Mapping salt-marsh vegetation by multispectral and hyperspectral remote sensing

Mapping salt-marsh vegetation by multispectral and hyperspectral remote sensing

Remote Sensing of Environment 105 (2006) 54 – 67 www.elsevier.com/locate/rse Mapping salt-marsh vegetation by multispectral and hyperspectral remote ...

2MB Sizes 0 Downloads 46 Views

Remote Sensing of Environment 105 (2006) 54 – 67 www.elsevier.com/locate/rse

Mapping salt-marsh vegetation by multispectral and hyperspectral remote sensing Enrica Belluco a , Monica Camuffo b , Sergio Ferrari a , Lorenza Modenese b , Sonia Silvestri c , Alessandro Marani b , Marco Marani a,⁎ a

Dipartimento IMAGE and International Centre for Hydrology ”Dino Tonini”, University of Padova, via Loredan 20, I-35131 Padova, Italy b Environmental Science Department, University of Venice, Dorsoduro 2137, I-30123 Venice, Italy c Servizio Informativo-CVN-Magistrato alle Acque di Venezia, San Marco 2803, I-30124 Venice, Italy Received 12 October 2005; received in revised form 8 June 2006; accepted 8 June 2006

Abstract Tidal marshes are characterized by complex patterns both in their geomorphic and ecological features. Such patterns arise through the elaboration of a network structure driven by the tidal forcing and through the interaction between hydrodynamical, geophysical and ecological components (chiefly vegetation). Intertidal morphological and ecological structures possess characteristic extent (order of kilometers) and smallscale features (down to tens of centimeters) which are not simultaneously accessible through field observations, thus making remote sensing a necessary observation tool. This paper describes a set of remote sensing observations from several satellite and airborne platforms, the collection of concurrent ground reference data and the vegetation distributions that may be inferred from them, with specific application to the Lagoon of Venice (Italy). The data set comprises ROSIS, CASI, MIVIS, IKONOS and QuickBird acquisitions, which cover a wide range of spatial and spectral resolutions. We show that spatially-detailed and quantitatively reliable vegetation maps may be derived from remote sensing in tidal environments through unsupervised (K-means) and supervised algorithms (Maximum Likelihood and Spectral Angle Mapper). We find that, for the objective of intertidal vegetation classification, hyperspectral data contain largely redundant information. This in particular implies that a reduction of the spectral features is required for the application of the Maximum Likelihood classifier. A large number of experiments with different feature extraction/selection algorithms show that the use of four bands derived from Maximum Noise Fraction transforms and four RGBI broad bands obtained by spectral averaging yield very similar classification performances. The classifications from hyperspectral data are somewhat superior to those from multispectral data, but the close performance and the results of the features reduction experiments show that spatial resolution affects classification accuracy much more importantly than spectral resolution. Monitoring schemes of tidal environment vegetation may thus be based on high-resolution satellite acquisitions accompanied by systematic ancillary field observations at a relatively limited number of reference sites, with practical consequences of some relevance. © 2006 Elsevier Inc. All rights reserved. Keywords: Salt marshes; Vegetation; Hyperspectral data; Multispectral data

1. Introduction Coastal intertidal areas are transition zones between marine and terrestrial systems characterized by high biodiversity and primary production, which play a central role in mediating sea action on the coast, the effects of floods on estuarine areas, and in buffering nutrient fluxes from the land.

⁎ Corresponding author. Tel.: +39 49 8275449; fax: +39 49 8275446. E-mail address: [email protected] (M. Marani). 0034-4257/$ - see front matter © 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2006.06.006

The state and evolutionary trends of such systems are the result of complex interactions among hydrodynamic, ecological, hydrological and sediment transport processes, forced by tidal fluctuations (e.g. Marani et al., 2004). A key element in intertidal system dynamics is halophytic vegetation, i.e. plants which have evolved to develop and reproduce in highly hypoxic and hypersaline soils (e.g. Adam, 1990; Cronk & Fennessy, 2001). Halophytes colonize salt marshes, areas located above the mean sea level but below the mean high water level, and thus flooded according to local tidal periodicities. Salt-marsh vegetation is largely responsible for the stability of these areas,

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

through feedbacks involving hydrodynamic and sediment circulations. Plant roots, in fact, stabilize the soil, while the aboveground biomass importantly reduces water flow velocity and dampens wind-induced waves, thus effectively impeding sediment resuspension and erosion (e.g. Pethick, 1984; Leonard & Luther, 1995). Furthermore, the biomass produced by halophytes often constitutes the largest contribution to the local incoming flux of soil and thus is crucial in allowing marsh accretion to keep pace with soil compaction, subsidence and sealevel rise (Cahoon & Reed, 1995; Day et al., 1999; Reed, 2002). On the other hand, the net effect of deposition and erosion processes determines local topography, which, together with tidal forcing and subsurface water flow, in turn determines the edaphic conditions constraining vegetation development and selecting halophytic species (e.g. Chapman, 1964; Beeftink, 1977; Bockelmann et al., 2002; Silvestri & Marani, 2004; Ursino et al., 2004; Silvestri et al., 2005). This complex ecogeomorphological feedback cycle induces a spatial distribution of halophytes which is characteristically organized in patches of a single species, or of a typical species association: a phenomenon known as zonation (Chapman, 1964; Pignatti, 1966; Silvestri & Marani, 2004). Because of its key dynamic role, its intrinsic ecological importance and its value as an ecogeomorphological indicator, it is not surprising that halophytic vegetation is of central interest in studies concerning intertidal processes and in management schemes attempting to counteract coastal squeeze phenomena so relevant throughout the world (related to the common situation in which the coastal margin is squeezed between a usually artificial fixed landward boundary and the rising sea level; e.g. Cahoon et al., 1995; Bernhardt & Koch, 2003; Cox et al., 2003; Hughes & Paramor, 2004;Wolters et al., 2005). Quantitative, accurate and repeatable observations of vegetation space–time distributions are therefore of self-evident importance. Such observations must cover spatial scales ranging between tens of centimeters and some kilometers, and temporal scales from a single season to several years. Remote sensing is thus ideally suited for the task, and there has recently been a growing interest in the application of remote sensing methods to halophytic vegetation mapping (e.g. Dale et al., 1986; Johnston & Barson, 1993; Donoghue et al., 1994; Eastwood et al., 1997; Smith et al., 1998; Thomson et al., 1998a,b, 2004; Munyati, 2000; Silvestri et al., 2002, 2003; Shuman & Ambrose, 2003; Marani et al., 2003, 2004, in press). Previous approaches to remote sensing mapping of salt-marsh vegetation often focus on study sites exhibiting relatively limited species diversity or attempt to discriminate just broad vegetation communities, each including several species (e.g. ‘pioneer species’, ‘creek margin vegetation’, etc., e.g. see Smith et al., 1998; Brown, 2004). Much of the current literature furthermore deals with intertidal environments characterized by relatively large-scale vegetation structures (e.g. typical of meso- to macrotidal environments), with patch sizes comparable to the usual satellite sensor resolution (10 m–30 m) (e.g. Ramsey & Laine, 1997). Other studies use ground reference data which are limited in terms of their total amount or are characterized by a rather small areal extension as compared to sensor resolution (e.g. Schmidt

55

et al., 2004). Under these conditions reference data for classifier training and validation are not suitable for a conclusive assessment of the accuracy of the resulting vegetation maps. In fact, to our knowledge, no previous work has quantitatively addressed, on the basis of extensive ground reference data, the accuracy of classifications of highly spatially heterogeneous intertidal vegetation using a large set of last-generation airborne and satellite sensors with up to 1 m resolution. Moreover, previous remote sensing studies of intertidal vegetation deal with single ‘snapshot’ acquisitions and thus do not allow the appreciation of mapping reliability on scenes acquired at different times and seasons, under different atmospheric conditions, with different tidal levels and in the presence of different vegetation development stages. In this framework, we explore the possibility of reliably discriminating different salt-marsh vegetation species by the use of several multispectral and hyperspectral data sets acquired within the Venice Lagoon (Italy) and of concurrent detailed field observations. The objective is to set a quantitative context for vegetation mapping applications in tidal environments and to test the performance of widely used classification procedures. This is relevant both as an assessment of the separability of the information classes of interest and as a reference for studying and monitoring schemes relying on widely applied classification tools. 2. Study sites and data description The Lagoon of Venice (Fig. 1), located in north–east Italy, is a water body with a surface of about 550 km2 and an average depth of approximately 1.1 m, characterized by a semidiurnal tide with a range of about 1.4 m. The Lagoon is connected to the Adriatic sea by three inlets and receives freshwater inputs from a few tributaries, contributing a quite small water flux, but a relatively large associated input of solutes (e.g. nutrients from agricultural areas) (e.g. Consorzio Venezia Nuova, 2005). Because of major river diversions, performed between the 14th and the 17th centuries, and the construction of concrete jetties at the inlets, completed at the beginning of the 20th century, the net sediment balance of the Venice Lagoon is currently negative. The Lagoon is thus experiencing a transformation toward a marine system, with important environmental implications. Even though the general trend is quite clear, the mechanisms shaping the landscape of the Venice Lagoon, as for many other tidal environments (e.g., Bird, 1985; Finkl, 1996; Friedrichs & Perry, 2001; Leatherman, 2003; Zhang et al., 2004), are not well understood and, while some of its parts experience intense erosion, some salt marshes are actually accreting. The Venice Lagoon is thus a representative example of a relevant intertidal environment subject to complex changes, which requires a better understanding of its ecological and morphological dynamics and reliable and easy-to-implement monitoring schemes. 2.1. Study sites Fig. 1 shows the location, within the Venice Lagoon, of the study site analyzed here, the San Felice salt marsh. This site is extracted from a larger set of study marshes (TIDE, 2005)

56

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

Fig. 1. The San Felice salt marsh in the northern part of the Venice Lagoon (Italy).

located in the northern part of the Lagoon, which are characterized by ‘healthy’ halophytic vegetation (i.e. not showing evidence of dieback), relatively small rates of general erosion and, in some cases, by local accretion. The San Felice salt marsh is located about 2 km from the northern inlet of the Lagoon (Lido inlet), along the homonymous channel, and was frequently and densely surveyed during a five-year study period of reference (2000–2004). Its elevation ranges from about 0.01 m above the mean sea level (a.m.s.l.) to 0.68 m a.m.s.l. (with an average of 0.26 m a.m.s.l.), and its area is mainly colonized by four halophytic species, Spartina maritima (hereafter ‘Spartina’), Limonium narbonense (hereafter ‘Limonium’), Sarcocornia fruticosa (hereafter ‘Sarcocornia’) and Juncus spp. (hereafter ‘Juncus’) (nomenclature follows Caniglia et al. (1997)). These same species also dominate, to varying degrees, the remaining study marshes. Interestingly, Spartina was almost entirely displaced by Salicornia veneta (hereafter ‘Salicornia’) in 2004 at all study sites, showing that halophytic vegetation distribution may significantly vary over relatively short time scales, and the interest of reliable quantitative remote sensing monitoring schemes to characterize vegetation temporal changes. 2.2. Remote sensing observations We present and analyze data acquired during the period 2000–2004 in the Venice Lagoon (mostly within the TIDE project, TIDE, 2005), comprising observations from several multispectral and hyperspectral sensors, and a set of field ancillary observations systematically acquired within about 1 week from each remote sensing acquisition. The objective is to determine the optimal field procedures and sensor configurations with related quantitative assessments of the resulting accuracy. To minimize directional reflectance effects all hyperspectral flights were performed along directions in the principal plane and under clear sky conditions (except one flight line of the ROSIS acquisition, which was cloudy) within 2 h of solar noon.

Flight planning also attempted to reconcile these constraints with the requirement that salt marshes should not be flooded during acquisitions. These were thus planned for periods of relatively low tides. A final, but clearly fundamental, constraint was given by weather conditions and thus several potential time windows were selected satisfying the above constraints to suitable degrees. The remote sensing data set available is comprised of: • ROSIS (Reflective Optics System Imaging Spectrometer, of DLR, data acquired within the HySens project). ROSIS acquires 115 spectral bands in the Visible (VIS) and Near Infrared (NIR) part of the spectrum (spectral range: 415.5–875.5 nm with a band width of 4 nm). The acquisition was performed on 8 July 2000 with a ground resolution of 1 m. • CASI (Compact Airborne Spectrographic Imager). This sensor allows the selection of a number of bands between 400 nm and 950 nm as a function of bandwidths and flight altitude. Fifteen bands were selected in the VIS and NIR (band intervals in nm are: 437.15–445.65; 484.4–494.8; 542.7–554.1; 614.65–623.35; 631.65–640.35; 648.25–655.95; 662.5–668.3; 677.15–682.05; 688.65–697.35; 705.75–710.65; 745.85–750.75; 757.25–762.15; 774–781.8; 817.15–823.05; 860.9–869.7) on the basis of previous literature (Smith et al., 1998; Thomson et al., 1998b) and of the observed characteristics of vegetation spectra collected during previous field campaigns using a handheld spectrometer (e.g. TIDE, 2005). Two separate CASI acquisitions over the study site were performed (29 September 2002 and 8 February 2003), with a ground resolution of 1.3 m. The data acquired in September 2002 are analyzed here. • MIVIS (Multispectral Infrared and Visible Imaging Spectrometer). This sensor covers the ranges: 433 nm–833 nm (20 bands, 20 nm resolution); 1150 nm–1550 nm (8 spectral bands, 50 nm resolution); 2000 nm–2500 nm (64 spectral bands, 8 nm resolution); 8200 nm–12,700 nm (10 bands, 400 to 500 nm resolution). Five MIVIS acquisitions were

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

performed over the study site (20 July 2002, 29 September 2002, 5 July 2003, 29 March 2004, and 30 June 2004), with a 2.6 m ground resolution (for all bands). The present paper analyzes the data acquired in July 2003 and June 2004 and considers all the spectral information available in the reflective as well as in the emissive part of the spectrum. The thermal infrared bands can in particular convey useful information to discriminate halophytic vegetation species, possibly due to differences in canopy structure, exposed soil and soil water content. • IKONOS. The IKONOS data used (made available by the Venice Water Authority, ‘Magistrato alle Acque’) comprise 4 bands (450–520; 520–600; 630–690; 760–900 nm) in the visible and near infrared part of the spectrum. The original IKONOS resolution is 4 m, but the data used here had first been downscaled to a 1 m resolution on the basis of the 1 mresolution panchromatic band (pan-sharpening, e.g. Zhang, 2002). The IKONOS data were acquired on 26 June 2001. • QuickBird. QuickBird 4-band data (450–520; 520–600; 630–690; 760–900 nm) have a ground resolution of 2.88 m. Six QuickBird acquisitions were available over the selected study site: 16 May 2002, 10 February 2003, 25 July 2003, 10 October 2003, 13 September 2004, and 8 June 2005. The present paper analyzes the data acquired in July 2003. The flight or acquisition timing and general characteristics of the data analyzed in the present paper are summarized in Table 1. The period of the year was an important factor in the selection of the acquisitions to be analyzed in detail, due to the fact that halophytic species bloom during the summer. Initial classification experiments showed, in fact, that intertidal species are most easily discriminated when at their full development stage. Another important variable was the tidal level, which is indicated in Table 1 for a reference location (Punta della Salute, in the city center). 2.3. Field observations Extensive sets of field observations were acquired concurrently with remote sensing campaigns to provide accurately located and quantitative ground reference data. For each cover

57

type of interest, we randomly selected ground reference areas (Regions of Interest, ROIs) with an extent larger than several times the pixel size throughout the acquisition scene. This procedure was rather difficult to implement because of the smallest scale of variability of species spatial distribution (order of tens of centimeters) and to the often small size of some vegetation patches. However, in all cases (i.e. for all campaigns and all sensors) the size of the ROIs was at least 3 pixels by 3 pixels, even for the coarsest multispectral data. The boundaries of the ROIs were accurately delimited using either differential GPS (DGPS) or a laser theodolite (minimum accuracy of ±1 cm), while their superficial properties were quantitatively characterized in terms of relative cover. Even though the resolution of sub-pixel scale variability is not explicitly attempted here, ROI characterization was designed to provide detailed subpixel scale information. For each ROI, species presence and their relative covers were estimated using a standard Braun–Blanquet visual method (e.g. Mueller-Dombois & Ellenberg, 1974; Silvestri et al., 2002), which assigns vegetation presence with reference to a sub-division of the 0%–100% range into five intervals. In order to construct a quantitative and less subjective ground reference dataset, relative ground cover was also estimated by acquiring several photographs within each ROI from a nadir-looking digital camera mounted on a 2.5 m-high pole (resulting in a resolution of 2 mm). Relative covers were estimated by overlaying a regular grid (with cell size of 25 cm) onto the digital images and by identifying the type of cover characterizing the pixels at the intersections of the grid (at least 96 within every image of about 5.8 m2) (e.g. Thomson et al., 1998a). Interestingly, the Braun–Blanquet estimates of relative covers made by different trained operators were always very consistent with the more objective, camera-based values. For the scope of the present paper, pixel relative cover information was reduced to the determination of the dominant cover type for each ROI. In this context each pixel is considered to belong to a single class if its area is occupied for more than 60% by a single cover type. Otherwise the pixel is excluded from the reference data set. Majority mapping of vegetation species is thus the subject of the classification procedures described hereafter.

Table 1 Characteristics of the sensors considered and of the corresponding acquisitions

Platform Spatial resolution Spectral range Radiometric resolution Bands Altitude Acquisition date Flight time (GMT) Tidal level Visibility Azimuthal angle Zenithal angle

ROSIS

CASI

MIVIS

Airborne 1m 415–875 nm 12 bit 115 1000 m 08/07/2000 8:00 − 16 cm 6.5 km 103° 46°

Airborne 1.3 m 437–870 nm 16 bit 15 840 m 29/09/2002 8:55 +18 cm 48 km 138° 56°

Airborne 2.6 m 433–12,700 nm 12 bit 102 1300 m 05/07/2003 9:30 −28 cm 56.5 km 128° 31°

Airborne 2.6 m 433–12,700 nm 12 bit 102 1300 m 30/06/2004 11:30 +17 cm 42 km 190° 23°

IKONOS

QuickBird

Satellite P: 1 m M: 4 m P: 450–900 nm M: 450–880 nm 11 bit 1+4 680 km 26/06/2001 10:00 − 20 cm 12 km 140° 27°

Satellite P: 0.72 m M: 2.88 m P: 450–900 nm M: 450–900 nm 16 bit 1+4 450 km 25/07/2003 9:45 +8 cm 47 km 136° 32°

Tidal level, visibility, azimuthal angle and zenithal angle values are given for an intermediate time between the beginning and the end of acquisition. P = panchromatic; M = multispectral.

58

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

The ground reference dataset constructed in this manner is constituted by a set of geocoded ROIs for each of the classes of interest, which include the five dominant halophytic species (Juncus, Spartina, Limonium, Sarcocornia, Salicornia) with the addition of bare soil and water classes. These seven classes are known to describe the greater part of surface types within the study site. Finally, ancillary field observations also include sun-photometric measurements (using a CIMEL CE 318 operating in the following bands: 440 nm, 670 nm, 870 nm, 936 nm, 1020 nm, see TIDE, 2005 for details) performed during hyperspectral overflights and horizontal visibility observations for all multispectral/hyperspectral acquisitions (as described in Table 1). 3. Methods

Table 2 Number of ROIs (NR) and total number of pixels (NP) in ROIs for each vegetation type and sensor Sensor ROSIS

Lim

NR = 3 NP = 317 CASI NR = 4 NP = 470 MIVIS 2003 NR = 3 NP = 197 MIVIS 2004 NR = 7 NP = 244 IKONOS NR = 3 NP = 317 QuickBird NR = 3 NP = 165

Jun

Spa

Sar

Sal

Soil

NR = 2 NP = 150 NR = 1 NP = 40 NR = 2 NP = 36 NR = 4 NP = 45 NR = 2 NP = 150 NR = 2 NP = 36

NR = 4 NP = 463 NR = 2 NP = 174 NR = 1 NP = 44 NR = 1 NP = 16 NR = 4 NP = 463 NR = 1 NP = 44

NR = 4 NP = 451 NR = 3 NP = 207 NR = 2 NP = 58 NR = 3 NP = 68 NR = 4 NP = 451 NR = 2 NP = 53

– – – – – – NR = 2 NP = 48 – – – –

NR = 2 NP = 188 NR = 2 NP = 32 NR = 2 NP = 16 NR = 3 NP = 41 NR = 2 NP = 187 NR = 2 NP = 16

Lim = Limonium; Jun = Juncus; Spa = Spartina; Sar = Sarcocornia; Sal = Salicornia.

3.1. Preliminary data processing ROSIS and IKONOS acquisitions were atmospherically corrected using MODTRAN (in its ATCOR implementation, Richter & Schlapfer, 2002), while the remaining datasets were transformed into reflectance values by the use of the 6S model (Vermote et al., 1997, using a ‘Maritime’ type of atmospheric profile) on the basis of sun-photometer-derived atmospheric optical thickness or horizontal visibility observations according to availability. The atmospheric correction is known not to influence the result of classifications obtained using within-scene training and validation ROIs (e.g. Hoffbeck & Landgrebe, 1994), but was here performed to produce a homogeneous dataset for future direct comparisons or analyses, e.g. based on vegetation indexes. After the atmospheric correction, geometric correction schemes were applied to data from airborne sensors to minimize the distortions induced by perturbations in aircraft altitude and flight direction. While ROSIS and CASI data were corrected by the data producers using a full set of ancillary information (onboard GPS and inertial system data), MIVIS data were corrected using the PARGE model (Schlapfer & Richter, 2002), based just on the onboard GPS acquisitions. Finally, all multispectral and hyperspectral data were accurately geocoded. In order to maximize the accuracy of image geocoding, a set of Ground Control Points (GCP) was selected on the basis of a reference aerial photograph (0.16 m resolution) acquired in the year 2000 and previously geometrically corrected in the Gauss– Boaga Italian reference system, by means of a large number of visible markers laid out throughout the scene. GCPs for remote sensing data geocoding were selected using the few existing buildings (mainly fisherman huts) and easily distinguishable points along marsh edges or at tidal creek intersections. More than 150 points per km2 were used with an approximately homogeneous distribution. Geocoding was performed using a nearestneighbor resampling and yielded Root–Mean–Square errors smaller than 0.5 pixels in all cases. ROIs for the vegetation and soil classes were constructed by selecting just the inner pixels within each ground reference area, as border pixels are more likely to be mixed. The characteristics of the resulting ROIs are summarized in Table 2. It should be noted that, because of their very characteristic spectral signature, maps

of water pixels are very accurate (and all very similar) using any classifier. We thus adopted the K-means classification of the water class as a reference. This was then used as a mask excluded from the further classification experiments performed for the vegetation and soil classes on the remaining pixels. The number of reference pixels in Table 2 is of course dependent on the resolution of the sensor and on the vegetation species. Different halophytes, in fact, are characterized by different overall presence and by different patch sizes (e.g. Juncus is relatively less frequent and its patches are typically smaller than for all other plants), thus limiting the number and size of ROIs available. The set of pixels within the reference areas in Table 2 was divided into a training set, for classifier calibration, and a validation set, to provide an independent test of classification performance. Calibration and validation reference pixels varied according to sensor resolution, but they were always selected from separate vegetation patches to ensure statistical independence and to allow assessments of the actual generalization capabilities of the classifiers. Three classifiers are here examined: The unsupervised Kmeans classification procedure (e.g. Tou & Gonzalez, 1974), and the supervised Spectral Angle Mapper (SAM, e.g. Kruse et al., 1993) and Maximum Likelihood (ML, e.g. Hoffbeck, 1995) algorithms. These classifiers were selected because they are most widely and generally applied, while specific applications to intertidal vegetation are still relatively uncommon. Also, their application poses different constraints on the spectral characteristics of the data, and on the type and amount of reference information required. Indications on the relative and absolute performance of these classifiers applied to halophytic vegetation mapping under different conditions and with different remotely sensed data are thus highly desirable. In order to explore and optimally use the information content of the different hyperspectral data sets considered, we performed analyses using: i) the original spectral bands; ii) a progressively reduced number of bands, selected using a feature selection and a feature extraction algorithm, based, respectively, on the Bhattacharyya distance (e.g. Jimenez & Landgrebe, 1998) and on the Maximum Noise Fraction (MNF) transform (Green et al., 1988); iii) a reduced number of broader bands

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

59

obtained by spectral averaging, mimicking multispectral RGBI data. The latter strategy is interesting because it allows the quantification of the accuracy that would be obtained from a multispectral sensor (typically from a satellite platform) with the same spatial resolution of the hyperspectral sensor considered. 3.2. Unsupervised classifications The K-means unsupervised classification procedure was applied by tentatively fixing a number of classes and by evaluating the resulting map by visual appraisal and by comparison with ground reference data. The results obtained by using about 50 classes, which are then merged together on the basis of calibration ROIs and of the operator direct knowledge of the site (Figs. 2(c) for CASI and 3(c) for IKONOS), show that the information classes of interest are spectrally separable. 3.3. Supervised classifications The reference spectra necessary for the application of the SAM classifier were defined by simply averaging the spectra corresponding to the pixels within the training sets. In the case of the ML algorithm, the training spectra were used to compute the statistics (mean and covariance matrix for all the spectral channels) necessary for its application. The application of the SAM classifier requires the definition of threshold spectral angles above which an unknown spectrum

Fig. 3. Classifications of San Felice IKONOS June 2001 data using all reference areas (TEST2). ((a) Maximum Likelihood, (b) Spectral Angle Mapper, (c) K-means).

is left unclassified (e.g. Kruse et al., 1993). These parameters were determined, for each class of interest, by optimizing the classifier accuracy as characterized by the Confusion Matrix (e.g. Congalton and Green, 1999), and by verifying that classification results were consistent with direct field observations. Sample results of the application of the SAM algorithm to the San Felice marsh are shown in Figs. 2(b) 3(b) and 4. The calibration of the SAM algorithm requires a relatively small amount of training spectra (e.g. as compared to Maximum Likelihood) to provide its maximal performance. The performance of the ML classifier tends to be appreciably degraded when the number of spectral bands available is increased, because of the large number of covariances that need be estimated (Hughes' effect, Hughes, 1968; Swain & Davis, 1978). The quantity of ground reference data required by the ML classifier thus rapidly increases with the number of bands. As discussed above, in order to reduce the dimensionality of the data and thus minimize the Hughes' effect, we applied ML to spectral data obtained using feature selection and extraction algorithms, whose optimality was evaluated on the basis of validation areas and direct field knowledge. 3.4. Accuracy assessment Fig. 2. Classifications of San Felice CASI September 2002 data using all reference areas (TEST2). ((a) Maximum Likelihood, (b) Spectral Angle Mapper, (c) K-means).

The performance of different classifiers is evaluated both in terms of visual comparisons with general field information on

60

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

Fig. 4. Spectral Angle Mapper vegetation maps (TEST2) for the ROSIS (a), MIVIS 2003 (b), MIVIS 2004 (c), and QuickBird (d) data considered.

known vegetation structures and of the Confusion Matrix. A very important first evaluation of classifier results, in fact, uses the large amount of direct information on the spatial distribution of vegetation collected during the field campaigns. The use of this information was facilitated by the fact that different halophytes colonize quite distinct portions of the marsh: Spartina and Salicornia colonize lower areas in the middle of the marsh and, moving to higher grounds towards the nearest channel, one usually first encounters Limonium-dominated areas, and then Sarcocornia zones on even higher soils (e.g. Silvestri et al., 2005). The resulting species patches (zonation), several of which were identified during field activities, constitute a quite stringent criterion for the evaluation of the vegetation maps produced. It was thus possible to discard or accept classifications characterized by relatively good validation statistics on the basis of their consistency with macroscopic features of vegetation distribution. The direct quantitative comparison of classifier performances is based on the Confusion Matrix, whose information will in the following be often summarized by the Overall Accuracy, A, defined as the ratio of the number of validation pixels that are

classified correctly to the total number of validation pixels irrespective of the class (e.g. Foody, 2002). A further important Confusion Matrix statistics used here is the Kappa coefficient, K, which describes the proportion of correctly classified validation sites after random agreements are removed (Rosenfield & Fitzpatrick-Lins, 1986). Comparisons between classifications were performed by testing differences in the Kappa values for statistical significance at the 95% significance level (Congalton, 1983; Hudson and Ramm, 1987). Because the K-means, SAM and ML classifiers involve different procedures and require a different amounts of calibration ROIs, the comparative evaluation of their performance cannot be based on a single choice of calibration/validation data sets. Different types of accuracy tests were thus performed. The first test (TEST1) consists of determining by trial-and-error, for each classifier and each sensor, the ‘optimal’ selection of training ROIs (with the constraint of ensuring that the training ROIs were about half of the total available reference areas) and of classifier parameters, which maximize classification performance. In TEST1, the minimum size of the training set required to obtain the maximal classifier performance was determined. The results of TEST1 constitute an estimate of the best classification performance that should be expected when an independent validation set is available. Training and validation ROIs are non-overlapping, as described above, and allow a statistically significant assessment of the classifier performance and generalization capabilities. The fact that, in TEST1, ‘optimal’ training and validation ROIs are selected separately for each classifier, does not allow the comparison of the performance of the different classification schemes on a common validation set. A second type of test was thus performed (TEST2), based on a resubstitution approach (e.g. Fukunaga, 1990), which uses all available ROIs both for training and validation. Training and validation spectra for TEST2 are identical, to determine the ability of each classifier to discriminate the information classes of interest within the ground reference dataset. However, TEST2 allows a coherent comparison of different algorithms on a single calibration–validation set and the estimation of an upper bound for the performance attainable by a classifier. TEST3 consists of the evaluation of classifier performances using ‘optimal’ training sets (i.e. the training sets of TEST1) and all ROIs as validation sets (i.e. validation ROIs used in TEST2). Comparisons between the results of TEST2 and TEST3 provide indications as to the generalization capabilities of the classifiers. Similar performances of TEST2 and TEST3 indicate that, given a limited number of training spectra, a classifier correctly identifies a proportion of spectra close to the ‘upper-limit’ proportion it can correctly identify when trained on the whole set of reference ROIs (and thus under optimal conditions). On the contrary, a decreased performance in TEST3 with respect to TEST2, indicates partial failure of the classifier to identify a general classification rule applicable to spectra which it was not trained on. The calibration of the K-means classifier was performed using all the information available to the operator (including direct field information) because all the ground reference observations are required to group the initial spectral classes produced by the algorithm into meaningful information classes

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

corresponding to different vegetation species. Because the Kmeans maps are validated on the entire ROI set, TEST2 and TEST3 are in this case identical.

Tables 3 and 4 show the ‘validation’ confusion matrices for CASI and IKONOS using the validation ROIs of TEST2. The confusion matrices for the remaining sensors are not shown here for brevity, but the overall classifier performance for each sensor and each TEST is summarized in Table 5. Table 5 describes the results of the classifications obtained using the original spectral dataset for K-means and SAM (except for the ROSIS data, in which case the ‘Blue’ part of the spectrum was noisy and had to be excluded to obtain optimal results, as described in detail later) and the first 4 MNF bands for ML (which resulted to be optimal in almost all cases, please see below). Comparisons with the performance obtained with the numerous band configurations explored are given at appropriate points in the detailed discussions below. 4.1. ROSIS classifications

Table 3 TEST2 Confusion Matrix for CASI classifications of San Felice salt marsh using the K-means (KM), Spectral Angle Mapper (SAM) and Maximum Likelihood (ML) classifiers

Lim

Jun

Spa

Sar

Soil

Total

Lim

Jun

Spa

Sar

Soil

Total

ML SAM KM ML SAM KM ML SAM KM ML SAM KM ML SAM KM ML SAM KM

Lim

Jun

Spa

Sar

Soil

Total

285 276 269 7 12 7 16 0 12 9 29 27 0 0 2 317 317 317

3 15 4 125 131 69 3 0 3 19 0 12 0 4 62 150 150 150

10 14 6 3 6 1 448 423 406 0 18 32 0 0 16 463 463 463

3 39 116 3 19 13 1 0 3 443 393 318 1 0 1 451 451 451

3 3 0 16 24 9 2 3 0 2 0 3 164 157 175 187 187 187

304 347 395 154 192 99 470 426 424 473 440 392 165 161 256 1568 1568 1568

Visual inspection and experiments with SAM classifications, exhibiting a decrease in accuracy when the bands covering the ‘Blue’ part of the spectrum are used, indicate that the first 27 ROSIS bands are affected by significant noise. As an example, SAM Table 5 Overall Accuracy (A) and Kappa coefficient (K) summarizing the Confusion Matrix characteristics for all sensors and validation TESTs considered TEST1

Lim

Jun

Spa

Sar

Soil

Total

455 412 417 1 0 0 2 2 30 12 56 23 0 0 0 470 470 470

0 0 2 39 35 28 1 5 6 0 0 0 0 0 4 40 40 40

6 4 1 8 10 4 160 150 153 0 8 0 0 2 16 174 174 174

4 11 7 0 0 0 0 0 0 203 196 200 0 0 0 207 207 207

0 0 0 0 0 0 0 0 0 0 0 0 32 32 32 32 32 32

465 427 427 48 45 32 163 157 189 215 260 223 32 34 52 923 923 923

Lim = Limonium; Jun = Juncus; Spa = Spartina; Sar = Sarcocornia.

TEST2

TEST3

A

K

A

K

A

K

97.8% 85.8% Same as TEST2 93.4% 90.9% Same as TEST2 93.4% 76.3% Same as TEST2 89.0% 80.5% Same as TEST2 93.4% 74.6% Same as TEST2 91.1% 81.2% Same as TEST2

0.97 0.81 Same as TEST2 0.90 0.86 Same as TEST2 0.89 0.65 Same as TEST2 0.83 0.71 Same as TEST2 0.91 0.67 Same as TEST2 0.86 0.72 Same as TEST2

ROSIS

ML SAM KM

99.2% 92.6% –

0.99 0.90 –

99.0% 93.9% 90.4%

0.99 0.92 0.87

CASI

ML SAM KM

92.6% 96.1% –

0.89 0.94 –

96.3% 89.4% 89.9%

0.94 0.84 0.85

MIVIS03

ML SAM KM

88.3% 89.2% –

0.84 0.83 –

94.0% 82.0% 81.2%

0.91 0.72 0.69

MIVIS04

ML SAM KM

80.8% 85.5% –

0.59 0.76 –

93.5% 84.2% 77.0%

0.90 0.77 0.67

IKONOS

ML SAM KM

97.2% 89.0% –

0.96 0.85 –

94.8% 88.0% 78.9%

0.93 0.84 0.73

QuickBird

ML SAM KM

96.6% 89.5% –

0.95 0.86 –

92.3% 81.5% 86.9%

0.88 0.72 0.80

Test areas (pixel)

ML SAM KM ML SAM KM ML SAM KM ML SAM KM ML SAM KM ML SAM KM

Test areas (pixel)

Lim = Limonium; Jun = Juncus; Spa = Spartina; Sar = Sarcocornia.

The tidal level was quite low at the time of the ROSIS acquisition (about −0.16 m a.m.s.l., compared to an average marsh level of 0.26 m a.m.s.l.), resulting in wide areas of exposed bare soil and in a small probability of water standing on the marsh surface. These are likely optimal conditions, in particular because possible effects of variable soil moisture are minimized. The data used are composed of two adjacent flight lines. Unfortunately the second flight line (in the upper-right part of the scene in Fig. 4 (a)) was acquired under cloudy conditions, with visible effects in the corresponding vegetation maps.

Classes

Table 4 TEST2 Confusion Matrix for IKONOS classifications of San Felice salt marsh using the Maximum Likelihood (ML), Spectral Angle Mapper (SAM) and Kmeans (KM) classifiers Classes

4. Results

61

The classifications referred to here are based on the original set of bands for SAM and K-means (except ROSIS classifications, which were performed by excluding the noisy bands in the ‘Blue’ region) and on the first 4 MNF bands for ML.

62

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

classifications using all 115 bands yielded A = 71.3% (TEST2), while the use of 88 bands (in the range 523.5 nm–875.5 nm) resulted in A = 93.9% and K = 0.92 (see Table 5). The application of the Maximum Likelihood algorithm requires a reduction in the number of spectral bands because of the well-known sensitivity of the estimation of band covariances on the size of the calibration set with respect to the number of bands (e.g. Hughes, 1968). We first performed classification experiments using a progressively increasing number of bands selected by maximizing the Bhattacharyya distance between the distributions of spectra defined by reference areas of each class. The classifications show a progressive increase in accuracy with the number of bands up to 4 to 6 bands. When the number of bands is increased further, the overall accuracy hardly shows any improvement and the coherence of the vegetation maps with known macroscopic features of species spatial arrangement tends to decrease. This circumstance may be related to the existence of a tradeoff between the number of bands, the amount of non-redundant spectral information, and the Hughes' effect. The approximately optimal configuration was found to correspond to the selection of 4 spectral bands (547.5–551.5, 667.5–671.5, 703.5–707.5, 767.5–771.5 nm, note that the feature selection algorithm excludes the noisy bands in the ‘Blue’ part of the spectrum), yielding A = 98.3% and K = 0.98 (TEST2). We also reduced the number of bands by spectral averaging into broader bands and by the use of the MNF transform. Band averaging to an RGBI spectral configuration produced satisfactory classifications on the basis of both Confusion Matrix and visual inspection. For comparison with the Bhattacharyya-selected bands, the classification using 4 averaged broad bands (451.5– 523.5; 523.5–603.5; 631.5–691.5; 759.5–875.5 nm, a band configuration similar to the QuickBird and IKONOS sensors) yielded A = 97.6% and K = 0.97 (TEST2, the difference in K not being statistically significant at the 95% confidence level). Note that the RGBI configuration includes the noisy bands in the ‘Blue’ part of the spectrum. Averaging to four, equally spaced, broad bands excluding the ‘Blue’ channels (555.5– 635.5; 635.5–715.5; 715.5–795.5; 795.5–875.5 nm) improves this performance to A = 99.9% and K = 1.00 (TEST2, with a good visual correspondence with known vegetation structures), showing that indeed the classification accuracy on broad bands is high when noise is eliminated. Classifications on MNFtransformed data also confirm that a greater number of bands does not contribute an amount of information capable of balancing the greater uncertainty related to the larger number of statistics needed to apply the ML algorithm (Hughes' effect). The best ML classification performance in this case was obtained using the first four bands produced by the MNF transformation (A = 99.0%, K = 0.99 for TEST2, see Table 5, virtually indistinguishable from the broad band classification). We also applied a Principal Component Analysis (PCA), as a reference for MNF results. A classification based on the first four PCA bands yield: A = 96.5% and K = 0.95. The slightly decreased performance may be ascribed to the different noise component present within each ROSIS band, as PCA and MNF results should be the same when the noise variance is the same in all bands (e.g. Green et al., 1988).

The SAM algorithm does not require a reduction of the number of bands, as it does not make use of possibly ill-estimated band covariances. The best results were obtained when the noisy bands in the ‘Blue’ region of the spectrum were excluded from the original band set, as discussed above (A = 93.9%, K = 0.92 for TEST2, see Table 5). The results are also in visual good agreement and consistent with direct field information, except for a minimal number of pixels which are erroneously assigned to the Juncus class. Experiments in band number reduction were carried out also with the SAM algorithm, using MNF transforms and band-averaging. Such experiments indicate that the SAM classification performance (characterized both visually and statistically) is appreciably decreased when a limited number of bands is used. For example, use of the first four MNF bands yields A = 94.8%, K = 0.93 (TEST2), but the corresponding spatial vegetation distribution is not entirely consistent with the overall observed vegetation distribution. Use of four averaged RGBI bands (representing QuickBird- and IKONOS-like bands as described above) results in A = 80.9% and K = 0.75 (TEST2), likely due to the inclusion of the noisy ‘Blue’ region bands in the first averaged band (the accuracy increases to A = 87.4% and K = 0.84, when the four equallyspaced broad bands excluding the ‘Blue’ region are used). Application of the K-means unsupervised classifier using 88 bands, i.e. excluding the first 27 noisy bands, results in A = 90.4% and K = 0.87. K-means performances are thus inferior to both SAM (using the same spectral information) and ML (using a smaller set of bands). As may be seen in Table 5, TEST2 and TEST3 results are quite similar for the ML classifier, indicating that this algorithm is able to effectively classify spectra not used for its training. The SAM algorithm, on the contrary, does not seem to be able to efficiently extract general information from the limited training set of TEST3. The overall comparison of classification accuracies according to the different tests indicates that ML tends to perform better than SAM in the ROSIS case. From a more qualitative viewpoint, all maps indicate an excessive presence of Juncus, which is in contrast with direct field knowledge. This observation may be interpreted by considering that Juncus-dominated areas tend to contain nonnegligible amounts of other species. Therefore, because Juncus training pixels are highly mixed, it is likely that the classifiers tend to interpret as Juncus any mixed pixel, whose spectral reflectance is decisively different from the remaining vegetation classes. However, the maps retrieved exhibit a remarkable overall agreement with observed vegetation spatial structures, particularly for Limonium and Sarcocornia. 4.2. CASI classifications The tidal level at the time of the CASI acquisition considered (29 September 2002) was quite high (the highest among all the acquisitions considered), at about 18 cm a.m.s.l., resulting in a greater presence of water particularly in the smaller channels (see Fig. 2). The San Felice site was covered using two flight lines, acquired under slightly different illumination conditions, with consequences for classifications results.

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

The application of the ML classifier again showed that the Hughes' effect dominates when a large number of bands is used. We thus experimented with different band selections according to the Bhattacharyya distance between the classes. An optimal performance, based on the Confusion Matrix (A = 95.1% and K = 0.93, TEST2) and comparisons to known vegetation structures, is obtained when just 4 bands are selected: 437.15– 445.65; 542.7–554.1; 677.15–682.05; 860.9–869.7 nm. Band averaging experiments yield an optimal performance (A =96.0% and K =0.94, TEST2) when CASI bands are averaged down to 4 RGBI (QuickBird- or IKONOS-like) spectral bands: 437.15– 494.80; 542.70–554.10; 614.65–697.35; 705.75–869.70 nm. Again, this may be ascribed to a tradeoff between the amount of spectral information used and the uncertainty in the estimation of an increasing number of band covariances. The ML classification on MNF-transformed bands yielded A = 96.3% and K = 0.94 (TEST2, see Table 5). The results are identical (also visually, see Fig. 2 (a)) to the broad band classification. A classification on 4 PCA bands gives A = 96.6% and K = 0.94 (TEST2). The very similar performances obtained using the PCA and MNF bands suggest that CASI bands are homogeneously affected by a relatively small noise component. It is worth noting that, overall, the use of averaged bands yields performances which are equivalent to those obtained from feature selection or extraction algorithms. The application of the SAM classifier to the full 15-band set gives A = 89.4%, K = 0.84 for TEST2, though the associated vegetation map likely overestimates the presence of Limonium with respect to Sarcocornia (Fig. 2 (b)). Tests of the SAM algorithm on the averaged RGBI spectral bands, yielding satisfactory results for ML, gave significantly poorer results (A = 80.2%, K = 0.70 for TEST2) than in the case of the full 15band data. Classifications based on a decreasing number of MNFselected bands are consistently characterized by poorer performances both statistically (for the 4 most signal-containing MNF bands: A = 89.2%, K = 0.83) and visually. TEST3 results for ML are just slightly inferior to those for TEST2, while the differences for SAM are not statistically significant. This indicates that both classifiers can reasonably generalize the spectral information contained in the training set. K-means accuracies are comparable (A = 89.9% and K = 0.85, Fig. 2 (c)) to SAM performances (at the 95% confidence level) and the corresponding vegetation map is in good agreement with the overall species distribution. In summary, the CASI classifications explored yield vegetation distributions in satisfactory agreement with observations. As for ROSIS, the ML algorithm outperforms the SAM and Kmeans classifiers when the number of bands is reduced by spectral averaging or feature extraction/selection schemes. It is worth noting here that ML classifications using all the band configurations explored (e.g. see Fig. 2 (a)) show difficulties in discriminating bare soil and Spartina (e.g. in some portions from the wide expanses of low-elevation bare soil areas in the upper part of the marsh). This is probably due to the fact that Spartina is composed of almost-vertical stems and that a great portion of soil is thus visible from above. The resulting spectrum is therefore a combination of Spartina's own

63

signature and the signature of the soil. On the other hand, the marsh surface sediment is covered by variable proportions of microalgae (microphytobenthos, e.g. Aspden et al., 2004), containing significant quantities of chlorophyll. The spectral reflectance of marsh soil is thus a combination of a characteristic soil signature and of a vegetation-like spectrum (with the typical red edge), contributed by the microphytobenthos, and may thus be very similar to the signature of Spartina (e.g. Thomson et al., 2003). 4.3. MIVIS classifications 4.3.1. MIVIS 2003 The tidal level during the acquisition (5 July 2003) was very low (the minimum level among all acquisitions considered), at −0.28 m a.m.s.l., resulting in wide areas of exposed bare soil and in relatively little soil moisture variability. The application of the ML classifier using the bands selected by maximizing the intra-class Bhattacharyya distance (513–533 nm; 593–613 nm; 773–793 nm; 8.20–8.60 μm) gives A = 96.6%, K = 0.95 (TEST2). The selection of a thermal band indicates that the emissive part of the spectrum indeed contains useful information to enhance vegetation species separability. The classification result is however in evident disagreement with field observations of vegetation structure, mainly because Sarcocornia is often misclassified as Juncus and Spartina is mistakenly mapped as soil. The high performance indicated by the Confusion Matrix statistics is therefore in this case misleading. This can partly be attributed to the relatively small number of reference areas available for MIVIS acquired in 2003, which, combined with the quite coarse resolution of the MIVIS data (2.6 m), results in a limited amount of pixels falling within the ground reference areas for each class (e.g. see Table 2). In this case direct information on vegetation structures, however subjective, is invaluable in the accuracy assessment process (e.g. Foody, 2002). However, the effects of geometric resolution are not likely to entirely explain the poorer performance of MIVIS classifications if one considers the noteworthy coherence of QuickBird vegetation maps (which have an even coarser resolution, see below the detailed discussion) with respect to field observations. Speculatively, one may suggest the poorer performance of MIVIS classifications to be related to inaccuracies in geometric corrections of MIVIS data (e.g. due the lack of an onboard inertial system). The application of the ML classifier to 4 averaged RGBI broad bands (453–533; 533–613; 633–693; 753–833 nm) yields A = 94.3% and K = 0.91 (TEST2) and vegetation distributions again in disagreement with observed features. Similarly visually unrealistic results are obtained when the first 4 MNF bands (A = 94.0% and K = 0.91, TEST2, see Table 5) or PCA bands (A = 88.9% and K = 0.83, TEST2) are used. SAM results for the original set of bands give A = 82.0% and K = 0.72, with a slight improvement in the estimated abundance of Sarcocornia with respect to ML vegetation maps (see Fig. 4 (b)). Also the application of K-means to the MIVIS 2003 data produces vegetation maps in disagreement with the known

64

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

distribution of species (especially due to an underestimation of Sarcocornia), with statistics also indicating a low performance (A = 81.2% and K = 0.69). The Confusion Matrix statistics suggest a slight superiority of ML, particularly on the basis of TEST2 and TEST3. This superiority is not confirmed by comparisons to the overall known species distributions, which e.g. indicate diffuse misclassifications of bare soil and Spartina areas in ML maps. In conclusion, in spite of the statistics results, the direct comparisons to known vegetation structures indicate that SAM classifications are somewhat more robust in capturing the overall vegetation and soil patterns in the case of MIVIS 2003. 4.3.2. MIVIS 2004 The tidal level was quite high (0.17 m a.m.s.l.) during the acquisition (30 June 2004) and thus a large part of bare soil areas, which were visible in the MIVIS 2003 acquisition, were covered by water in the MIVIS 2004 image. MIVIS 2004 data are quite interesting because they document a sudden change in vegetation patterns, which were otherwise relatively stable in the period 2000–2003. The vegetation maps from all classifiers, in fact, show the almost complete replacement of Spartina by Salicornia (coral color), which was not previously present in the San Felice salt marsh. This event is confirmed by field inspections and has occurred on a Lagoon-wide scale. The application of ML to the 4 bands selected according to the Bhattacharyya distance (653–673 nm; 693–713 nm; 1.15–1.20 μm; 1.45–1.5 μm) gives A = 90.9% K = 0.87 (TEST2). The selection of two thermal bands confirms that this part of the spectrum contains useful information for halophytic vegetation mapping. The vegetation map shows an excessive presence of Juncus. The ML results from 4 averaged RGBI bands (453–533; 533–613; 633–693; 753–833 nm) yield A = 88.1% and K = 0.82 (TEST2), while the application to the 4 most significant MNF bands gives A = 93.5% and K = 0.90 (TEST2, also see Table 5). In all cases, Juncus is clearly overestimated (results not shown here for brevity). The K-means (A = 77.0%, K = 0.67, TEST2) and SAM (A = 84.2%, K = 0.77) maps obtained from the original set of MIVIS bands (Fig. 4 (c)) indicate very similar patterns for the newly appeared Salicornia. Even though the Confusion Matrix statistics indicate a marginally better performance of ML (Table 5), comparisons with known vegetation structures are suggestive of a superior performance of the SAM scheme, in particular because ML indicates an excessive presence of Juncus. Due to the relatively coarse resolution the amount of reference MIVIS pixels is smaller than for the remaining hyperspectral data. Therefore the field information on the overall species distribution is in this case important to conclude that SAM results should be considered superior to those from ML and K-means. 4.4. IKONOS classifications The tidal level during the acquisition was about − 0.20 m a.m.s.l., thus minimizing the spatial variability of water presence over the marsh.

The application of ML, SAM, and K-means to the IKONOS data (as well as to QuickBird data, as seen below) is much more straightforward than in the case of hyperspectral sensors, as it does not require feature selection or extraction procedures. The maps obtained from the three classifiers (Fig. 3) exhibit vegetation structures which are in good agreement with one another and are coherent with ground reference observations (Table 4) and direct field knowledge. Considering that IKONOS data have been pansharpened to a 1 m resolution from the original multispectral 4 m resolution, the convincing classification results are indirect evidence of a good performance of the pan-sharpening procedures. All classifiers indicate a more widespread presence of Spartina than seen in any other acquisition and relatively few bare soil areas. This may in part be due to the cited difficulty in discriminating bare soil from Spartina areas, but direct field information suggest that this mainly reflects an actual abundance of Spartina presence resulting from the interannual variability of species presence. Overall, the ML algorithm outperforms the K-means and SAM schemes (as confirmed by statistical significance tests performed on K values at the 95% significance level). However, ML performance is, as usual, very sensitive to the number of training pixels used (experiments not shown for brevity). The decreased performance measured in TEST3 with respect to TEST2 (Table 5) reasonably indicates that the SAM classifier has a greater difficulty in generalizing the representative spectral properties of the information classes of interest, possibly due to the limited number of bands available. 4.5. QuickBird classifications The tidal level during the acquisition (25 July 2003) was 0.08 m a.m.s.l. It is interesting to compare the results of QuickBird classifications to those from MIVIS 2003, acquired nearly in the same period (Fig. 4 (b) and (d)). The overall patterns are quite similar and consistent with observations (except for a pronounced fragmentation in the MIVIS 2003 map, possibly due to the inaccuracies of MIVIS data discussed previously). The ML map derived using the training set of TEST1 (not shown) tends to depart from K-means (not shown) and SAM classifications (mainly due to a misclassification of Sarcocornia and Spartina in different areas) and is the least accurate on the basis of visual appraisal. The Confusion Matrix statistics, on the contrary, indicate a superior performance of ML (A = 96.6%, K = 0.95 for TEST1, see Table 5) with respect to the remaining two algorithms. This again points to the importance of comparisons with direct field information, particularly when the relatively coarser resolution (2.88 m) causes the pixels to be more highly mixed, and reduces the number of training pixels, thereby reducing the statistical significance of the Confusion Matrix. In fact, the use of all reference areas (TEST2) for ML training reconciles the indications of the Confusion Matrix and of visual appraisal, yielding vegetation distribution estimates in agreement with other classifiers and with known vegetation structures. K-means and SAM classifications are in satisfactory agreement and capture well the general known vegetation

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

structures, which are mostly coherent with those from MIVIS 2003. The performances of the ML and SAM classifiers are similar for TEST2 and TEST3, showing that both algorithms can effectively generalize the information contained in a limited training set. Once again, QuickBird classifications exhibit the greatest inaccuracies in mapping Juncus. This species is indeed correctly identified by the classifiers where present, but its extent is usually importantly overestimated. 5. Discussion and conclusions In general, it may be concluded that, for the San Felice marsh as for the other study sites not explicitly discussed here, vegetation maps obtained from remote sensing observations at the 1 m scale are in qualitative and quantitative agreement with direct observations. When a large amount of reference information is available (with respect to the number of bands used) the ML classifier outperforms SAM and K-means. In the case of hyperspectral sensors the optimal application of ML requires the reduction of the number of spectral channels used, in order to obtain reliable estimates of band covariances and avoid incurring in the Hughes' effect. The classification experiments performed using i) a feature selection algorithm (based on maximizing the Bhattacharyya distance between the classes), ii) features extraction procedures (MNF and PCA), and iii) simple band averaging, show that indeed there exists a tradeoff between the number of bands, the signal-to-noise ratio (which was a critical factor e.g. for ROSIS and MIVIS) and the amount of ground reference pixels. The overall balance is obtained when adopting a higher spatial resolution, a small number of bands (4 in our case study) and a decreased signal-to-noise ratio, at the cost of a reduced spectral resolution. Considering the limited amount of ground reference information usually available particularly in coastal applications, a higher spatial resolution has the advantage of resulting in a larger number of reference pixels to be used in classifier training (even more so in the case of highly heterogeneous vegetation distributions, which limit the possible extent of reference areas). A higher spatial resolution also reduces the within-pixel heterogeneity, thus increasing their spectral separability (e.g. the case of Juncus in the present application). These conclusions are also supported by the observation that classifications of MIVIS hyperspectral data (characterized by the relatively coarse resolution of 2.6 m) have comparable, and at times poorer, performances with respect to multispectral sensors with a high spatial resolution (particularly IKONOS with a resolution of 1 m). We found MNF and band averaging to be the most effective methods of features reduction. MNF was efficient in extracting information-containing bands and in identifying the noise component in the general case of unequal noise content over different bands (differently from PCA, e.g. see ROSIS classifications). Spectral averaging to obtain RGBI data sets allows performances which are very close to those obtained using MNF transformed bands. The satisfactory performance of classifications of band-averaged data and their similarity with multispectral

65

data classifications is a further support of the prevalent importance of spatial with respect to spectral resolution. When the spatial resolution is relatively low, and significantly reduces the number of reference pixels (e.g. MIVIS and QuickBird), the validation using statistics from the Confusion Matrix may be misleading. In this situation, direct field information, though of a more subjective nature, assumes a crucial role in providing adequate constraints for classifier calibration and validation. When reference spectra are relatively less numerous, the use of SAM and K-means provides more robust and reliable classification results with respect to ML, because they can exploit the entire spectral information available and reduce the Hughes' effect. The grouping of spectral classes into information classes to produce a K-means classification requires quite a significant amount of information, which, given the small typical scale of spatial heterogeneity of halophytic vegetation, is often not available in the form of geocoded ROIs. The operator must thus make decisions on the basis of less objective broad knowledge of vegetation patch position or of species characteristics (e.g., the tendency of some species to colonize the marsh interior, rather than areas near creeks). K-means may thus be useful when the marsh area to be classified is known in its broad spatial vegetation structures and few reference spectra are available. Irrespective of the classifier adopted, the classifications of ROSIS and CASI hyperspectral data are somewhat superior to those from multispectral observations, which have however comparable performances. This circumstance and experiments with features reduction schemes, suggest that much of the information contained in hyperspectral data is redundant for halophytic vegetation classification. In light of the greater importance of geometric resolution with respect to spectral resolution, and considering the smallscale spatial variability of halophytes, we conclude (differently from previous studies, e.g. Thomson et al., 2003) that the use of high-resolution multispectral satellite sensors is highly suited to effectively map intertidal vegetation. This conclusion is of some relevance for studying and monitoring schemes which require repeated observations and thus may benefit from the easier acquisition of satellite data and their relatively lower cost. Regarding the different halophytic species occurring in the study sites considered, classification results indicate that Limonium and Sarcocornia are generally quite correctly and consistently identified using all sensors. The comparison of classifier results for all the acquisitions described indicates that a robust identification of Juncus is quite problematic. This is likely due to the fact that patches of Juncus are usually quite small (order of a few square meters), compared to the resolutions of the sensors. It is therefore quite difficult to identify a statistically significant number of reference pixels. Calibration and validation reference areas are thus inevitably composed of a large proportion of mixed pixels, whose spectra are not truly representative of Juncus. This circumstance generates both commission and omission errors: i) the reference spectra used for classification are not ‘truly pure’ and may be similar to spectra from pixels with ‘mixed vegetation’ (which are thus erroneously classified as Juncus); ii) pixels containing Juncus rarely display

66

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67

a ‘pure’ Juncus spectrum, due to their ‘mixed’ nature, and may thus be mistakenly attributed to another class. Remote sensing mapping of Spartina may also be at times uncertain. This can be explained by considering that in Spartina areas plant density and structure are such that a substantial amount of soil is visible. The spectral signature of a Spartina area is thus a combination of the spectral signature of these species and that of soil. This mixture of spectra may be confused with the spectral signature typical of a soil containing microphytobenthos, which carries significant quantities of chlorophyll, thus causing an overestimation of Spartina areas. This situation is mostly relevant for the lowest areas of the marsh, where typically both Spartina and bare soil areas occur. In conclusion, remote sensing has been shown to be a useful tool for salt-marsh vegetation mapping and for the quantitative characterization of its spatial distribution. Remote sensing classifications should be considered the technique of choice for saltmarsh study and monitoring. In particular, remotely sensed vegetation maps allow the exploration of a wide range of scales of interest in intertidal areas and do not involve the inaccuracies associated with extensive interpolation and extrapolation of (necessarily point) field observations. Furthermore, direct field mapping does not provide ‘instantaneous’ descriptions of vegetation distribution, requiring very long periods of time to be completed in large areas, and e.g. cannot quantitatively describe events such as the observed replacement, at all study sites, of Spartina by Salicornia between years 2003 and 2004. On the contrary, the ability of remote sensing classifications to describe the distribution of halophytes in space and time illustrates the suitability of remote sensing classifications for an accurate monitoring of the spatial structure and the time evolution of saltmarsh vegetation. Acknowledgements This research was funded by the TIDE EU RTD Project (EVK3-CT-2001-00064). The ROSIS data analyzed were acquired through the Hysens project 2000 DLR-EU. The authors wish to thank the Magistrato alle Acque di Venezia for making the IKONOS data available. The authors also acknowledge the contribution of five anonymous reviewers, whose comments substantially improved the quality of the manuscript. References Adam, P. (1990). Saltmarsh ecology. Cambridge: Cambridge University Press. Aspden, R., Vardy, S., & Paterson, D. M. (2004). Salt marsh microbial ecology: Microbes, benthic mats and sediment movement. In S. Fagherazzi, M. Marani, & L. Blum (Eds.), Ecogeomorphology of Tidal Marshes. Coastal and Estuarine Monograph Series. (pp. 115−136) Washington DC: American Geophysical Union. Beeftink, W. G. (1977). The coastal salt marshes of western and northern Europe: An ecological and phytosociological approach. In V. J. Chapman (Ed.), Wet Coastal Ecosystems (pp. 109−155). Amsterdam: Elsevier. Bernhardt, K. G., & Koch, M. (2003). Restoration of a salt marsh system: Temporal change of plant species diversity and composition. Basic and Applied Ecology, 4(5), 441−451. Bird, E. C. F. (1985). Coastline Changes. New York: Wiley and Sons.

Bockelmann, A. C., Bakker, J. P., Neuhaus, R., & Lage, J. (2002). The relation between vegetation zonation, elevation and inundation frequency in a Wadden Sea salt marsh. Aquatic Botany, 73, 211−221. Brown, K. (2004). Increasing classification accuracy of coastal habitats using integrated airborne remote sensing. EARSeL Proceedings, 3(1), 34−42. Cahoon, D. R., & Reed, D. J. (1995). Relationships among marsh surface topography, hydroperiod, and soil accretion in a deteriorating Louisiana salt marsh. Journal of Coastal Research, 11, 357−369. Cahoon, D., Reed, D., & Day, J. (1995). Estimating shallow subsidence in microtidal salt marshes of the southeastern United States: Kaye and Barghoorn revisited. Marine Geology, 128, 1−9. Caniglia, G., Contin, G., Fusco, M., Anoè, N., & Zanaboni, A. (1997). Confronto su base vegetazionale tra due barene della laguna di Venezia. Fitosociologia, 34, 111−119. Chapman, V. J. (1964). Coastal vegetation. New York: The Macmillan Company. Congalton, R. G. (1983). Assessing landsat classification accuracy using discrete multivariate analysis statistical techniques. Photogrammetric Engineering and Remote Sensing, 49(12), 1671−1678. Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotely sensed data: Principles and practices (pp. 137). Boca Raton, Florida: Lewis Publishers. Consorzio Venezia Nuova (2005). Consorzio Venezia Nuova web site, www.salve.it/ Cox, R., Wadsworth, R. A., & Thomson, A. G. (2003). Long-term changes in salt marsh extent affected by channel deepening in a modified estuary. Continental Shelf Research, 23(17–19), 1833−1846. Cronk, J. K., & Fennessy, M. S. (2001). Wetland plants: Biology and ecology. Boca Raton, Florida: CRC Press. Dale, P. E. R., Hulsman, K., & Chandica, A. L. (1986). Classification of reflectance on colour aerial photograph and sub-tropical salt-marsh vegetation types. International Journal of Remote Sensing, 7(12), 1783−1788. Day, Jr., J. W. Rybczyk, J., Scarton, F., Rismondo, A., Are, D., & Cecconi, G. (1999). Soil accretionary dynamics, sea-level rise and the survival of wetlands in Venice Lagoon: A field and modelling approach. Estuarine, Coastal and Shelf Science, 49, 607−628. Donoghue, D. N. M., Thomas, D. C. R., & Zong, Y. (1994). Mapping and monitoring the intertidal zone of the east coast of England using remote sensing techniques and a coastal monitoring GIS. Marine Technology Society Journal, 28, 19−29. Eastwood, J. A., Yates, M. G., Thomson, A. G., & Fuller, R. M. (1997). The reliability of vegetation indices for monitoring saltmarsh vegetation cover. International Journal of Remote Sensing, 18(18), 3901−3907. Finkl, C. W. (1996). What might happen to America's shorelines if artificial beach replenishment is curtailed: A prognosis for southeastern Florida and other sandy regions along regressive coasts. Journal of Coastal Research, 12(1), R3−R9 (WIN). Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of Environment, 80, 185−201. Friedrichs, C. T., & Perry, J. E. (2001). Tidal salt marsh morphodynamics. Journal of Coastal Research, 27, 7−37. Fukunaga, K. (1990). Introduction to Statistical Pattern Recognition, 2nd ed. New York: Academic Press. Green, A. A., Berman, M., Switzer, P., & Craig, M. D. (1988). A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Transactions on Geoscience and Remote Sensing, 26(1), 65−74. Hoffbeck, J. P., & Landgrebe, D. A. (1994). Effect of radiance-to-reflectance transformation and atmosphere removal on maximum likelihood classification accuracy of high-dimensional remote sensing data. Proceedings of the 14th Annual International Geoscience and Remote Sensing Symposium IGARSS'94 (pp. 2538−2540). California: Pasadena. Hoffbeck, J.P. (1995). Classification of high dimensional multispectral data. PhD thesis, Purdue University, West Lafayette, Indiana. Hudson, W. D., & Ramm, C. W. (1987). Correct formulation of the kappa coefficient of agreement. Photogrammetric Engineering and Remote Sensing, 53(4), 421−422. Hughes, G. F. (1968). On the mean accuracy of statistical pattern recognizers. IEEE Transactions on Information Theory, IT-14, 55−63. Hughes, R. G., & Paramor, O. A. L. (2004). On the loss of saltmarshes in southeast England and methods for their restoration. Journal of Applied Ecology, 41(3), 440−448.

E. Belluco et al. / Remote Sensing of Environment 105 (2006) 54–67 Jimenez, L., & Landgrebe, D. (1998). Supervised classification in high dimensional space: Geometrical, statistical, and asymptotical properties of multivariate data. IEEE Transactions on Systems, Man, and Cybernetics, 28(1), 39−54. Johnston, R. M., & Barson, M. M. (1993). Remote-sensing of australian wetlands — An evaluation of landsat TM data for inventory and classification. Australian Journal of Marine and Freshwater Research, 44(2), 235−252. Kruse, F. A., Lekoff, A. B., Boardman, J. W., Heiderbrecht, K. B., Shapiro, A. T., Barloon, P. J., et al. (1993). The spectral image processing system (SIPS)interactive visualization and analysis of imaging data. Remote Sensing of Environment, 44, 145−163. Leatherman, S. P. (2003). Shoreline change mapping and management along the US East Coast. Journal of Coastal Research, 38, 5−13. Leonard, L. A., & Luther, M. E. (1995). Flow hydrodynamics in tidal marsh canopies. Limnology and Oceanography, 40(8), 1474−1484. Marani, M., Belluco, E., Ferrari, S., Silvestri, S., D’Alpaos, A., Lanzoni, S., Feola, A., Rinaldo, A. (2006). Analysis, synthesis and modelling of high-resolution observations of saltmarsh ecogeomorphological patterns in the Venice lagoon. Estuarine, Coastal and Shelf Science, in press. Marani, M., Lanzoni, S., Silvestri, S., & Rinaldo, A. (2004). Tidal landforms, patterns of halophytic vegetation and the fate of the lagoon of Venice. Journal of Marine Systems, 51(1–4), 191−210. Marani, M., Silvestri, S., Belluco, E., Camuffo, M., DAlpaos, A., Defina, A., et al. (2003). Patterns in tidal environments: Salt-marsh channel networks and vegetation. Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Toulouse, France, 21–25 July 2003. Mueller-Dombois, D., & Ellenberg, H. (1974). Aims and Methods of Vegetation Ecology. New York: John Wiley and Sons. Munyati, C. (2000). Wetland change detection on the Kafue Flats, Zambia, by classification of a multitemporal remote sensing image dataset. International Journal of Remote Sensing, 21(9), 1787−1806. Pethick, J. (1984). An introduction to Coastal Geomorphology. London: Arnold Pb. Pignatti, S. (1966). La vegetazione alofila della laguna veneta. Istituto Veneto di Scienze, Lettere ed Arti, Memorie, XXXIII- I, Venezia (pp. 174). Ramsey, E. W., & Laine, S. C. (1997). Comparison of landsat thematic mapper and high resolution photography to identify change in complex coastal wetlands. Journal of Coastal Research, 13(2), 281−292 (SPR). Reed, D. J. (2002). Sea-level rise and coastal marsh sustainability: Geological and ecological factors in the Mississippi delta plain. Geomorphology, 48(1–3), 233−243. Richter, R., & Schlapfer, D. (2002). Geo-atmospheric processing of airborne imaging spectrometry data. Part 2: Atmospheric/topographic correction. International Journal of Remote Sensing, 23(13), 2609−2630. Rosenfield, G. H., & Fitzpatrick-Lins, K. (1986). A coefficient of agreement as a measure of thematic classification accuracy. Photogrammetric Engineering and Remote Sensing, 52, 223−227. Schlapfer, D., & Richter, R. (2002). Geo-atmospheric processing of airborne imaging spectrometry data. Part 1: Parametric orthorectification. International Journal of Remote Sensing, 23(13), 2609−2630. Schmidt, K. S., Skidmore, A. K., Kloosterman, E. H., van Oosten, H. H., Kumar, L., & Janssen, J. A. M. (2004). Mapping coastal vegetation using an expert system and hyperspectral imagery. Photogrammetric Engineering and Remote Sensing, 70(6), 703−715.

67

Shuman, C. S., & Ambrose, R. F. (2003). A comparison of remote sensing and ground-based methods for monitoring wetland restoration success. Restoration Ecology, 11(3), 325−333. Silvestri, S., Defina, A., & Marani, M. (2005). Tidal regime, salinity and saltmarsh plant zonation. Estuarine, Coastal and Shelf Science, 62, 119−130. Silvestri, S., & Marani, M. (2004). Salt marsh vegetation and morphology, modelling and remote sensing observations. In S. Fagherazzi, M. Marani, & L. K. Blum (Eds.), The Ecogeomorphology of Tidal Marshes. Coastal and Estuarine Studies, Vol. 59 (pp. 266) hardbound, ISBN 0-87590-273-1, AGU Code CE0592731. Silvestri, S., Marani, M., & Marani, A. (2003). Hyperspectral remote sensing of salt marsh vegetation, morphology and soil topography. Physics and Chemistry of the Earth, 28(1–3), 15−25. Silvestri, S., Marani, M., Settle, J., Benvenuto, F., & Marani, A. (2002). Salt marsh vegetation radiometry. Data analysis and scaling. Remote Sensing of Environment, 80, 473−482. Smith, G. M., Spencer, T., Murray, A. L., & French, J. R. (1998). Assessing seasonal vegetation change in coastal wetlands with airborne remote sensing: An outline methodology. Mangroves and Salt Marshes, 2, 15−28. Swain, P. H., & Davis, S. M. (1978). Remote Sensing: The Quantitative Approach. New York: McGraw-Hill. Thomson, A. G., Eastwood, J. A., Yates, M. G., Fuller, R. M., Wadsworth, R. A., & Cox, R. (1998). Airborne remote sensing of intertidal biotopes: BIOTA I. Marine Pollution Bulletin, 37(3–7), 164−172. Thomson, A. G., Fuller, R. M., Sparks, T., Yates, M., & Eastwood, J. A. (1998). Ground and airborne radiometry over intertidal surfaces: Waveband selection for cover classification. International Journal of Remote Sensing, 19(6), 1189−1205. Thomson, A. G., Fuller, R. M., Yates, M. G., Brown, S. L., Cox, R., & Wadsworth, R. A. (2003). The use of airborne remote sensing for extensive mapping of intertidal sediments and saltmarshes in eastern England. International Journal of Remote Sensing, 24(13), 2717−2737. Thomson, A. G., Huiskes, A., Cox, R., Wadsworth, R. A., & Boorman, L. A. (2004). Short-term vegetation succession and erosion identified by airborne remote sensing of Westerschelde salt marshes, The Netherlands. International Journal of Remote Sensing, 25(20), 4151−4176. TIDE (2005). TIDE project web site, www.tideproject.org Tou, J. T., & Gonzalez, R. C. (1974). Pattern Recognition Principles. Reading, Massachusetts: Addison-Wesley Publishing Company. Ursino, N., Silvestri, S., & Marani, M. (2004). Subsurface flow and vegetation patterns in tidal environments. Water Resources Research, 40(5), W05115, doi:10.1029/2003WR002702 Vermote, E. F., Tanré, D., Deuzé, J. L., Herman, M., & Morerette, J. J. (1997). Second simulation of the satellite signal in the solar spectrum, 6S: An overview. IEEE Transactions on Geoscience and Remote Sensing, 35(3), 675−686. Wolters, M., Garbutt, A., & Bakker, J. P. (2005). Salt-marsh restoration: Evaluating the success of de-embankments in north–west Europe. Biological Conservation, 123(2), 249−268. Zhang, Y. (2002). Problems in the fusion of commercial high-resolution satellite images as well as landsat 7 images and initial solutions. International Archives of Photogrammetry and Remote Sensing, 34, 4. Zhang, K. Q., Douglas, B. C., & Leatherman, S. P. (2004). Global warming and coastal erosion. Climatic Change, 64(1–2), 41−58.