Next Article in Journal
Figures of Merit for Indirect Time-of-Flight 3D Cameras: Definition and Experimental Evaluation
Next Article in Special Issue
Comparison of Object-Based Image Analysis Approaches to Mapping New Buildings in Accra, Ghana Using Multi-Temporal QuickBird Satellite Imagery
Previous Article in Journal / Special Issue
Object-Based Image Analysis of Downed Logs in Disturbed Forested Landscapes Using Lidar
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

An Object-Based Classification of Mangroves Using a Hybrid Decision Tree—Support Vector Machine Approach

Department of Geography and Center for Galapagos Studies, University of North Carolina at Chapel Hill, Campus Box 3220, Chapel Hill, NC 27599, USA
Current Address: Department of Geography and Global Environmental and Climate Change Centre, McGill University, Burnside Hall Room 705, 805 Sherbrooke St. West, Montreal, QC H3A 2K6, Canada
Remote Sens. 2011, 3(11), 2440-2460; https://doi.org/10.3390/rs3112440
Submission received: 20 September 2011 / Revised: 8 November 2011 / Accepted: 10 November 2011 / Published: 17 November 2011
(This article belongs to the Special Issue Object-Based Image Analysis)

Abstract

:
Mangroves provide valuable ecosystem goods and services such as carbon sequestration, habitat for terrestrial and marine fauna, and coastal hazard mitigation. The use of satellite remote sensing to map mangroves has become widespread as it can provide accurate, efficient, and repeatable assessments. Traditional remote sensing approaches have failed to accurately map fringe mangroves and true mangrove species due to relatively coarse spatial resolution and/or spectral confusion with landward vegetation. This study demonstrates the use of the new Worldview-2 sensor, Object-based image analysis (OBIA), and support vector machine (SVM) classification to overcome both of these limitations. An exploratory spectral separability showed that individual mangrove species could not be spectrally separated, but a distinction between true and associate mangrove species could be made. An OBIA classification was used that combined a decision-tree classification with the machine-learning SVM classification. Results showed an overall accuracy greater than 94% (kappa = 0.863) for classifying true mangroves species and other dense coastal vegetation at the object level. There remain serious challenges to accurately mapping fringe mangroves using remote sensing data due to spectral similarity of mangrove and associate species, lack of clear zonation between species, and mixed pixel effects, especially when vegetation is sparse or degraded.

Graphical Abstract

1. Introduction

1.1. Context

Mangroves are an assemblage of tropical and sub-tropical halophytes (i.e., salt tolerant) woody plants. Mangrove forests are among the most productive forest ecosystems in the world and unique in linking terrestrial and marine systems through the inter-tidal zone [1]. Despite the low tree species diversity and simple canopy structure, mangroves provide many valuable ecosystems goods and services such as carbon sequestration, habitat for terrestrial fauna as well as economically important fisheries, and coastal hazard mitigation [2]. Mangrove forests can range from vast swamps across large estuarine systems such as the Ganges River Delta to strips of vegetation along the fringe of arid coastlines.
Globally, satellite remote sensing has played an important role in mapping and monitoring mangroves [3,4]. Mapping and monitoring mangrove forests is critically important for numerous scientific areas such as carbon stock estimates of tropical coastal nations, effectively managing commercial fisheries and their mangrove nurseries, and understanding the dynamics of vegetation-coastal geomorphology and coastal hazard mitigation. Furthermore, mangroves can provide unique habitat for rare species such as the mangrove finch in the Galapagos Islands of Ecuador.
Previous studies have reported remote sensing classification accuracies between mangroves and other landcover ranging from 75% to 90%, though many studies have omitted accuracy assessments (see [4] for an in-depth review of satellite remote sensing or [5] for a more general review). There remain a number of challenges to accurately detect mangroves including spectral similarity between mangroves and nearby landward tropical vegetation including in arid or marginal environments [6,7,8,9] and the effect of mixed pixels for fringe mangroves [10]. Detection of individual mangrove species presents an even greater challenge. Traditional remote sensing approaches generally have failed to detect individual species [11]. While Vaiphasa et al. [12] and Wang and Sousa [13] were able to discriminate between mangrove species in hyperspectral laboratory studies, real-world results have been mixed. Almost all recent studies utilize very high resolution imagery, though a wide variety of different techniques have been tested including fuzzy classifications [14], neural networks [15,16], support machine vectors [17], post-classification data fusion [18] and OBIA [15,19,20,21]. Studies using only multispectral data have generally reported moderate to poor results. For example, Neukermans et al. [14] report an overall accuracy of 72 percent based on the mapping of four mangrove species and the surrounding land cover using Quickbird multispectral imagery and a fuzzy classification scheme. Similarly, Wang et al. [19] report an overall classification accuracy of nearly 75 percent or less for each of three mangrove species using Quickbird or IKONOS imagery with a maximum likelihood classification (MLC) technique.
The incorporation of spatial information either in the form of OBIA or pixel-based image texture (e.g., grey-level co-occurrence matrix or lacunarity) improves the classification accuracy [15,16,17,19,20]. Spatial information seeks to extract repeated patterns in canopy structure that can be indirectly related to species. This approach has merit as mangrove genera often differ greatly in form and structure [22]. Spatial metrics are very sensitive to edge effects and work best over continuous canopies. In the case of fringe or basin mangroves, mangrove species zonation is often not as distinct as in other environments, and high edge length to area ratio makes edge effects a serious challenge. Thus, to effectively map and monitor fringe mangrove forests, especially at local and regional scales, the challenges of spectral confusion and likely limited effectiveness of spatial metrics are constraining factors. Previous studies have reported a range of classification accuracies. Wang et al. [15] report that a hybrid OBIA-MLC classification outperforms either individual approach, but accuracy for individual species still ranged from 74% to 98%. Both Huang et al. [17] and Myint et al. [20] report accuracies greater than 90% using spatial data as part of the classification, or as an input into the image object segmentation process.

1.2. Study Objective

The objective of this study was to map fringe and basin mangrove forests at the species level. First, an analysis of spectral separability of vegetation using Jeffries-Matusita separability measure was conducted to distinguish between vegetation types or groups and to evaluate the differences between Quickbird and Woldview-2 for multispectral analysis. Based on these results, a hybrid OBIA-SVM approach was designed to enhance vegetation separability. An object-based decision tree classification was used to classify classes other than dense coastal vegetation that are not central to this study. A support vector machine classification was used to classify dense coastal vegetation between true mangroves and mangrove associates. The accuracy of the results was analyzed at the object level and field plot or point level for individual vegetation types.

1.3. Background—Object Based Image Analysis

Pixel-based analysis is generally conceptually simple and methods are generic across sensors. However, pixels are often not the unit of interests, but rather the default unit of measurement. For example, individual crowns and canopy gaps consist of multiple pixels and produce spatial-autocorrelation within objects that can be detected using high resolution imagery [23]. OBIA seeks to create “meaningful” objects by segmenting an image into groups of pixels with similar characteristics based on spectral and spatial properties [24]. In OBIA, segmented objects become the unit of analysis, from which spectral statistics, such as spectral band means and standard deviation, or spatial information, such as image texture, can be used for further analysis including image classification. In the software eCognition, user-defined scale, shape, and compactness parameters make OBIA particularly useful for creating objects with heterogeneous pixels such as desert with sparse vegetation. OBIA has been widely applied for forest remote sensing studies [25,26,27,28] and has been successfully applied to mangrove studies [15,20,29]. However, OBIA has not been explicitly applied to fringe mangroves.

1.4. Background—Support Vector Machine

SVM is a machine-learning technique that is well adapted to solving non-linear, high dimensional space classifications [30]. For remote sensing, SVM is a useful tool for multispectral and hyperspectral classifications in which spectral separability is less than perfect. The mathematical formulation of SVM is described by Vapnik [31] and a detailed assessment of SVM for remote sensing is described by Huang et al. [32]. Though still a novel method for remote sensing, SVM has been applied in many other fields such as biology, biochemistry, and economics. SVM differs from traditional classification approaches by identifying the boundary between classes in n-dimensional spectral-space rather that assigning points to a class based on mean values. SVM creates a hyperplane through n-dimensional spectral-space that separates classes based on a user defined kernel function and parameters that are optimized using machine-learning to maximize the margin from the closest point to the hyperplane. Figure 1 illustrates the difference between a maximum likelihood classification and a SMV. By identifying the hyperplane that separates two classes (represented by the red and blue dots) rather than using the distance between class spectral means (the black dots), SVM can produce a more accurate classification. A penalty parameter allows the SVM to vary the degree of training data misclassified due to possible data error when optimizing the hyperplane. While there are many possible kernels, four common kernels found in remote sensing packages are linear, polynomial, radial basis function, and sigmoid. Finding the best kernel and parameters can be difficult, though Hsu et al. [33] suggest starting with a radial basis function and testing a range of parameters to identify an effective model. In a recent study by Yang [34], it is shown that for most land cover classes, the radial basis function is the best kernel with a penalty parameter of 100.
Figure 1. A stylized example of a cluster-type (left) and support vector machine likelihood classification (right) between two hypothetical classes shown in blue and red.
Figure 1. A stylized example of a cluster-type (left) and support vector machine likelihood classification (right) between two hypothetical classes shown in blue and red.
Remotesensing 03 02440 g001
Several studies have demonstrated the great potential for SVM. Pal and Mather [30] found that SVM outperforms maximum likelihood and artificial neural network classifiers using Landsat TM and is well suited for small training sets and high-dimensional data. Foody and Mathur [35] found SVM outperforms discriminate analysis and decision-tree algorithms for airborne sensor data. Li et al. [36] applied SVM to an OBIA with better results than standard fuzzy logic classification. Only a single study has applied SVM for analysis of mangroves. Huang et al. [17] applied SVM as part of a fusion methodology of spectral and image texture data to map mangroves although the effectiveness of SVM for multispectral classification of mangroves remains untested.

2. Methodology

2.1. Study Area

The research was conducted on Isabela Island in the Galapagos Archipelago, Ecuador. The Galapagos Islands, located 1,000-km off the coast of Ecuador, are an archipelago consisting of 13 large islands, 4 of which have human populations, and 188 small islands and rocks (Figure 2). The Galapagos Islands were declared a national park in 1959 (the park consists of 97% of land area), a UNESCO World Heritage Site in 1978, and a UNESCO Biosphere Reserve in 1987. The Galapagos Islands lie on the western edge of the Atlantic-East Pacific mangrove complex. Mangrove forests consist of three true species common in this region: Rhizophora mangle (red), Avicennia germinans (black), and Laguncularia racemosa (white), and as well as the associate species such as Conocarpus erectus (button or buttonwood mangrove) and Hippomane mancinella (manzanillo), or other halophytes growing on nearby sand flats or dunes [37].
Figure 2. Quickbird false color composites for the Puerto Villamil and Cartago study areas on Isabela Island in the Galapagos Islands.
Figure 2. Quickbird false color composites for the Puerto Villamil and Cartago study areas on Isabela Island in the Galapagos Islands.
Remotesensing 03 02440 g002
In the Galapagos Islands, mangrove forest form dense, but small patches in protected coves and lagoons along an otherwise barren or arid coast. Mangrove forests in the study site can be described primarily as fringe mangroves forming along the coastline or basin mangroves along hyper-saline lagoons. Mangroves grow on a range of substrates from aa lava to sand or silty-clay. For a more detailed description of the arid coastal environment in the Galapagos Islands, see Van der Werff and Andersen [37].
Figure 3. Examples of vegetation near Puerto Villamil (from upper left, clockwise): (A) tall black mangroves near a fresh water spring, (B) red mangroves growing on lava shoreline, (C) mixed arid vegetation and mangroves along a hyper-saline pond, (D) tall red mangroves mixed with white and black mangroves on a saline pond.
Figure 3. Examples of vegetation near Puerto Villamil (from upper left, clockwise): (A) tall black mangroves near a fresh water spring, (B) red mangroves growing on lava shoreline, (C) mixed arid vegetation and mangroves along a hyper-saline pond, (D) tall red mangroves mixed with white and black mangroves on a saline pond.
Remotesensing 03 02440 g003
This study focuses on two study areas on Isabela Island: Puerto Villamil and Cartago (Figure 2). The Puerto Villamil study site is located on the southern end of Isabela Island extending west from the town of Puerto Villamil. The study area contains some features unique to the Galapagos Islands including the largest lagoon complex in the Galapagos Islands, the longest sand beach, and complex geologic topography along the coastline. Figure 3 shows several examples of the mangroves in different settings from the study area. Salinity varies greatly across the study site as both fresh water springs and hyper-saline ponds occur in relatively close proximity. Field observations show that while mangrove species form patterns of zonation based on salinity and/or wave action, the mangrove and associate species co-occur in close proximity due to micro-topographical geological features (i.e., lava coastline). To the west of Puerto Villamil, the elevation increases quickly away from the shoreline towards the Sierra Negra or Cerro Azul volcanoes and the vegetation changes from barren/arid to semi-arid/semi-humid along this elevation transition. It is important to note that unlike large riverine mangrove forests like those along the coast of mainland Ecuador, the pattern of zonation between mangrove and mangrove associate species is truncated and highly variable due to the small inter-tidal zone and the geologic rather the fluvial coastal geomorphology. The Cartago study area is located on the eastern edge of Isabela Island. This area has the largest mangrove forest patches in the Galapagos. Unlike the Puerto Villamil study area, Cartago lacks lagoons or vegetation away from the coast as the study area lies on a relatively flat lava field to the east of Sierra Negra.

2.2. Field Data

Field data were collected during the summer of 2009 near the town of Puerto Villamil. Due to conservation policies within the Galapagos National Park, non-destructive sampling was required. Mangroves form stands with dense aerial roots and branches, making many areas inaccessible. An opportunistic sampling scheme was conducted due to logistical constraints and efforts were made to sample a wide range of conditions for each species (Table 1). A wide range of conditions were sampled from lava to sand substrates, fresh water springs to hypersaline ponds, and short shrubs to trees over 20 m tall. Canopy height, substrate conditions, and mangrove species were recorded at nine points for 48, 10-m diameter plot. Plot location was recorded using a Trimble GeoXT GPS unit and differentially corrected to a 95% horizontal positional accuracy of less than 1.5 m. To extend the extent of the sampled area, an additional 481 species and height point measurements were collected Point locations were measured using a compass and laser range finder from a known GPS position. Due to the limited accuracy of the analog compass (±1°), a maximum of 100 m from the observer was set for all points collected. All field data points are considered representative for a 3-m diameter circle.
Table 1. Vegetation field data by species at plots and individual points.
Table 1. Vegetation field data by species at plots and individual points.
SpeciesPlots*PointsTotalPercent
MA (Mangrove Associates)AC (Acacia)827353.472
MZ (Manzanillo)437504.960
OV (Other Vegetation)1617333.274
BW (Buttonwood)565511111.012
TM (True Mangroves)RM (Red Mangrove)12017429429.167
WM (White Mangrove)14624338938.591
BM (Black Mangrove)6630969.524
Total4555531,008

2.3. Remote Sensing Data

Details of the Quickbird and Worldview-2 imagery are shown in Table 2. The Quickbird imagery was cloud-free over coastal areas, while the Worldview-2 imagery had a few clouds over the study area. Thus, the Quickbird imagery was used for the first level of analysis. All imagery was geometrically corrected using the ENVI Rational polynomial coefficients (RPC) with ground control points (GCP) orthorectification correction algorithm. Since all mangroves grow within the inter-tidal zone, the elevation was assumed to be at mean sea-level across the image. The root mean square error (RMSE) was found to be less than 1.5 m using 16 independent GCPs. All imagery was radiometrically corrected using a Dark Object Subtraction. Since consistent dark objects could not be identified between images, a 1% threshold value for each band was used. Solar angle was not found to be substantially different between images. All imagery was resampled to a resolution of 2 m using a cubic convolution interpolation. Several band ratios were computed to assist with classification. Band ratios were selected based on exploratory analysis using visual interpretation and the feature optimization tool in eCognition. The selected band ratios were NIR/Red (i.e., Simple Ratio) and NIR/Blue for the Quickbird imagery, and NIR2/Red, Red Edge/Green, and Yellow/Coastal Blue for the Worldview-2 imagery.
Table 2. Imagery details for the Quickbird and Worldview-2 sensors.
Table 2. Imagery details for the Quickbird and Worldview-2 sensors.
SensorQuickbirdWorldview-2
Acquisition Date27 August 200801 October 2010
Spatial Resolution (m)
Pan0.60.5
MSS2.42.0
Spectral Channels (nm)
Coastal BlueN/A400–450
Blue450–520450–510
Green520–600510–580
YellowN/A585–625
Red630–690630–690
Red EdgeN/A705–745
NIR-1760–900770–895
NIR-2N/A860–1,040

2.4. Spectral Separability

Spectral separability analysis compares the spectral signature of classes and determines the degree to which those classes can be distinguished. Spectral separability analysis is a commonly used exploratory analysis approach for selecting classes and training data for classification. Spectral separability was calculated using Jeffries-Matusita (J-M) Distance that measures the divergence between spectral means [38,39]. The J-M distances in ENVI is squared so that the distance values range from 0 to 2, where values greater than 1.9 are highly separable, and value less than 1.0 require class clumping or new training data for traditional mean-based classification methods.

2.5. Object-Based Image Analysis

Image segmentation and decision-tree classification were conducted using eCognition Developer 8. eCognition groups pixels based on spectral and spatial properties [24]. A two-level segmentation was used to first classify general land cover classes, and then refine the coastal vegetation classes. The first-level segmentation used shape = 0.5, compactness = 0.5, and scale = 25. The object size varied based on the heterogeneity of the area and the land cover class. For example, lava objects had a mean area of 656 m2, with a standard deviation of 625 m2 compared ocean objects had a mean area of 1,385 m2 with a standard deviation of 1,926 m2. A second image level was segmented based on the Worldview-2 imagery using shape = 0.5, compactness = 0.9, and scale = 10 for only the dense coastal vegetation classification from level 1. The mean level 1 object size was 144 m2 with a standard deviation of 102 m2.

2.5.1. Decision-Tree Classification

The decision tree classification is shown in Figure 4. Class rules were identified using interactive visual interpretation of threshold values based on training data, existing map, and expert knowledge of the study area. Upland and coastal vegetation were separated using a distance rule of 250 m from open water based on field observations. While there was little confusion between these general land cover classes, there was considerable confusion between lava and shallow water over lava (e.g., ponds, coastline). In these cases, objects were manually edited using expert image interpretation.
Figure 4. Object-based image analysis (OBIA) Decision Tree (rectangle = image; diamond = rule; oval = class).
Figure 4. Object-based image analysis (OBIA) Decision Tree (rectangle = image; diamond = rule; oval = class).
Remotesensing 03 02440 g004
Although overall cloud cover was less than 15% in the Worldview-2 image, clouds and cloud-shadow were classified and removed from the vegetation analysis. The remaining dense coastal vegetation objects were exported to ArcGIS 9.3.1 with the mean values of each band and band ratio for each object. In ArcGIS, the shapefiles were converted to raster stacks for analysis in ENVI 4.8. It should be noted that during exploratory analysis, object-level standard deviation and texture (i.e., grey-level co-occurrence matrix) were also calculated, but they were not found to substantially improve classification results, while raising concerns of model over-fitting with higher data dimensionality. Additionally, the eCognition fuzzy nearest neighbor classification, using the feature optimization tool to select input data for the classification, did not produce acceptable results for the true mangrove classification. When using only mean spectral information, there was insufficient separability between classes. The addition of standard deviation or skewness of spectral data or image texture, separability increased, but classification results showed strong overfitting of the classification to training data.

2.5.2. Support Vector Machine Classification

The SVM classification was conducted using ENVI. Calibration and validation objects were selected based on field data; homogenous objects were verified through visual assessment. The distribution of the objects is shown in Table 3. The objects were systematically divided between calibration and validation datasets based on the object ID created systematically across the image during segmentation. This ensured an equivalent geographic distribution of calibration and validation data. An SVM radial basis function (RBF) kernel was applied using the default parameters (gamma = 0.091 and a penalty parameter of 100). The penalty parameter is particularly important for non-separable classes. Equation (1) shows the RBF kernel:
K(xi,xj) = exp(−g||xi − xj||2), g > 0
where g is the user-defined gamma.
ENVI conducts pair-wise iterations of SVM and assigns fuzzy class membership. Classes are assigned using the highest membership. Exploratory analysis did not show improved results with other gamma or penalty values.
Table 3. Distribution of objects used to calibrate and validate the support vector machine (SVM) classification.
Table 3. Distribution of objects used to calibrate and validate the support vector machine (SVM) classification.
True MangrovesMangrove AssociatesTotal
Calibration14354197
Validation7324101
Total21678298

2.6. Accuracy Assessment

The accuracy of the SVM classification was assessed in several ways. First, accuracy is assessed at the object-level using an error confusion matrix. The overall, producer’s, and user’s accuracy was calculated, in addition to the kappa statistic. The area under the curve (AUC) of the receiver operating characteristic (ROC) was also computed based on fuzzy membership [40]. This statistic illustrates the accuracy of the classification relative to a perfect classification (AUC = 1) and a random classification (AUC = 0.5) based on the rate of false positives. Second, an error confusion matrix was created for the individual vegetation types at the field point level and all classes from both the decision tree and SVM classification. To further illustrate the relationship between the field data and SVM, a boxplot distribution of fuzzy membership to true mangroves was computed for each field vegetation class.

3. Results and Discussion

3.1. Spectral Separability Analysis

Class spectral separability at the pixel level for all vegetation field data points is shown in Table 4 for the Quickbird (A) and Woldview-2 (B) imagery. The spectral separability between vegetation classes for the Quickbird imagery was moderate to poor. Not a single value was found to be greater than the suggested threshold of 1.9, though many values were greater than 1.8. More importantly, there was not a consistently high separability for any individual species. This indicates that the ability to discriminate vegetation types with high accuracy using Quickbird imagery is very unlikely. Although the Worldview-2 imagery had better spectral separability than the Quickbird imagery (Table 3(C)), likely due to the greater number of spectral bands, only manzanillo (MZ) was consistently separable from mangroves. Separability between mangrove species was particularly low, especially between red and white mangroves. This result is consistent with field measurements taken from a handheld spectroradiometer during the field campaign [41]. The spectral overlap and confusion between species was consistent with the accuracy assessment of previous studies using various classification techniques. Neukermans et al. [14] reported an overall accuracy of 72% using a fuzzy classification and Wang et al. [19] reported an overall accuracy of 75% using a maximum likelihood technique, although the user's accuracy for some species was as low as 55%. During the exploratory analysis of this study, a MLC classification failed to detect two separate true mangrove and mangrove associate classes.
Table 4(D) shows spectral separability of true mangroves (TM) and mangrove associates (MA) for all vegetation field points and dense vegetation objects from OBIA segmentation and decision tree classification. In most cases, spectral separability increased with the inclusion of band ratios and spatial information through image segmentation into objects. However, the maximum value of 1.665 demonstrates that there is considerable spectral overlap between the two classes and that non-traditional classification methods are likely required (i.e., SVM). The use of an object-based approach worked best for dense vegetation objects. Results for sparse vegetation objects (not shown) did not improve the separability results and were sometimes worse than the pixel-based analysis due to the inclusion of background substrate reflectance. Moreover, the moderate to poor spectral separability indicates the in ability to discern between species using this imagery which includes noise from non-leaf surfaces such as braches and background substrate. Given, pure leaf reflectance from finer scale imagery or spectral unmixing, spectral separability may be higher.
Table 4. Spectral Separability (Jeffries-Matustia Distance) for individual species using Quickbird (A) or Worldview-2 (B) imagery, the difference between the Quickbird and Worldview-2 imagery (C) and a comparison between pixels and objects for both types of imagery using true mangrove and mangrove associate classes (D).
(A) Quickbird
ACMZOVBWRMWMBM
AC1.8921.3421.4551.8371.7031.543
MZ1.8921.7341.6901.5931.8141.355
OV1.3421.7340.9941.7251.6731.233
BW1.4551.6900.9941.2561.1850.702
RM1.8371.5931.7251.2560.5081.129
WM1.7031.8141.6731.1850.5081.258
BM1.5431.3551.2330.7021.1291.258
(B) Worldview-2
ACMZOVBWRMWMBM
AC1.9631.6471.6981.8201.4981.785
MZ1.9631.8611.9251.9001.9431.647
OV1.6471.8611.5321.6221.5841.381
BW1.6981.9251.5321.6171.3361.634
RM1.8201.9001.6221.6170.8661.226
WM1.4981.9431.5841.3360.8661.540
BM1.7851.6471.3811.6341.2261.540
(C) Difference
ACMZOVBWRMWMBM
AC0.0710.3040.243−0.017−0.2050.242
MZ0.0710.1270.2350.3070.1290.292
OV0.3040.1270.537−0.103−0.0890.148
BW0.2430.2350.5370.3610.1510.932
RM−0.0170.307−0.1030.3610.3580.097
WM−0.2050.129−0.0890.1510.3580.282
BM0.2420.2920.1480.9320.0970.282
(D) Pixel vs. Object for TM and MA
(D) Pixel vs. Object for TM and MA
QBQB w/BRWVWV w/ BRQB = Quickbird
All Veg Points0.6641.1410.7341.084WV = Worldview-2
Dense Veg Objects0.8391.1181.3211.665BR = Band Ratios

3.2. Classification

The classification is illustrated in Figure 5. Table 5 shows the proportion of each land cover type. For both study areas, lava and ocean are the dominant cover types and coastal vegetation comprises about 5.5 km2 or 12% and 8% of Puerto Villamil and Cartago images, respectively. The composition of coastal vegetation differs between the two study areas. The Puerto Villamil study area is mostly sparse vegetation with mangrove associates and true mangroves comprising a minority of coastal vegetation. The Cartago study area is mostly true mangroves with much less sparse vegetation and almost no mangrove associate species present. However, much of the dense coastal vegetation was obstructed by clouds or cloud shadow in the Worldview-2 image (24.8%).
Figure 5. Land cover classification for Puerto Villamil (top) and Cartago (bottom).
Figure 5. Land cover classification for Puerto Villamil (top) and Cartago (bottom).
Remotesensing 03 02440 g005
The satellite classification has shown that true mangroves are widespread and the dominant vegetation cover in the Cartago study area, while true mangroves are part of a wider range of vegetation in the Puerto Villamil study area. In both study areas, mangroves grow along the sheltered coastline and thrive where there is likely subsurface freshwater from the humid highlands emerging along the coast as springs. Several of these springs are found near the town of Puerto Villamil and these reflect the large dense mangrove patches observed.
Table 5. Land Cover Classification for all classes (A) and coastal vegetation (B).
(A) 
(A) 
Puerto VillamilCartago
CoverAreaPercentAreaPercent
OC17.930438.146329.224943.4184
PD1.54003.2763*
LV14.115030.029132.451148.2114
SD0.80321.70870.03970.0589
UV6.849114.57120.27220.4043
QBC0.09850.20960.00000.0000
SCV2.87966.12621.25521.8648
DCVC0.17850.37981.01191.5033
MA1.50913.21060.00060.0009
TM1.10102.34233.05444.5379
Total47.00412.058867.3107.9069
(B) Coastal vegetation classes and total.
(B) Coastal vegetation classes and total.
Puerto VillamilCartago
CoverAreaPct CVAreaPct CV
SCV2.87960.50801.25520.2358
DCVC0.17850.03151.01190.1901
MA1.50910.26620.00060.0001
TM1.10100.19423.05440.5739
Total5.6685.322
OC = Ocean  UV = Upland Vegetation
PD = Pond  QBC = Quickbird Clouds
LV = Lava  SCV = Sparse Coastal Veg
SD = Sand  DCVC = Dense Coast Vegetation w/Clouds
TM = True Mangrove MA = Mangrove Associates
The differences in land cover reflect the differences in the climatic and geomorphic environment. The Puerto Villamil study area is along the southern edge of the Sierra Negra volcano and has considerably more cloud cover during the year than the Cartago study area (unpublished MODIS data). Furthermore, elevation increases rapidly from the coast to the area west of Puerto Villamil, where mists and fog increase with elevation providing moisture to plants. In the Cartago study area, the elevation remains near sea-level with little available moisture, as observed from the barren lava beds, except along the coast where there are likely isolated fresh water springs fed by rain in the humid highlands on Sierra Negra.
Another major difference between the two study areas is the presence of ponds and lagoons. The Puerto Villamil study area contains the largest lagoon system in the Galapagos Islands. The hydrologic connectivity of these lagoons is complex as some lagoons are hyper-saline and others are nearly fresh water [41]. The ranges of hydrologic conditions near Puerto Villamil are likely the cause of the range of vegetation types (i.e., true mangroves vs. mangrove associates) and vegetation conditions (i.e., LAI and canopy height) observed. As Song et al. [42] observed in this study area, salinity can have an observable impact on remote sensing-derived photosynthetic productivity.
The lagoon complex is the result of volcanic topographic features seemingly unique to that area. In contrast, the structure of the Cartago coastline reflects a more fluvial pattern of inter-tidal channels. The relatively simple topography and hydrology and more arid environment near Cartago has led to isolated but large, dense mangrove patches around protected coves and likely fresh water springs. Future research is needed to investigate the link between hydrologic conditions including subsurface flow and coastal vegetation.

3.3. Accuracy Assessment

The accuracy assessment was considered at two levels: (1) validation objects for a typical assessment of just the SVM classification, and (2) validation field points to understand the accuracy from the decision tree classification and the sub-object level.

3.3.1. SVM

The overall accuracy of the SVM classification between true mangroves and mangrove associates was 94.4% with a kappa statistic of 0.863. The greatest source of error was the misclassification of mangrove associates as true mangroves (Table 6). The producer’s and user’s error were consistent for each class and greater than 90% in all cases. The AUC-ROC was 0.991 for true mangroves and 0.987 for mangrove associates. The overall accuracy of the classification was very good and better than most previous mangrove studies [4] and thus demonstrates the ability of this approach to accurately distinguish between true and associate mangroves in fringe and basin environments.
Table 6. Classification confusion matrix (A) and classification accuracy (B) of SVM classification.
(A)TM (Pix)TM (%)MA (Pix)MA (%)Total
Unclassified00000
TM3,15296.041289.693,280
MA1303.961,19390.311,323
Total3,2821001,3211004,603
(B)Prod. (Pix)Prod.(%)User (Pix)User (%)
TM3,152/3,28296.043,152/3,28096.1
MA1,193/1,32190.311,193/1,32390.17

3.3.2. Total Classification vs. Field Points

Table 7 shows the distribution of validation vegetation field data points (i.e., points that did not fall within objects used for the SVM classification). The majority of field points were classified either as sparse coastal vegetation or lava. This is indicative of the arid environment of the study area, particularly in an OBIA classification framework where the spectral signature of the object and not individual pixels are used for classification. Furthermore, given the sparse nature of the vegetation, the spectral signal strength of these vegetation points was relatively weak, preventing further classification with detailed spectral unmixing, requiring hypersepctral data. Of the true mangrove species, about 50% of red and black mangroves points were classified as sparse vegetation, while nearly all of the white mangrove points were classified as sparse. It should be noted that since the field sampling scheme was opportunistic due to Galapagos National Park regulations about cutting vegetation, sampling was likely biased towards less dense and more accessible vegetation, especially for the field plots.
Table 7. Accuracy Assessment of Decision Tree and SVM Classification from Field Data by number of points (A) and percent of points (B).
Remote Sensing Classes
Field Data Classes(A)TMMAQBCSCVDCVCLVOCPDSDUVTotal
AC20018011000031
MZ5270200000034
OV49014010000037
BW186052012040092
RM84807102311800205
WM431015901207700337
BM92201606000053
Total165730332018282900789
Remote Sensing Classes
Field Data Classes(B)TMMAQBCSCVDCVCLVOCPDSDUVTotal
AC6.450.000.0058.060.0035.480.000.000.000.00100
MZ14.7179.410.005.880.000.000.000.000.000.00100
OV10.8124.320.0037.840.0027.030.000.000.000.00100
BW19.576.520.0056.520.0013.040.004.350.000.00100
RM40.983.900.0034.630.0011.220.498.780.000.00100
WM12.760.300.0047.180.0035.612.082.080.000.00100
BM16.9841.510.0030.190.0011.320.000.000.000.00100
Total0.2090.09300.42100.2310.0100.037001
Figure 6 shows the fuzzy SVM true mangrove classification distribution for each vegetation type. Manzanillo, other vegetation, red mangrove, and white mangrove had appropriate membership to true mangroves. However, black mangrove and buttonwood had low and high membership, respectively, indicating misclassification of these vegetation types.
There are many aspects to this error. First, there is a spatial scale mismatch between the field points and the objects in that many points may occur in a single object. For example, field plots dominated by black mangroves (~11 of 29 dense black mangrove points) were misclassified as mangrove associates due to the misclassification of a single object. Moreover, single points may not be representative of a whole object, not to mention the inherent geometric error in both the field data and remote sensing imagery that may results in a data classification mismatch. Second, fuzzy classification or mixed classes were not explicitly considered. Only pure objects were used for calibrating the classification as the exact composition of mixed objects was unknown due to the lack of a tree census. While mangroves often have detectable patterns of zonation, this is not always observed [43]. This was certainly the case in some parts of the study area where edaphic and topographic conditions changed rapidly over small distance such that for a given field plot, multiple species were present.
Figure 6. Boxplot of true mangrove (TM) fuzzy membership for validation field points.
Figure 6. Boxplot of true mangrove (TM) fuzzy membership for validation field points.
Remotesensing 03 02440 g006
This second point demonstrates a gap in the current knowledge of methods in remote sensing. While there are several papers that assess methods of image segmentation and object classification [44], there is not a good assessment of linking field sampling schemes with OBIA of natural landscapes where visual interpretation is not as straight forward as human landscapes (e.g., buildings, roads, impervious surfaces). Field sampling protocols for remote sensing have been largely designed for pixel-based analysis from a legacy of 25-m pixels from Landsat and SPOT. The type of sampling for pixel based analysis does not lend itself to assessing whole objects created after field data collection, especially fuzzy membership of heterogeneous objects. Two alternative sampling schemes such as large-scale quadrat sampling or mapping boundaries of homogeneous patches may be more appropriate of OBIA, but this type of field data collection is difficult and time-consuming in the best of conditions, let alone in dense mangrove swamps. Another alternative is the visual interpretation of ultra-high resolution (i.e., pixel size less than 5 cm) airborne imagery for a small subset of the study area to generate calibration/validation data. Recent advancements in radio-controlled helicopter platforms are particularly promising for this application. Furthermore, this type of sampling would be prohibitively destructive in the Galapagos National Park. Future research is needed to assess effective and efficient field sampling schemes for use with OBIA.

4. Conclusions

Effective monitoring and management of mangrove forests requires accurate and repeatable measures of forest extent and species composition. While previous studies have successfully mapped mangrove extent and species, these studies have largely ignored fringe mangroves. This study has addressed this issue. Spectral separability analysis revealed that the spectral signatures between mangrove species and even associate species were moderately separable using Quickbird or Worldview-2 imagery at both the pixel and object level. The best separability was found using dense vegetation objects, indicating that even for very high resolution imagery, the multispectral signature of non-vegetation components for sparse vegetation produce mixed pixel effects that seriously limit multispectral analysis. Using a hybrid decision-tree and SVM approach, true mangrove species and associate mangrove species were classified with an accuracy of 94% at the object level. However, when the classification was assessed at the species level, the accuracy was poor for some species such as black and buttonwood mangrove, which had accuracies of 29% and 25%, respectively. This research demonstrates that while non-linear machine-learning classification techniques such as support vector machine, in combination with OBIA of very-high spatial resolution data, can provide marked improvement in the classification of vegetation types, there is a necessity for greater spectral resolution to distinguish between the subtle differences between individual species. Given these findings, future research should focus on hyperspectral image analysis to improve spectral separability between species and LiDAR to enhance image segmentation based on canopy structure as well as spectral properties.

Acknowledgements

This research was supported by a Doctoral Dissertation Research Improvement Grant from the National Science Foundation (Award 0927164), the Center for Galapagos Studies and the Department of Geography at the University of North Carolina at Chapel Hill. The Galapagos National Park and Charles Darwin Research Station provided support in the field. Special thanks to Stephen J. Walsh, Conghe Song, Aaron Moody, Dean Urban, George P. Malanson, and the anonymous reviewers for comments on the manuscript.

References

  1. Hogarth, P. The Biology of Mangroves and Seagrasses; Oxford University Press: Oxford, UK; New York, NY, USA, 2007. [Google Scholar]
  2. Alongi, D.M. Present state and future of the world’s mangrove forests. Environ. Conserv. 2002, 29, 331. [Google Scholar] [CrossRef]
  3. Giri, C.; Ochieng, E.; Tieszen, L.L.; Zhu, Z.; Singh, A.; Loveland, T.; Masek, J.; Duke, N. Status and distribution of mangrove forests of the world using earth observation satellite data. Global Ecol. Biogeogr. 2011, 20, 154–159. [Google Scholar] [CrossRef]
  4. Heumann, B. Satellite remote sensing of mangrove forests: Recent advances and future opportunities. Progr. Phys. Geogr. 2011, 35, 87–108. [Google Scholar] [CrossRef]
  5. Kuenzer, C.; Bluemel, A.; Gebhardt, S.; Quoc, T.V.; Dech, S. Remote sensing of mangrove ecosystems: A review. Remote Sens. 2011, 3, 878–928. [Google Scholar] [CrossRef]
  6. Al Habshi, A.; Youssef, T.; Aizpuru, M.; Blasco, F. New mangrove ecosystem data along the UAE coast using remote sensing. Aquat. Ecosyst. Health Manage. 2007, 10, 309–319. [Google Scholar] [CrossRef]
  7. Benfield, S.L.; Guzman, H.M.; Mair, J.M. Temporal mangrove dynamics in relation to coastal development in Pacific Panama. J. Environ. Manage. 2005, 76, 263. [Google Scholar] [CrossRef] [PubMed]
  8. Gao, J. A hybrid method toward accurate mapping of mangroves in a marginal habitat from SPOT multispectral data. Int. J. Remote Sens. 1998, 19, 1887–1899. [Google Scholar] [CrossRef]
  9. Simard, M.; De Grandi, G.; Saatchi, S.; Mayaux, P. Mapping tropical coastal vegetation using JERS-1 and ERS-1 radar data with a decision tree classifier. Int. J. Remote Sens. 2002, 23, 1461–1474. [Google Scholar] [CrossRef]
  10. Manson, F.J.; Loneragan, N.R.; McLeod, I.M.; Kenyon, R.A. Assessing techniques for estimating the extent of mangroves: Topographic maps, aerial photographs and Landsat TM images. Mar. Freshwater Res. 2001, 52, 787–792. [Google Scholar] [CrossRef]
  11. Ramsey, E.W.; Jensen, J.R. Remote sensing of mangrove wetlands: Relating canopy spectra to site-specific data. Photogramm. Eng. Remote Sensing 1996, 62, 939. [Google Scholar]
  12. Vaiphasa, C.; Ongsomwang, S.; Vaiphasa, T.; Skidmore, A.K. Tropical mangrove species discrimination using hyperspectral data: A laboratory study. Estuar. Coast. Shelf Sci. 2005, 65, 371–379. [Google Scholar] [CrossRef]
  13. Wang, L.; Sousa, W.P. Distinguishing mangrove species with laboratory measurements of hyperspectral leaf reflectance. Int. J. Remote Sens. 2009, 30, 1267–1281. [Google Scholar] [CrossRef]
  14. Neukermans, G.; Dahdouh-Guebas, F.; Kairo, J.G.; Koedam, N. Mangrove species and stand mapping in GAzi bay (Kenya) using Quickbird satellite imagery. J. Spat. Sci. 2008, 53, 75–86. [Google Scholar] [CrossRef]
  15. Wang, L.; Sousa, W.P.; Gong, P. Integration of object-based and pixel-based classification for mapping mangroves with IKONOS imagery. Int. J. Remote Sens. 2004, 25, 5655–5668. [Google Scholar] [CrossRef]
  16. Wang, L.; Silvan-Cardenas, J.L.; Sousa, W.P. Neural network classification of mangrove species from multi-seasonal ikonos imagery. Photogramm. Eng. Remote Sensing 2008, 74, 921–927. [Google Scholar] [CrossRef]
  17. Huang, X.; Zhang, L.P.; Wang, L. Evaluation of morphological texture features for mangrove forest mapping and species discrimination using multispectral IKONOS imagery. IEEE Geosci. Remote Sens. Lett. 2009, 6, 393–397. [Google Scholar] [CrossRef]
  18. Vaiphasa, C.; Skidmore, A.K.; de Boer, W.F. A post-classifier for mangrove mapping using ecological data. ISPRS J. Photogramm. 2006, 61, 1–10. [Google Scholar] [CrossRef]
  19. Wang, L.; Sousa, W.P.; Gong, P.; Biging, G.S. Comparison of IKONOS and QuickBird images for mapping mangrove species on the Caribbean coast of Panama. Remote Sens. Environ. 2004, 91, 432–440. [Google Scholar] [CrossRef]
  20. Myint, S.W.; Giri, C.P.; Le, W.; Zhu, Z.L.; Gillette, S.C. Identifying mangrove species and their surrounding land use and land cover classes using an object-oriented approach with a lacunarity spatial measure. GISci. Remote Sens. 2008, 45, 188–208. [Google Scholar] [CrossRef]
  21. Krause, G.; Bock, M.; Weiers, S.; Braun, G. Mapping land-cover and mangrove structures with remote sensing techniques: A contribution to a synoptic GIS in support of coastal management in North Brazil. Environ. Manage. 2004, 34, 429–440. [Google Scholar] [CrossRef] [PubMed]
  22. Tomlinson, P.B. The Botany of Mangroves; Cambridge University Press: Cambridge, UK; New York, NY, USA, 1986; p. 413. [Google Scholar]
  23. Woodcock, C.E.; Strahler, A.H. The factor of scale in remote sensing. Remote Sens. Environ. 1987, 21, 311–332. [Google Scholar] [CrossRef]
  24. Benz, U.C.; Hofmann, P.; Willhauck, G.; Lingenfelder, I.; Heynen, M. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. 2004, 58, 239–258. [Google Scholar] [CrossRef]
  25. Chubey, M.S.; Franklin, S.E.; Wulder, M.A. Object-based analysis of IKONOS-2 imagery for extraction of forest inventory parameters. Photogramm. Eng. Remote Sensing 2006, 72, 383–394. [Google Scholar] [CrossRef]
  26. Desclee, B.; Bogaert, P.; Defourny, P. Forest change detection by statistical object-based method. Remote Sens. Environ. 2006, 102, 1–11. [Google Scholar] [CrossRef]
  27. Hay, G.J.; Castilla, G.; Wulder, M.A.; Ruiz, J.R. An automated object-based approach for the multiscale image segmentation of forest scenes. Int. J. Appl. Earth Obs. Geoinf. 2005, 7, 339–359. [Google Scholar] [CrossRef]
  28. Wulder, M.A.; White, J.C.; Hay, G.J.; Castilla, G. Towards automated segmentation of forest inventory polygons on high spatial resolution satellite imagery. Forest. Chron. 2008, 84, 221–230. [Google Scholar] [CrossRef]
  29. Conchedda, G.; Durieux, L.; Mayaux, P. Object-Based Monitoring of Land Cover Changes in Mangrove Ecosystems of Senegal. In Proceedings of 4th International Workshop on the Analysis of Multi-Temporal Remote Sensing Images, Louvain, Belgium, 18–20 July 2007; pp. 44–49.
  30. Pal, M.; Mather, P.M. Support vector machines for classification in remote sensing. Int. J. Remote Sens. 2005, 26, 1007–1011. [Google Scholar] [CrossRef]
  31. Vapnik, V. The Nature of Statistical Learning Theory; Springer: New York, NY, USA, 1995. [Google Scholar]
  32. Huang, C.; Davis, L.S.; Townshend, J.R.G. An assessment of support vector machines for land cover classification. Int. J. Remote Sens. 2002, 23, 725–749. [Google Scholar] [CrossRef]
  33. Hsu, C.; Cahng, C.; Lin, C. A Practical Guide to Support Vector Classification; Department of Computer Science, National Taiwan University: Taipei, Taiwan, 2010. [Google Scholar]
  34. Yang, X. Parameterizing Support Vector Machines for Land Cover classification. Photogramm. Eng. Remote Sensing 2011, 77, 27–38. [Google Scholar] [CrossRef]
  35. Foody, G.M.; Mathur, A. The use of small training sets containing mixed pixels for accurate hard image classification: Training on mixed spectral responses for classification by a SVM. Remote Sens. Environ. 2006, 103, 179–189. [Google Scholar] [CrossRef]
  36. Li, H.T.; Gu, H.Y.; Han, Y.S.; Yang, J.H. Object-oriented classification of high-resolution remote sensing imagery based on an improved colour structure code and a support vector machine. Int. J. Remote Sens. 2010, 31, 1453–1470. [Google Scholar] [CrossRef]
  37. Van Der Werff, H.H.; Adsersen, H. Dry coastal ecosystems of the Galapagos Islands. In Dry Coastal Ecosystems: Africa, America, Asia and Oceania; van der Maarel, E., Ed.; Ecosystems of the World 2B; Elsevier Science Ltd.: Amsterdam, The Netherlands, 1993; pp. 459–475. [Google Scholar]
  38. Schmidt, K.S.; Skidmore, A.K. Spectral discrimination of vegetation types in a coastal wetland. Remote Sens. Environ. 2003, 85, 92–108. [Google Scholar] [CrossRef]
  39. Swain, P. Remote sensing. In The Handbook of Pattern Recognition and Processing; Young, T., Fu, K., Eds.; Academic Press: Orlando, FL, USA, 1986; pp. 613–627. [Google Scholar]
  40. Metz, C.E. Basic principles of ROC analysis. Semin. Nucl. Med. 1978, 8, 283–298. [Google Scholar] [CrossRef]
  41. Song, C.; White, B.; Heumann, B.W. Hyperspectral remote sensing of salinity stress on red (Rhizophora mangle) and white (Laguncularia racemosa) mangroves on Galapagos Islands. Unpublished data.
  42. Song, C.; White, B.; Heumann, B. Hyperspectral remote sensing of salinity stress on red (Rhizophora mangle) and white (Laguncularia racemosa) mangroves on Galapagos Islands. Remote Sens. Lett. 2011, 2, 221–230. [Google Scholar] [CrossRef]
  43. Ward, G.A.; Smith, T.J.; Whelan, K.R.T.; Doyle, T.W. Regional processes in mangrove ecosystems: Spatial scaling relationships, biomass, and turnover rates following catastrophic disturbance. Hydrobiologia 2006, 569, 517. [Google Scholar] [CrossRef]
  44. Blaschke, T.; Lang, S.; Hay, G. Object-Based Image Analysis: Spatial Concepts for Knowledge-Driven Remote Sensing Applications; Springer: Berlin/Heidelberg, Germany, 2008. [Google Scholar]

Share and Cite

MDPI and ACS Style

Heumann, B.W. An Object-Based Classification of Mangroves Using a Hybrid Decision Tree—Support Vector Machine Approach. Remote Sens. 2011, 3, 2440-2460. https://doi.org/10.3390/rs3112440

AMA Style

Heumann BW. An Object-Based Classification of Mangroves Using a Hybrid Decision Tree—Support Vector Machine Approach. Remote Sensing. 2011; 3(11):2440-2460. https://doi.org/10.3390/rs3112440

Chicago/Turabian Style

Heumann, Benjamin W. 2011. "An Object-Based Classification of Mangroves Using a Hybrid Decision Tree—Support Vector Machine Approach" Remote Sensing 3, no. 11: 2440-2460. https://doi.org/10.3390/rs3112440

Article Metrics

Back to TopTop