological expert knowledge it is possible identify in the study area ecological units directly related to Annex I habitats (Table5). Each ecological unit is linked to a distinctive set of characteristic habitats through ecological expert knowledge. Consequently, the system allows identifying and assessing the habitats listed in Annex I of Habitats Directive

Ecologically significant units of analysis were defined (Table 5), based on the *EUNIS* pan-European classification proposed by the EEA and the CORINE Biotopes System of Classification (www.eea.europa.eu, last accessed, May 2011). By using these systems, the approach can be applicable to other European regions and it will produce cross-comparable

The generation of *ecological unit* mapping is based on the selection of key input variables (spectral, derived and ancillary variables) as a function of the main characteristics of the target habitats and on the use of a standard maximum likelihood classification (MLC) algorithm (Swain and Davis 1978). The target habitats are those defined in the Annex I of the Habitats Directive for the Atlantic Biogeographical Region. Because of the *not direct* correspondence between spectral classes and habitat types, we propose the combination of ecological expert knowledge to find the relationship between both of them (step 1) along with the selection of suitable input variables (step 2) in order to achieve the best possible

The study area was the Biosphere Reserve of *Terras do Miño* in the Northwest of Iberian Peninsula (Figure 3).The classification process was undertaken by the use of multi-temporal

Input variables were rectified to the Universal Transverse Mercator Projection (UTM 29T) using the European Datum 1950 (ED50) and resampled to 30m grid size. Training samples were taken by fieldwork. They were located with a global positioning system (GPS) differential receiver. Training sites were selected as to be large and representative enough to characterize each target class and provide efficient and unbiased estimators using stratified sampling. At least 50 additional points per ecological unit were surveyed on the field for

and derived maps) that are difficult to incorporate into a statistical classifier.

**habitat-key variables (Advances from Martinez et al., 2010)** 

92/43/EC through the identification of land covers by remote sensing.

results assessment. These data were not used in the training process.

results.

classification of habitats.

Landsat ETM+ images and ancillary data.

Fig. 2. Flowchart of the methodological steps for the model applied to Biosphere Reserve *Terras do Miño* 

Early spring (March 26, 2002), late spring (May 26, 2001) and summer (August 17, 2002) images were selected for this model to account for the seasonal trends in vegetation communities. Images were geometrically registered using ground control points (GCP), first order transformations and nearest neighbor interpolation. The August 2002 image was georeferenced to 1/25,000 digital maps, produced by the National Geographical Institute of Spain and used as a reference for geometrical correction of the other images. Atmospheric correction was based on the dark-object technique proposed by Chavez(Chavez 1996). Correction for effects of ground slope and topographic orientation was computed using the Lambertian cosine method initially proposed by Teillet et al. (Teillet et al. 1982) and later modified by Civco (Civco 1989). To model illumination conditions a Digital Elevation Model (DEM) was generated using contour lines from 1/5000 digital cartography. The thermal band was not included in the classification processes.

It was hypothesized that derived and ancillary variables would provide critical information for landscape classification and enable the identification of complex habitats. For example, topographic features and vicinity to fluvial corridors have an important influence on the distribution of natural and semi-natural habitats; therefore, the discrimination of this type of habitats should be favored by those variables.

The input dataset for classification processes included satellite derived variables (reflectance, vegetation indices, texture measures and spectral mixture analysis) along with continuous ancillary data. Reflectance bands were included using principal components transform (Mather 2004) in order to reduce the dimensionality of the dataset and optimize the number of training samples. Vegetation indices were: NDVI (Normalized Difference Vegetation Index) (Rouse et al. 1974) and NDII (Normalized Difference Infrared Index; using Landsat TM Band 5) (Hunt and Rock 1989). Texture measures (homogeneity using Band 3 of each Landsat ETM+ image) were calculated using the co-ocurrence matrix as designed by Haralick et al. (Haralick et al. 1973). The co-ocurrence matrix was computed from a window of 3x3 pixels, which was considered an optimum size for measuring neighbor conditions. Linear spectral mixture analysis (SMA) (Mather 2004) generated

Assessing Loss of Biodiversity in Europe

1/5000 river map using raster processing.

(endmember-mz)+ FMo 3x3

MLC1 75,56 0,733 50,126\*

class, [PCs]: Principal Components,

are significantly different.

\*MLC2

Comparison with MLC1

[FMo]: modal filter

although the model was allowed to produce a residual image.

Through Remote Sensing: The Necessity of New Methodologies 35

non\_photosynthetic vegetation (NPV). SMA models the reflectance of each pixel as a linear combination of reflectance of those four components. It was assumed constrains of no negative values and that the four components explained the whole variation of reflectance,

Slope gradient was calculated from the DEM. Proximity to rivers was calculated from the

Results were assessed by cross-tabulation with a sample of pixels included in test plots. Global, user and producer accuracies were evaluated using an error matrix (Congalton and Green 2009). Additionally the Kappa analysis (Congalton and Green 2009) was used to evaluate the accuracy of the results: we used KHAT statistic to measure how well the remotely sensed classification agrees with the reference data, the Z statistic to determine the significance of a matrix error and the Z pairwise comparison to decide if two KHAT values

**(%) KHAT** 

82,75 0,811

**CODE VARIABLES GA** 

PCs (may+march+august) + NDVI (my,ag) + NDII5 (my,ag) + prox. streams + SLOPE+ Homogeneity- B3 (my, mz, ag)+ WATER

MLC1 PCs (may+march+august) 75,56 0,733

Legend: [GA] Gloabl accuracy (%). [MLC]: Maximun likelihood classification, [\*] without *Bare Land*

**CODE GA (%) KHAT Z PAIR-WISE Z SCORES\*\*** 

Legend: [MLC]: Maximun likelihood classification [\*] Significant at the 95% confidence level. [\*\*]

The results of this model showed 82.75% global accuracy after the application of a modal filter. The best result provided a Kappa value (Congalton and Green 2009) of 0.811 with a Z value indicating very good agreement between classification results and the reference data. Tables 6 and 7 shows the accuracy assessment for two processes: i) one of them based on the principal components of the three images (MLC1); ii) the second one also includes the group of ancillary and derived variables (MLC2). Some variables like slope, distance to rivers, NDVI, NDII and homogeneity showed its valuable potential (Table 9). The combination of all of them in the best MLC trial produced a significant increase in global accuracy along with an increase on user and producer accuracies for the most part of the classes of habitats (Table 8). MLC2 showed user and producer accuracies above 70% and 80% in the most part

Table 6. Global accuracies for parametric multi-temporal processes (MLC algorithm)

MLC2 82.75 0,811 62,089\* 3,496569\*

Table 7. Kappa analysis for parametric multi-temporal processes (MLC algorithm)

of habitats. Only WH and forests showed producer or user accuracies less than 60%.

endmember spectra, which are defined as the proportion of each pixel covered by a basic spectral class. The included endmembers were water, soil, green vegetation (GV), and


Table 5. Ecological units directly related to CD 92/43/EEC habitats in the proposed model

34 Biodiversity Loss in a Changing Planet

endmember spectra, which are defined as the proportion of each pixel covered by a basic spectral class. The included endmembers were water, soil, green vegetation (GV), and

**ECOLOGICAL UNIT NATURA 2000 Code** 

**Standing Water** 3110/3120/3130

**WH** Bogs (Raised and blanket bogs) and Atlantic wet heaths 7130/7110\*/7120

**RF** Alluvial and riparian forests 91D0\*/91E0\*/91F0

**HM** Tall and mid-herb humid meadows 6430,6410

**DF** Deciduous oak forests 9230

**DH** Siliceous rocky habitats and dry heaths 4030/8220

and grasslands 6410/6510

grasslands 6410/6510

Table 5. Ecological units directly related to CD 92/43/EEC habitats in the proposed model

*Natural –Seminatural Landscape* 

**Inland no-wooded wetlands** 

**Inland wooded wetlands** 

*Anthropic Landscape* 

**Forest plantations P** *Pine sp.* groves

**BL** Bare land

**Other natural and seminatural forests** 

**Rocky habitats and other heaths** 

**E** *Eucalyptus sp.* plantations

**TF** Rural system mainly made up of pastures

**TR** Rural system mainly made up of corn and pasture in rotations

**CG** Traditional rural mosaic with fenced fields, dominated by crops

**WG** Traditional rural mosaic with fenced fields, dominated by wet

**B** Buildings for agricultural, forestry and industrial use

**Transformed rural landscape**

**Traditional rural landscape**

**Man-made landscape** 

**Ur** Urban areas (villages, towns) **ME** Mining exploitations

**I** Communication infrastructures

**Running Water W** Water courses **(Main Habitat)** 

3140/3150/3160 3260/3270

7230, 4020\*

non\_photosynthetic vegetation (NPV). SMA models the reflectance of each pixel as a linear combination of reflectance of those four components. It was assumed constrains of no negative values and that the four components explained the whole variation of reflectance, although the model was allowed to produce a residual image.

Slope gradient was calculated from the DEM. Proximity to rivers was calculated from the 1/5000 river map using raster processing.

Results were assessed by cross-tabulation with a sample of pixels included in test plots. Global, user and producer accuracies were evaluated using an error matrix (Congalton and Green 2009). Additionally the Kappa analysis (Congalton and Green 2009) was used to evaluate the accuracy of the results: we used KHAT statistic to measure how well the remotely sensed classification agrees with the reference data, the Z statistic to determine the significance of a matrix error and the Z pairwise comparison to decide if two KHAT values are significantly different.


Legend: [GA] Gloabl accuracy (%). [MLC]: Maximun likelihood classification, [\*] without *Bare Land* class, [PCs]: Principal Components, [FMo]: modal filter

Table 6. Global accuracies for parametric multi-temporal processes (MLC algorithm)


Legend: [MLC]: Maximun likelihood classification [\*] Significant at the 95% confidence level. [\*\*] Comparison with MLC1

Table 7. Kappa analysis for parametric multi-temporal processes (MLC algorithm)

The results of this model showed 82.75% global accuracy after the application of a modal filter. The best result provided a Kappa value (Congalton and Green 2009) of 0.811 with a Z value indicating very good agreement between classification results and the reference data.

Tables 6 and 7 shows the accuracy assessment for two processes: i) one of them based on the principal components of the three images (MLC1); ii) the second one also includes the group of ancillary and derived variables (MLC2). Some variables like slope, distance to rivers, NDVI, NDII and homogeneity showed its valuable potential (Table 9). The combination of all of them in the best MLC trial produced a significant increase in global accuracy along with an increase on user and producer accuracies for the most part of the classes of habitats (Table 8). MLC2 showed user and producer accuracies above 70% and 80% in the most part of habitats. Only WH and forests showed producer or user accuracies less than 60%.

Assessing Loss of Biodiversity in Europe

in Galicia (NW Spain)

accurate identification of many ecological units.

Through Remote Sensing: The Necessity of New Methodologies 37

The special interest of this approach comes from the use of an international classification system (EUNIS) that will allow cross-comparable spatial and temporal assessments and make the methodology extrapolated to other regions. The definition of ecological units goes farther than the simple *land cover* idea and it allows the definition of a direct relation win AnnexI habitats and consequently with species. Finally the use of a standard maximum likelihood algorithm based on the selection of key input variables makes possible the

Fig. 3. Localization of the Biosphere Reserve *Terras do Miño* and de SCI *Parga-Ladra-Támoga*

The output classification of this model and the CLC map (5th level) were spatially compared with a habitat map of the Site of Community Importance (SCI) Parga-Ladra-Támoga in the Northwest of Iberian Peninsula (which belongs the Biosphere Reserve of *Terras do Miño)*. This map was elaborated by photo interpretation through aerial photography with different scales ranging from 1/20000 until 1/2000 (Ramil et al. 2005). It also based on expertise fieldwork and its minimum mapping unit was 0,5ha. The map was the reference to evaluate

Again, spatial inconsistencies between both sources (CLC and the model applied to Biosfere Reserve *Terras do Miño*) were evaluated using this map by the analysis of coincidences in the

this site as a candidate to belong to *Natura 2000* ecological network.

cells of two different grids (UTM based): 1 and 10 km2.


Table 8. User and producer accuracies for parametric multi-temporal processes (MLC algorithm)


principal components of the three images

Table 9. Improvement in global accuracy for multi-temporal analyses after adding to the classification the layers showed in the table.

The contribution of topographical variables and vegetation indices to the habitat mapping accuracy is appropriate for the analyses; the combination of vegetation indices was relevant in the analyses, with improvements in global accuracy (to a maximum of 2.35% of accuracy increase when both NDVI and NDII were combined in the process) (Table 9). The topographical variables (Slope and MDE) improved also the global accuracy of multitemporal classifications, although to a lesser extent than variables like homogeneity. Texture measures and SMA components did not provide significant improvements in global accuracy in parametric methods. However they showed to be suitable for improving the discrimination of some particular classes like HM, WH or Ur. Therefore, they could be considered as interesting and helpful variables for nonparametric methods in order to get good discrimination of some classes.

WG 60,41 87,00 74,22 95,00 TR 97,19 100,00 96,30 100,00 CG 84,62 74,00 85,85 87,50 HM 43,94 72,5 73,81 77,50 DF 59,74 44,23 76,39 52,88 RF 52,88 55,0 59,84 73,00 WH 53,33 26,67 73,91 56,67 P 97,92 90,38 96,23 98,08 E 94,74 90,0 90,91 100,00 TF 84,26 91,0 91,01 81,00 TI 100,00 91,67 100,00 91,67 W 100,00 100,00 100,00 100,00 ME 100,00 100,00 100,00 83,33 Ur+I+B 83,33 100,00 86,96 100,00 BL 66,67 30,0 ---- ---- DH 82,28 81,25 83,56 76,25

Legend: [MLC]: Maximun likelihood classification

Table 8. User and producer accuracies for parametric multi-temporal processes (MLC algorithm)

**Improvement in** 

DEM + Slope 76,05 0,49 0,738 50,792\* 0,256709 Prox. to streams 76,34 0,78 0,742 51,213\* 0,412576 NDVI and NDII\*\* 77,91 2,35 0,758 53,603\* 1,248175 Homogeneity\*\* 77,22 1,66 0,751 52,467\* 0,867394 Endmember Water 75,75 0,19 0,735 50,442\* 0,102807 Legend: [\*] Significant at the 95% confidence level. [\*\*] Three images. [\*\*\*] In relation to the trial with

Table 9. Improvement in global accuracy for multi-temporal analyses after adding to the

The contribution of topographical variables and vegetation indices to the habitat mapping accuracy is appropriate for the analyses; the combination of vegetation indices was relevant in the analyses, with improvements in global accuracy (to a maximum of 2.35% of accuracy increase when both NDVI and NDII were combined in the process) (Table 9). The topographical variables (Slope and MDE) improved also the global accuracy of multitemporal classifications, although to a lesser extent than variables like homogeneity. Texture measures and SMA components did not provide significant improvements in global accuracy in parametric methods. However they showed to be suitable for improving the discrimination of some particular classes like HM, WH or Ur. Therefore, they could be considered as interesting and helpful variables for nonparametric methods in order to get

**multi-temporal\*\*\* KHAT Z** 

**Kappa analysis** 

**Pair-wise Z scores\*\*\*** 

**MLC1 MLC2 user producer user producer** 

**ECOLOGICAL UNITS** 

**Variables Global** 

principal components of the three images

good discrimination of some classes.

classification the layers showed in the table.

**accuracy** 

The special interest of this approach comes from the use of an international classification system (EUNIS) that will allow cross-comparable spatial and temporal assessments and make the methodology extrapolated to other regions. The definition of ecological units goes farther than the simple *land cover* idea and it allows the definition of a direct relation win AnnexI habitats and consequently with species. Finally the use of a standard maximum likelihood algorithm based on the selection of key input variables makes possible the accurate identification of many ecological units.

Fig. 3. Localization of the Biosphere Reserve *Terras do Miño* and de SCI *Parga-Ladra-Támoga* in Galicia (NW Spain)

The output classification of this model and the CLC map (5th level) were spatially compared with a habitat map of the Site of Community Importance (SCI) Parga-Ladra-Támoga in the Northwest of Iberian Peninsula (which belongs the Biosphere Reserve of *Terras do Miño)*. This map was elaborated by photo interpretation through aerial photography with different scales ranging from 1/20000 until 1/2000 (Ramil et al. 2005). It also based on expertise fieldwork and its minimum mapping unit was 0,5ha. The map was the reference to evaluate this site as a candidate to belong to *Natura 2000* ecological network.

Again, spatial inconsistencies between both sources (CLC and the model applied to Biosfere Reserve *Terras do Miño*) were evaluated using this map by the analysis of coincidences in the cells of two different grids (UTM based): 1 and 10 km2.

Assessing Loss of Biodiversity in Europe

**6. Conclusions** 

suitable RS and ancillary data.

be, at least not in a direct way).

data: at spatial and temporal levels.

that it is uncertain to use CLC as a proxy of habitat maps.

definite nor any that has been standardized across Europe.

Through Remote Sensing: The Necessity of New Methodologies 39

Results show that, although both sources (CORINE and the model) are based on LANDSAT images with 30m of spatial resolution, the inclusion of decisive variables in the classification processes along with the identification of ecological units was crucial. And it is again proved

To meet the requirements of European policies such as *Natura 2000* Network and the 2020 EU Biodiversity Strategy the development of more cost and time effective monitoring strategies are mandatory. Remote sensing (RS) techniques contribute significantly to biodiversity monitoring and several approaches have been proposed to get on-going requirement for spatially explicit data on the ecological units, and the value and threats against natural and semi-natural habitats (Bock et al. 2005; Weiers et al. 2004), but no

The major obstacles to get standardized scientific monitoring methodologies for habitat monitoring form a complex patchwork. The immense versatility of RS, the full range of RS techniques and products, has led to numerous potential approaches but all of them are dependent of many factors: i) firstly the large variability in the quality of input variables, their semantic, thematic and geometrical accuracy; many approaches have assumed the suitability and representativity of the selected geospatial data; ii) secondly, the possible variability of the spectral, spatial and temporal resolutions; iii) finally, the availability of

There is no a simple relationship between habitats and biophysical parameters like land covers (Groom et al. 2006). *Habitat classes* are not the same that *land cover classes* and the inconsistencies and gaps when a land cover map, as CORINE Land Cover, is used as a surrogate of a habitat map are significant and it should be evaluated in each case. It is necessary to develop *ad hoc* criteria to get the objective of identifying and monitoring habitats from remote sensing. It should be found the optimal way (cost effective and in an acceptable time, and with an optimal level of accuracy) to get from one unit of land cover (which can definitely be detected directly by remote sensing) to a unit of habitat (which may

At the European Community level the appropriate criteria for getting that relation should be achieved through EUNIS system (Martínez et al. 2010; Moss and Davies 2002) since it is a common denominator that is compatible with the requirements of Annex I of the Habitat Directive. It will support the standardization because it makes possible cross-comparable

In regard to habitat identification through RS recent researches have suggested different relevant considerations and requirements: study areas specific approaches; ecological expert knowledge implemented as decision rules; the implementation/inclusion of key input variables selected following specific characteristics of individual habitats; the integration of ancillary data into the classification processes, related to shape, texture, context; the use of non-parametric algorithms implemented through binary classifications or decision trees that allow to include nominal, derived and ancillary geospatial data and also are advantageous with scarce training samples; (Bock et al. 2005; Boyd et al. 2006; Foody et al. 2007; Franklin et

On the other hand, insufficient integration at different scales is one of the constraints of the current biodiversity monitoring programmes (Pereira and Cooper 2006) and it is also urgent

al. 2001; Kerr and Ostrovsky 2003; Martínez et al. 2010; Mücher et al. 2009).


Table 10. Spatial correlations between CD 92/43/EEC habitats cartography, Corine Land Cover cartography (5th level) and the MODEL *Terras do Miño* in the SCI Parga-Ladra-Tamoga (NW Spain)

At 10km CLC shows a total gap in the most part of the habitats (Table 10). Only heaths and the woodlands with *Quercus spp.* (9230) have good correspondence. At 1 km CLC shows total or very high gap in any case. On the other hand, the model of *Terras do Miño* shows good results at 10km. At 1km, the most part of the habitats present good correspondence or moderate gap. Only two habitats present total gap which corresponds to water courses (3260 and 3270) which can be assigned to the constraints of the spatial resolution of the images.

Results show that, although both sources (CORINE and the model) are based on LANDSAT images with 30m of spatial resolution, the inclusion of decisive variables in the classification processes along with the identification of ecological units was crucial. And it is again proved that it is uncertain to use CLC as a proxy of habitat maps.
