Introduction

Earth Observation (EO) imagery can provide a continuous synoptic view of land cover/use (LC/LU) patterns and LC/LU changes which have an impact on biodiversity loss (Turner et al. 2003; Townsend et al. 2009). LC/LU classes are not a suitable tool in assessing biodiversity in comparison to habitat classes, which are linked to species, communities and biotopes (Bunce et al. 2012a). According to Lengyel et al. (2008), habitat mapping constitutes one of the possible links among EO data, biodiversity monitoring and ecosystem status assessment, as habitats are linked to species occurrence, and the choice of appropriate taxonomies is a key point in the translation of LC/LU to habitat maps. However, whereas a conversion from LC/LU to habitat classes can be of great help, differences in taxonomies and definitions between the two domains (LC/LU and habitats) have so far limited the establishment of a unified approach for such translation.

Concerning LC/LU, the Food and Agricultural Organisation (FAO) Land Cover Classification System (LCCS) (Di Gregorio and Jansen 2005) was found to be the most useful for translating EO-derived LC/LU classes to habitat categories (Tomaselli et al. 2013), since it allows a better description of natural habitats in comparison to other classification systems (e.g., CORINE, Bossard et al. 2000). Consequently, Tomaselli et al. (2013) proposed a framework to translate LCCS classes into habitat categories, with these described according to Annex I to the Habitats Directive (92/43 EEC). The Habitats Directive is the main European Union (EU) legal instrument concerning biodiversity and conservation of natural habitats; its Annex I provides a list of natural habitats of Community interest within the EU and it is of central importance for international reporting and Natura 2000 site management (Ladoux et al. 2000; Mehtälä and Vuorisalo 2007). The list does not include anthropogenic and artificial habitats and, although organized into nine groups, it appears that there is no way to identify natural habitats in the field other than by gaining familiarity with the list by local experts who are generally used to interpret the descriptions when answering to national priorities (Bunce et al. 2012b). In addition, common approaches for monitoring changes in habitats require also definitions and rules that are harmonised continentally and globally. To address such problems, Bunce et al. (2011, 2012b) introduced the General Habitat Categories (GHC) methodology and tested it throughout Europe through in-field campaigns carried out, in different biogeographic regions, on sampling 1 km2 basic survey areas (0.25 km2 in complex landscapes), with these being extremely costly The GHC methodology for habitat classification and monitoring comes as an ecological refinement of the land cover categorization used in LCCS (Bunce et al. 2011) relying on the concept of Life Forms (LFs), which are strictly related to plant structure and morphology. However, LCCS and GHC use different definitions of LFs. In LCCS LFs are defined by physiognomic aspect of plant (Kuechler and Zonneveld 1988), whilst GHC are based on the concept of LFs as proposed by Raunkiaer (1934), defined by the position of buds during unfavourable season. The composition of LFs and the relative abundance in the biological spectra are a direct link among flora, vegetation and environmental (mainly climate) variables. The mapping of GHC, which are reported in Table 1, is firstly based on the identification of LFs and Non-LFs (e.g. artificial areas, rocks) through a set of expert decision rules (Bunce et al. 2011). Then, additional qualifiers on environment, site, management and species composition are collected during in-field campaigns to express variations between elements having the same GHC code and to provide a translation from GHC to Annex I habitat maps (Bunce et al. 2012b). The majority of environmental qualifiers are unlikely to change quickly over time, so that they can be updated during a targeted monitoring process.

Table 1 List of General Habitat Categories (GHC)
Full size table
Table 2 Expert rules based on spatial and temporal relations and spectral indices towards advanced LCCS to GHC translation
Full size table

LC/LU maps obtained from VHR space imagery, if based on LCCS taxonomy, can be used to provide holistic GHC maps through translation rules (Kosmidou et al. 2014) at lower costs. However, such systematic translation between two different domains (i.e., LC/LU and habitats) has been proven challenging, largely because of discrepancies between class definitions in the related taxonomies. In particular, only few one-to-one relationships and several one-to-many and many-to-many LCCS to GHC class transitions were observed, as systematically identified and described in Figs. 1, 2, 3, and 4 in Kosmidou et al. (2014). These class relationships were crisp and the authors suggested the use of additional information, including vegetation height measurements and land use information, to resolve the ambiguity in the one-to-many and many–to-many relationships.

Fig. 1
figure 1

a Le Cesine (red box) and Lago Salso (blue box) test sites location. RGB (Red-NIR-Blue) composition of: b Quickbird image acquired on June 2009, c Worldview-2 image acquired on October 2010 c for Le Cesine site. RGB (Red-NIR-Blue) composition of Worldview-2 image acquired on d June 2010 and e on February 2011 for Lago Salso site. (Color figure online)

Full size image
Fig. 2
figure 2

LC/LU map available and the list of LCCS classes with the alphanumeric code subset considered in LCCS to GHC translation evidenced by bold characters for: a Le Cesine and b Lago Salso

Full size image
Fig. 3
figure 3

a Diagram for adjacency rule used in the first LCCS to GHC mapping step to discriminate vegetated urban and natural GHC. b Diagram of rules for the identification of vegetated urban GHC sub-categories. The red arrows represent ambiguities that cannot be resolved with the elements at hand. (Color figure online)

Full size image
Fig. 4
figure 4

Woody GHC categories disambiguation diagram adopted: a when LIDAR data are used for plant height measurements; b when texture is used as proxy for plant height information. The red arrows represent ambiguities that cannot be resolved without LIDAR data. (Color figure online)

Full size image

There are many studies in the literature, both feature based and pixel based, focusing on the issue of map legend comparison and disambiguation of one-to-many and many-to-many relations between legends for change detection and validation applications (Herold and Schmullius 2004; Comber et al. 2004; Herold et al. 2006; Fritz and See 2008 Herold et al. 2009). However most of them mainly focus on the comparison of map legends concerning the same thematic product (e.g., land cover) produced by using different taxonomies and methods. Fritz and See (2008) compared the Global Land Cover 2000 map with the MODIS land cover data set by firstly reconciling legends through a look-up table and then calculating the fuzzy degree of overlap between legend classes, where there is an overlap, by considering user-driven or external sources-driven importance of different land cover types and finally producing a fuzzy agreement map. Regarding the quantitative comparison of taxonomies, several studies have recently contributed to define a frame where the interoperability between taxonomies is assessed by introducing some semantic similarity measures of the different classification schemes (Feng and Flewelling 2004; Ahlqvist 2004, 2005, 2008; Fritz and See 2008).

The present paper focuses on the translation rules between taxonomies belonging to two different thematic domains (i.e. LC/LU and habitats) and introduces expert knowledge in the form of crisp rules to resolve the ambiguities in LCCS to GHC translation. LCCS and GHC classes and their relationships are described in terms of spatial (e.g., topological adjacency, enclosure) and temporal relations (i.e., plant phenology and water coverage), as well as class specific spectral signatures in EO data. Applying expert knowledge with reflectance and ancillary data allows better solutions to classification problems (Comber et al. 2004, 2005). In addition, the possibility to formalize spatial–temporal relations (Goodchild et al. 2007; Pierkot et al. 2013) allows for the identification of inconsistencies between different datasets.

The purpose of the paper is twofold: (1) to resolve the ambiguities reported through the one-to-many and many-to-many relations but not dealt with in (Kosmidou et al. 2014); (2) the assessment of texture measurements as surrogate of plant height information in comparison to Light Detection And Ranging (LIDAR) data measurements. Once obtained from LC/LU maps, GHC maps can be translated into Annex I habitats through a specific key (Bunce et al. 2012c), as well as to other habitat LFs based taxonomies outside Europe. The effectiveness of the proposed approach is demonstrated for two coastal Natura 2000 sites in Italy as described below but can be extended to other sites.

The proposed approach is considered novel in the sense that it incorporates spatial reasoning (adjacency rules) and temporal relations as well as multi-source data integration (e.g., aerial LIDAR and VHR satellite data) for habitat mapping. To date, in the framework of Mapping and Assessment of Ecosystems and their Services (Maes et al. 2013) there has been little work done in automatic translation of LC/LU to habitat classes based on the use of VHR EO data (fine scale) and expert knowledge.

This work was conducted within the three-year BIO_SOS (www.biosos.eu) project, funded within the European Union FP7-SPACE third call. The project developed a pre-operational system for cost effective and timely monitoring of changes in land cover and habitats within and along the borders of protected areas in support to policy makers.

Study sites and EO data

Two study sites, belonging to the Natura 2000 network and located in Southern Italy, Le Cesine (SCI code IT9150032, SPA code IT9150014) and Lago Salso (SCI code IT9110005, SPA code IT9110038) are shown in Fig. 1a.

Le Cesine site is a coastal area, mainly composed by a complex of coastal lagoons and characterized by a system of ponds, marshes and wet meadows. The woody vegetation in this area is a mosaic of Pinus halepensis stands and different types of Mediterranean maquis and garrigues. The site covers an area of about 2,148 ha. The wetland is one of the most important sites in southern Italy in terms of biodiversity and is formed by two large coastal lagoons and various channels, marshes and wet grasslands. Reeds and sedges communities are prone to fires spreading from the adjacent agricultural areas and from arsons. Marine erosion has caused a progressive reduction of the sandbank, with frequent ruptures of the dune belt and a progressive salinization of the lagoons and the related environments. Due to water salinization, sedges, rushes and reeds communities have been replaced, over time, by halophytic scrub. In addition, the progressive erosion of the sand bank is determining the reduction and fragmentation of the typical dune habitat types. Finally, agricultural practices taking place both inside and outside the protected area are likely to lead to the disappearance of temporary ponds, which are priority habitats according to the EU Habitats Directive. Along the coastline, the Natura 2000 site is not only exposed to conversion to agricultural practices, but also to illegal urban development and to an increasing tourist pressure.

The Lago Salso site falls in the northern part of the Natura 2000 site “Zone Umide della Capitanata” (IT9110005) and within the Gargano National Park. It consists mainly of a wetland characterized by brackish and freshwater that, at the end of 19th century, covered an area of more than 4,000 ha, and then dramatically reduced as a result of massive land reclamation. Nowadays, the wetland is diminished to an area of about 540 ha and is characterized by reed, sedge and rush communities. The main pressures are represented by agriculture intensification and farm expansion in the surrounding areas. In the last few years, parts of such agricultural areas have been prone to progressive overflowing, as an action for the recovery of part of the original wetland. Northward, Lago Salso is adjacent to typical salt marshes, characterized by salty soils only periodically flooded and by halophilous annual and shrub vegetation.

Pre-existing validated LC/LU maps available for both sites were used for producing and validating GHC maps. The pre-existing LC/LU map (scale 1:5,000) available for Le Cesine site was produced in CORINE taxonomy and validated by in-field campaigns undertaken in 2008–2009. The map was firstly converted from CORINE taxonomy to LCCS taxonomy based on the report of panel of the Global Observation of Forest and Land Cover Dynamics (Herold and Schmullius 2004; Herold et al. 2006, 2009; Tomaselli et al. 2013). The LC/LU map for Lago Salso was produced on the basis of a pre-existing vegetation map and orthophoto interpretation and was based on the LCCS taxonomy (scale 1:5,000). The mapping was validated by in-field campaigns carried out in 2011–2012, to verify the presence and distribution of both artificial and natural and semi-natural habitat types. For natural and semi-natural habitat types, information on vegetation composition and structure was also collected, to validate the GHC translation.

For each coastal wetland site, two multi-temporal VHR images were acquired in two different seasons (e.g., summer and autumn) according to the phenology of the most representative plant communities and the water seasonality. The images were used to extract the spectral indices useful to disambiguate the one-to-many LCCS to GHC relations along with spatial and temporal relations, as described in subsequent sections. For Le Cesine site, a QuickBird image acquired on June 2009 and a WorldView-2 image acquired on October 2010 were selected (Fig. 1b, c). For Lago Salso site, two WorldView-2 images were acquired on June 2010 and February 2011 (Fig. 1d, e). All the images were ortho-rectified, co-registered and calibrated in Top of Atmosphere (TOA) Reflectance values. The spatial resolution (2 m) was compatible with LCCS map scale. Due to differences in the dates related to in-field validation campaigns of LC/LU maps and EO image acquisitions, only reference samples corresponding to no changed classes were selected through visual image inspection to validate the GHC products. LIDAR data were provided by the Italian Ministry of Environment for both sites. The acquisitions were made in spring 2009.

Methodology

The Figs. 1, 2, 3, and 4 in Kosmidou et al. (2014) evidenced that only 39 out of the total 87 LCCS to GHC mapping relations are one-to-one. In this paper, most discrepancies between the two taxonomies were resolved, focusing on the ones related to the manner the land use information is perceived (i.e., a field close to a building changes the categorization in the GHC taxonomy from natural trees and shrubs (TRS) into urban woody (TRE), whereas in LCCS taxonomy it remains as such), and to the differences in the height criteria used to classify the trees/shrubs categories as described hereafter.

Features used for the translation

Based on the field patches indicated by the input LCCS, an object-based translation approach is adopted. Expert knowledge is used in the proposed framework to describe LCCS and GHC classes and their relationships in terms of spatial and temporal relations and spectral indices extracted from EO data. The description is clearly scale dependent; this means that the methodology can be adapted to any scale if the description of classes is provided at the appropriate required scale. The GHC surveillance and monitoring technique (Bunce et al. 2011) was designed to map GHC through in-field campaigns. Consequently the rules and class descriptions are provided for fine scale mapping. VHR EO data and automatic techniques can help in GHC mapping across the full extent of protected study areas with a significant reduction in the costs related to in-field campaigns. To the authors’ knowledge, no fine-scale mapping of GHC has previously been performed from EO data. A number of key elements of LCCS to GHC map disambiguation are discussed hereafter at fine scale, including:

  1. (1)

    Spatial topological relations among classes. These were considered to both take into account the spatial onation of wetlands and decide if a vegetated object had to be considered as part of an urban area. More specifically, the adjacency relation to urban categories was used.

  2. (2)

    Temporal relations. To deal with the main phenological (vegetative) stages of both vegetation and water seasonality, two satellite images had to be introduced, corresponding, respectively, to the dry and wet seasons.

  3. (3)

    Geometric attributes for extracting land use information. The requirement of regularity in the shape of a field (object) was used. This attribute was extracted within the eCognition software using the Border Index feature (Trimble 2011). Based on a rectangular approximation, this feature describes how jagged a segment is. The smallest rectangle enclosing the image object is created and the Border Index is calculated as the ratio between the perimeter length of the object and the smallest enclosing rectangle. These attributes were used to discriminate trees belonging to TRS from WOC GHC categories.

  4. (4)

    Prior spectral knowledge on class spectral signature. Spectral indices were used as inputs to the translation algorithm, such as the Blue/NIR Ratio (BNR) (Morris and Dupigny-Giroux 2010) and Green/Red Ratio (GRR) (Ritchie et al. 2010) defined as:

$$BNR = {{\rho \left( {\lambda_{Blue} } \right)} \mathord{\left/ {\vphantom {{\rho \left( {\lambda_{Blue} } \right)} {\rho \left( {\lambda_{NIR} } \right)}}} \right. \kern-0pt} {\rho \left( {\lambda_{NIR} } \right)}} ,$$
(1)
$$GRR = {{\rho \left( {\lambda_{Green} } \right)} \mathord{\left/ {\vphantom {{\rho \left( {\lambda_{Green} } \right)} {\rho \left( {\lambda_{{\text{Re} d}} } \right)}}} \right. \kern-0pt} {\rho \left( {\lambda_{{\text{Re} d}} } \right)}} ,$$
(2)

where ρ is the TOA value in the bands considered and it is calculated as the mean value inside the object and λBlue, λRed and λNIR, indicate the image bands lying in the blue, red and near-infrared part of the electromagnetic spectrum. The bands ratios used in GRR and BNR indices perform well in discriminating vegetated and water covered pixels, respectively.

  1. (5)

    LIDAR derived measurements. Due to the existing inconsistency in the two taxonomies between the definitions of plant height (Kosmidou et al. 2014), the information of plant height was directly extracted from the available LIDAR dataset and not taken from the height code element (if any) associated to the LCCS label. This information is useful for the disambiguation of Phanerophytes (Raunkiaer’s LF including perennial woody plants with buds located more than 30 cm above the soil surface, typically trees and large shrubs) which are divided in 5 GHC (i.e., LPH, MPH, TPH, FPH and GPH) categories on the basis of their height. Up to the date this work was conducted, 43.44 % of the Italian territory (131.118 km2) has been covered by LIDAR acquisitions (Costabile et al. 2013). The LIDAR acquisitions were made in May 2009 when plant growth and biomass were seasonally high to minimize laser penetration of the vegetation canopy. The dataset used consists of Digital Terrain Model (DTM), Digital Surface Model (DSM) (from the first and the last pulse) and cloud points. The DTM and DSM are provided with a spatial resolution of 2 m × 2 m and a vertical accuracy of ± 15 cm. The height of each object in the LCCS map was obtained as the difference between the DSM (from the first pulse) and the DTM.

  2. (6)

    Contextual features. The first order entropy (occurrence measure) texture (Anys et al. 1994), computed in the green band of the biomass peak image, was used to provide plant height information to solve problems when LIDAR data are not available. To compute first-order statistics at a given scale of interest, the pixel values within a moving window were used and the measurement was assigned to the central pixel. Two first-order texture measures (entropy and variance) were selected based on their established ability to characterize vegetation structure (Wuest and Zhang 2009; Wood et al. 2012; Petrou et al. 2012. These were then calculated using imagery acquired at the peak of the biomass. Different size windows were used to account for the heterogeneity of pixel values in relation to vegetation structure and height. The entropy measure, calculated with a kernel window 3 × 3 pixels sized on the image green band was experimentally identified as the most appropriate to discriminate plant height into two main classes in the analysis. The comparison of the results with the ones from LIDAR data was carried out in terms of the overall accuracy (OA) of the GHC maps.

The translation algorithm receives as inputs: an LCCS map, the multi-seasonal EO images and the set of translation rules. As well known (Di Gregorio and Jansen 2005; Tomaselli et al. 2013), any LCCS class in the map is described by a unique numerical code and a standard name. The alphanumeric elements of the LCCS code correspond to either specific classifiers (i.e., a set of independent diagnostic criteria used to identify each class) or environmental and technical attributes (Di Gregorio and Jansen 2005) useful for class identification. However, some qualifiers in the LCCS code are not related to the GHC semantic class description and consequently are not useful for LCCS to GHC mapping. For this reason, the algorithm considers only a specific sub-set of the alphanumeric elements in the code. More specifically, life form, leaf type and leaf phenology elements (classifiers) are selected in the code of both natural terrestrial vegetated (i.e., A12 category) and natural aquatic vegetated classes (i.e., A24). Regarding the cultivated classes (i.e., A11), additional codes related to the technical attributes (e.g., orchards and plantations) are considered; only the life form component is considered for cultivated aquatic classes (i.e., A23); surface aspect is considered for both primarily non vegetated artificial surfaces (i.e. B15) and bare areas (i.e. B16), whereas physical status is used for artificial water bodies (i.e., B27) and inland waterbodies (i.e. B28).

Figure 2a and b show the input LCCS maps available for Le Cesine and Lago Salso sites, respectively. The Figures include the list of LCCS classes. The elements of the code used for LCCS to GHC translation are evidenced in bold.

Expert rules

For translating LCCS classes into GHC, the algorithm applies a set of expert rules listed in Table 3 with the feature used in the text (second column), the feature/spectral index condition (third column) and the description of the rule conditions (second column). Figures 3, 4, and 5 depict in a concise graphical manner the rules for the LCCS to GHC translation. The rules are organized per group of output GHC categories (e.g., URB, TRS, HER), as described hereafter. In these figures, the LCCS input classes satisfying the same rules are listed in grey rectangular boxes within the diagram of each figure. Such grey boxes are in all figures indicating that LCCS to GHC mapping involve not only one (LCCS)-to-many (GHC) relations (only one LCCS class in the grey box) but also many (LCCS)-to-many (GHC) relations. To identify the final GHC class among the many GHC outputs (with these evidenced in the central line of each figure) corresponding to the same input LCCS class (or group of LCCS input classes when many-to-many LCCS to GHC relations hold), black nodes are introduced in the diagrams to represent alternative decision paths, based on the satisfaction of the rule conditions (see last column of Table 3). The grey arrows represent well defined disambiguation rules, whereas red arrows indicate the presence of ambiguities not resolved with the remote sensing techniques alone.

Table 3 LCCS classes and the corresponding feasible GHC classes located in Le Cesine and Lago Salso sites, based on the work in Kosmidou et al. 2014
Full size table
Fig. 5
figure 5

Herbaceous categories disambiguation. Red links evidence ambiguities still not resolved without in-field data

Full size image

Methodology processing chain

The output GHC categories are identified according to the following sequence:

  1. (1)

    Firstly, Artificial Urban (ART) and Urban Non-Vegetated (NON) categories are identified. This step just consists of re-labelling the LCCS objects (classes) originally labelled as buildings, roads, car parks, etc., since these classes correspond to the semantic description of ART category. All non-vegetated and non-artificial objects within an urban area in the LCCS map (identified with the first level LCCS classifiers with codes B15, B16 and B27), including urban water bodies, are re-labelled as NON (Bunce et al. 2011).

  2. (2)

    Then, GHC Urban vegetated categories are identified as correspondent to the vegetated classes (objects) in the input LCCS map adjacent to ART or NON objects for more than 50 % of their boundaries. In Fig. 3a, the discrimination between GHC Urban and the remaining GHC vegetated classes (i.e., both natural and cultivated) is shown. The discrimination among the vegetated URB sub-categories, which include Urban vegetables (VEG), Urban herbaceous (GRA) and (TRE) categories, is then based on plant height measurements from either LIDAR data or texture-based information rules reported in Table 3 (last four lines of the 5th column). Fig. 3b provides more details.

  3. (3)

    Plant height information extracted from LIDAR data, see Fig. 4a, or alternately from first order entropy texture measurements, see Fig. 4b, is engaged.

  4. (4)

    Within the TRS category, Coniferous (CON), Deciduous (DEC), Evergreen (EVR), Non-Leafy Evergreen (NLE) and Summer Deciduous (SUM) are finally discriminated within the output layers of the previous step 3), mainly based on phenological class properties exploiting spectral indices (i.e., P1 and P2 in Table 3) extracted from multi-temporal EO images.

  5. (5)

    Finally, Herbaceous (HER), Herbaceous Wetland (HER), Cultivated (CUL) and Sparse Vegetation (SPV) categories are obtained from the remaining vegetated LCCS objects, according to geometric features related to the shape of fields and temporal relations (Fig. 5). In particular, for Wetland Herbaceous categories (HER), expert knowledge based on spatial topological rules related to habitats spatial pattern has been used. It is well known that in coastal and wetland environments plant communities tend to reside in more or less regular belts, with specific zonation and that water regime is a major determinant of plant zonation patterns in wetlands. Water regime can be described by depth, duration, frequency, timing of flooded and dry phases in a wetland (Casanova and Brock 2000). Studying the two wetland sites, the following rule was identified: (i) submerged hydrophytes (SHY: plants that grow in aquatic conditions with the whole plant in water) are located in the inner part of a lagoon, flooded during the whole year; (ii) emergent hydrophytes (EHY: plants that grow in aquatic conditions and have emergent shoots out of the water) are arranged in an external (and adjacent) belt, flooded during part of the year; and (iii) helophytes (HEL: helophytes, plants that grow in waterlogged conditions) tend to form a further external belt, on soils flooded for short periods or only occasionally and waterlogged in the remaining part of the year. A schematic representation of the zonation rule would appear as: SHY → EHY → HEL.

As an example, the LCCS code “A24.A2.A6.E6” will give as output the GHC class HEL if the considered object simultaneously satisfies the following conditions (see Fig. 5): (i) Adjacency to ART or NON: “less than 50 %”; (ii) BNR of the summer image: “less than 1”; (iii) adjacency to SHY class: “0 %”.

Results and discussion

The lists of LCCS classes characterizing the two study sites are reported in Table 3 for both Le Cesine site and Lago Salso site. For each LCCS class, the first column of Table 3 includes a binary code of two elements indicating whether the class is present (1) or absent (0) in Le Cesine (first digit) and Lago Salso (second digit) sites; the second and third columns report the dichotomous (e.g. A11, A12, A24, B15, B16) and hierarchical LCCS code components; the last column includes possible GHC output classes corresponding to the same input LCCS class according to the framework introduced by Kosmidou et al. (2014). As evident in Table 3, the previously proposed framework results in many one-to-many and many-to-many relations in LC/LU to habitat translation to be resolved through the proposed rules of Table 2.

Figures 6 and 7 show the output GHC maps obtained from the pre-existing input LCCS maps applying the disambiguation rules in Table 2, both with (6a and 7a) and without (6b and 7b) LIDAR data, for Le Cesine and Lago Salso sites, respectively. For Le Cesine site, when comparing Fig. 6a with Fig. 8, which represents the output when applying the generic LCCS to GHC translation framework described in Kosmidou et al. (2014), it becomes evident that the majority of the ambiguity is resolved. Most of the resulting classes after applying the approach proposed in this paper are single GHC classes: an adequate number of multiple classes in Fig. 8 were resolved in Fig. 6a to single ones with the use of the additional features (spatial relations and temporal relations, context features and geometric properties) and expert translation rules embodied in the approach. This constitutes the main contribution of this study to the advancement of automatic complex translation between legends and domains. As an example, LCCS class A12.A1.A4.A10.B3.D1.E2_B9 in Le Cesine site, representing a “broadleaved deciduous medium–high closed shrubland,” may correspond to either MPH/DEC, TPH/DEC (deciduous medium or tall phanerophytes, respectively) or TRE (urban trees), as seen in Table 2 and in Fig. 8. Inclusion of the ‘adjacency to ART or NON urban categories’ rule helped in discriminating the urban TRE classes from the natural vegetation, i.e. MPH/DEC and TPH/DEC. Information on height provided by the LIDAR data made possible the further discrimination between the MPH/DEC and TPH/DEC categories, being 0.6–2 m and 2–5 high, respectively. As a note, the ‘adjacency to ART or NON’ rule intensely enhanced the class discrimination task, by particularly resolving ambiguities between artificial or managed and natural vegetated areas.

Fig. 6
figure 6

a GHC map produced for Le Cesine site: with the use of LIDAR data. b. GHC map produced for Le Cesine site: with the use of texture feature

Full size image
Fig. 7
figure 7

a GHC map produced for Lago Salso site: with the use of LIDAR data. b GHC map produced for Lago Salso site: with the use of texture feature

Full size image
Fig. 8
figure 8

The GHC map deduced from the pre-existing LCCS map, available for Le Cesine site, without the disambiguation rules proposed in the present paper and published as supplementary material by Kosmidou et al. (2014). The output GHC classes include many ambiguities (multiple GHC classes)

Full size image

To validate the maps, GHC reference samples were selected through visual inspection of the multi-seasonal VHR images and in situ campaigns. The samples were collected only in no-changed areas, with these identified as the ones characterized by the same LCCS class both on the ground at the date of survey, in the pre-existing LCCS map, and in the images used as input to the discrimination procedure. The reason was to avoid classification errors due to differences in the label of reference samples at different dates. For each site, classification confusion matrices were produced for the GHC maps obtained with and without LIDAR. The columns in the matrices represent ground truth classes and the rows the classification results.

Table 4 shows the confusion matrix obtained with LIDAR data for Le Cesine. All results are affected by the quality of the input LCCS map and also by the error in the LIDAR data. In addition, due to the object based classification approach, when trees and shrubs objects (TRS) are considered, the average plant height value from LIDAR includes also the height information of all background vegetated pixels (if any) contained in the object. This might generate an underestimation of the average plants height and, consequently, an incorrect discrimination of the TRS super-category into the sub-categories (e.g., LPH, MPH) associated to different height ranges. As Table 4 reports, an Overall Classification (OA) accuracy equal to 69.9 % with an error of 5.3 % is achieved. An additional problem might be related to the fact that, for HER (herbaceous wetland) object identification, the adjacency rule is effective when habitats are in an ideal conservation status; if such condition is not verified misclassifications may occur. In fact, relationships between flooding duration and plant distribution only partially explain the plant community zonation, because other micro-ecological factors may influence the vegetation pattern. Above all, unwise human activities usually produce habitat reduction, fragmentation and isolation, with deep changes in the spatial pattern of plant communities.

Table 4 Confusion matrix for Le Cesine site: with the use of LIDAR data, all plant height ranges are considered
Full size table

The results highlight how a set of rules, including spatial topological and temporal rules, can disambiguate one-to-many and many-to-many LC/LU to GHC relationships. Concerning wetland spatial patterns, the spatial topological rules used to resolve ambiguities, even though resulting from a generalization of some main ecological gradients, should be carefully evaluated for each site in analysis, in consideration of site-specific ecological conditions and processes. As an example, the salinization occurring in Le Cesine site in the last few years is determining a modification of plant communities and their spatial pattern, leading to an alteration of the spatial rules described above. So, new rules should be applied where the spatial pattern has been altered. In fact, plant communities of salt marshes are spatially arranged according to flooding period, soil salinity and moisture (Rogel et al. 2000; Molina et al. 2003). Also in this case, it is possible to generalize: submerged hydrophytes (SHY) in the inner part of the lagoon; therophytes (THE: annual plants; in this case pioneer communities rich in glasswort and other succulent plants of salt marshes, usually appearing during the late summer) forming narrow stripes or belts surrounding the water bodies and only in the summer period; and shrubby chamaephytes (SCH; in this case, succulent dwarf shrubs typical of salt marshes) arranged in external belts. Then, we could schematize such a zonation: SHY → THE → SCH. Future researches will test these rules in different contexts, in order to verify their effectiveness. Concerning temporal relations, expert knowledge on class water coverage seasonality was used and combined with topological information to solve some many-to-many relations evidenced in Fig. 5. As an example, the four LC/LU aquatic classes (A24) grouped in the second pink box of last line in Fig. 5 (i.e., A24/A2.A6.E6, A24/A2.A6.E7, A24/A2.A5.A6 and A24/A2.A5.E7) can correspond to three possible GHC depending on the presence of water, which was measured through spectral indices, in the two bi-seasonal satellite images considered, and also the adjacency to SHY category. According to the rules and legends in Table 3, the final GHC class can be: (i) SHY, if W1 holds or (ii) EHY, if W2 and S1 hold or (iii) HEL, if W2 and S2 hold. Temporal phenological information was used to solve discrepancy in LCCS and GHC terminology. Even though LCCS uses codes to identify natural woody trees broadleaved as evergreen (E1) and deciduous (E2) (i.e., full codes A12/A1.A3.D1 and A12/A1.A3.E2) it is not possible to identify GHC classes corresponding to summer or winter deciduous. To do that, the spectral Green/Red Ratio (GRR) values from the Post Biomass Peak (PoBP) image are considered according to the rule in Table 3. As evidenced in Figs. 3, 4, and 5 by red lines, there are ambiguities that cannot be resolved by using temporal, spatial and spectral information from EO data. As an example, in Fig. 5, the one-to-many relations corresponding to the association of natural herbaceous forbs (A12/A2.A5) to three GHC classes (i.e., LHE, HCH and GEO) cannot be solved from EO data observations because the discrimination is related to plant morphological and structural properties. In such case in-field inspection is mandatory for final GHC identification. As an additional example, in Fig. 4b it is not possible to resolve the ambiguities related to accurate plant height measurements, because the texture feature was able to discriminate only two coarse groups of plant height. When height information is measured through texture measurements, only two macro height clusters can be detected representing low and high textured vegetation. This means that it is possible to select a threshold in the texture image histogram for separating two macro clusters. In the analysis, to identify the plant height ranges corresponding to such clusters, the reference samples were grouped in two different ways to validate the GHC map and obtain the OA value of the associated confusion matrix. The first grouping merged all reference samples labelled as DCH, SCH, LPH and MPH (respectively, dwarf and shrubby chamaephytes, low and mid phanerophytes), corresponding to trees and shrubs with height <2 m, into one cluster and TPH, FPH (tall and forest phanerophytes) and GPH (Giant phanerophytes over 40 m) corresponding to trees and shrubs with height >2 m into a second cluster. With the first grouping, the OA was equal to 63 % with an error of 5.5 %. The second grouping merged only DCH, SCH and LPH sub-categories corresponding to trees and shrubs with height ≤0.5 m in the first cluster and MPH, TPH, FPH and GPH corresponding to height >0.5 m in the second cluster. The corresponding OA reached 78.4 % with an error of 5 %, as reported in Table 5. The latter result is comparable with the OA value (79.1 %, with an error of 4.7 %) obtained with LIDAR data when the same grouping of reference sample was adopted. As expected, this result is higher than the one in Table 4 (i.e. 69.9 %) when the full set of plant height values was considered and confirms the feasibility to use texture for height discrimination. Based on the experience gained on Le Cesine site, the translation of LCCS to GHC maps was carried out also for the second site, Lago Salso. The OA value for the second site with LIDAR was equal to 66.1 % with an error of 12.7 %. By grouping the reference samples as in the previous case (i.e., ≤0.5 and >0.5 m), the OA value obtained with texture and LIDAR resulted equal to 81.4 % with error 9.9 % and 77.9 % with error 10.6 %, respectively. However, the error values were higher than the ones of Le Cesine. This is due to the different number of reference samples available for Lago Salso and Le Cesine, corresponding to the 22 % and 10 % of the total number of classified objects, respectively.

Table 5 Confusion matrix for Le Cesine site: based on texture measurements once reference samples are grouped in two clusters (i.e., ≤ 0.5 and > 0.5 m.)
Full size table

Conclusions

The findings of this paper confirm the feasibility of automatic procedures for translating LC/LU maps into GHC maps and resolve the one-to-many and many-to-many relations identified in a previous paper (Kosmidou et al. 2014), through the combined use of EO data and expert knowledge rules to combine temporal, spatial and geometrical features of both LCCS classes and GHC classes. In addition, ambiguities in plant height definition are directly solved by using LIDAR data. However, when LIDAR is considered to be too costly or is not available at all, entropy texture feature from the image green band can help to separate plant heights into two coarse height groups (i.e.,≤0.5 and ≥0.5 m). As a result, based on expert knowledge, a number of multiple classes can be resolved to single ones and the method allows for the mapping of large areas without the need of any in-field ground truth for training the classifiers. This constitutes the main contribution of this study to the advancement of automatic complex translation between legends and domains and to the mapping of inaccessible areas where the collection of in-field data (ground truth) is impractical.

The methodology can be applied to any LC/LU validated map in LCCS taxonomy or translated in such taxonomy (e.g., from CORINE). Consequently, GHC maps can be produced also in the past if archive EO multi-temporal images (two seasons) close to the date of LC map validation can be found. To the authors’ knowledge, no fine-scale GHC mapping has previously been performed using EO data. So far, GHC maps have been only produced by in-field campaigns with the help of in-field computers to register GHC labels for each patch within a specific 1 km × 1 km grid. Therefore, the main contribution of the method described in this paper is the possibility to automatically provide GHC maps from validated VHR LCCS maps and EO data (at 2 m spatial resolution) for areas as large as needed and without the class ambiguities discussed but not resolved in the previous work (Kosmidou et al. 2014). Such extended GHC maps can be used to determine areas to which specific in-field campaigns should concentrate for locating the vegetation plots (Bunce et al. 2011) and collecting in situ data on species and environmental attributes. The in-field campaigns can contribute to the refinement of the remaining LCCS to GHC translation ambiguities, such as the one-to-many relations corresponding to the association of natural herbaceous forbs (A12/A2.A5) to three GHC classes (i.e., LHE, HCH and GEO). These often cannot be solved from EO data observations because their discrimination is related to specific plant morphological and structural properties. The fuzzification of the proposed rules can further improve their generalization (Petrou et al. 2013). Within the BIO_SOS project (www.biosos.eu), the proposed rule based disambiguation approach has been applied to other Natura 2000 sites in different areas (e.g. Wales, The Netherlands, Greece).

To conclude, the proposed framework is a very promising tool for the automatic translation of LCCS to GHC maps, offering a service for habitat mapping with consistent cost reduction of in-field campaigns in the age of rapid biodiversity decline. The GHC maps can be regularly updated to detect habitat changes for providing support to decision makers.