3. Spectral and backscattering analyses of wetlands using multi-source optical and SAR data

Wetlands are complex landscapes and ecologically share similar characteristics. However, each wetland type contains its own specifications, which can be effectively investigated using various satellite imageries. In this regard, both optical and SAR data are the most common remote sensing data, which have so far proved to be significantly helpful in discriminating wetland species. Numerous types of features can be extracted from multi-source optical and SAR data. However, since all the extracted features cannot be inserted into a classification algorithm, the most important features should be selected for classification. As such, the best optical and SAR satellites, spectral bands, spectral indices, SAR features, SAR channels, backscattering mechanisms, decomposition methods, and textural features can be defined for wetland studies. To this end, various separability measures have already been developed and employed for differentiating wetland classes.

Before separability analysis, several pre-processing steps should be performed on the datasets, the most important being variance analysis of field samples. This should be carried out on both individual classes and class pairs. For this, Eqs. (1) and (2) can be used, respectively.

$$Var = \frac{1}{N-1} \sum\_{i=1}^{N} \left(\mathbf{x}\_i - \boldsymbol{\mu}\right)^2 \tag{1}$$

$$F = \frac{Var\_B}{Var\_W} \tag{2}$$

the high variance of field samples of wetlands, the recommendation is to employ a nonparametric distance. After removing the poor features using variance analyses and obtaining the separability measures that each feature provides, the most effective features are inserted

Table 3 summarizes the results of separability analyses performed by U-test on five wetland classes (bog, fen, marsh, swamp, and shallow water) using multi-source optical (RapidEye,

> Tz: alpha\_s CP: alpha FD: volumescattering CP: anisotropy R2: HV/TP

> N\_derd CP: anisotropy R2: HH/HV R2: HH/TP

L8: NDVI L8: SAVI L8: NIR Brightness A: Red Brightness S2: NDVI

R: Red Edge Brightness R: NDWI R: NIR Brightness S2: Red Edge Brightness S2: NDWI

Table 3. The most important optical (provided in the lower left half of the table) and SAR (provided in the upper right half of the table) features for delineating each pair of wetland class in June and August, respectively (the features are

Bog Fen Marsh Swamp Shallow water

R2: HH/HV

R2: HH/TP R2: HH/HV Anisotropy12 A2: HH/HV

A Collection of Novel Algorithms for Wetland Classification with SAR and Optical Data

A2: HH/HV serd R2: HH/HV R2: HH/TP N\_serd

R2: HV/TP R2: HH/TP A2: HV serd

R: NIR Brightness R: NDWI R: Green Brightness R: Red Brightness A: Green Brightness

A: ASTER A2: ALOS-2 FD: Freeman-Durden

TP: total power

relative difference

Polarization-Asymmetry

R2: HH/HV

CP: anisotropy N\_derd R2: HH/VV FD: volume-scattering

N\_serd

http://dx.doi.org/10.5772/intechopen.80688

119

N\_derd CP: anisotropy R2: HH/VV R2: HH/HV N\_serd

CP: anisotropy S1: VV/HV N\_derd R2: VV/TP R2: VV/HV

N\_derd CP: anisotropy serd N\_serd

NDVI: normalized difference vegetation index

N\_serd: normalized single-bounce eigenvalues

into a classification algorithm to produce a highly accurate wetland map.

S1: HH/HV

Tz: alpha\_s FD: double-bounce CP: entropy S1: HH/HV

A: NIR Brightness R: Green Brightness S2: NDWI A: NDVI A: SAVI

L8: NDVI L8: SAVI L8: NIR Brightness A: Red Brightness A: NDWI

R: Red Edge Brightness R: NDWI R: NIR Brightness S2: NDWI S2: Red Edge Brightness

R: RapidEye R2: RADARSAT-2 CP: Cloude-Pottier NIR: near infrared

relative difference

index

NDWI: normalized difference water

serd: single-bounce eigenvalues

Bog CP: alpha

Brightness A: NDWI R: Red Edge Brightness L8: NIR Brightness S2: Red Edge Brightness

Brightness A: NDWI S2: Red Edge Brightness S2: NDWI R: NIR Brightness

Brightness L8: NIR Brightness L8: Green Brightness S2: NDVI S2: SAVI

Brightness A: NDWI R: Red Edge Brightness R: Green Brightness R: NDWI

Fen A: Green

Marsh A: Green

Swamp S2: Red Edge

Shallow water A: Green

N\_derd: normalized doublebounce eigenvalues relative

ordered based on their separability measures).

L8: Landsat-8 S2: Sentinel-2A S1: Sentinel-1 Tz: Touzi SAVI: soil adjusted vegetation index

difference

in which, xi indicates the value of a field sample; μ is the mean value of samples; N is the number of field samples in a feature; F indicates the Fisher-test; and VarB and VarW indicate the between and within variance values in each class pair, respectively. These two variance analyses are more important in the case of wetlands because they are complex environments, and thus, the field samples collected for a wetland class can contain high variance in satellite imagery, especially those acquired by the SAR systems. Figure 2 illustrates an optical spectral band and a SAR feature, for which the variations of field samples are high, and consequently, they should be removed before separability analyses as noisy and poor features.

So far, different separability measures have been developed, which can generally be classified into two categories: parametric and non-parametric. Unlike parametric methods (e.g. t-test), non-parametric techniques, such as Mann-Whitney U-test, do not assume a normal distribution of the samples and evaluate the separability of samples by their ranks [92]. Considering

Figure 2. Spectral and Backscattering values for field samples for two types of wetlands: (a) Fen, and (b) Shallow water.

the high variance of field samples of wetlands, the recommendation is to employ a nonparametric distance. After removing the poor features using variance analyses and obtaining the separability measures that each feature provides, the most effective features are inserted into a classification algorithm to produce a highly accurate wetland map.

various satellite imageries. In this regard, both optical and SAR data are the most common remote sensing data, which have so far proved to be significantly helpful in discriminating wetland species. Numerous types of features can be extracted from multi-source optical and SAR data. However, since all the extracted features cannot be inserted into a classification algorithm, the most important features should be selected for classification. As such, the best optical and SAR satellites, spectral bands, spectral indices, SAR features, SAR channels, backscattering mechanisms, decomposition methods, and textural features can be defined for wetland studies. To this end, various separability measures have already been developed and

Before separability analysis, several pre-processing steps should be performed on the datasets, the most important being variance analysis of field samples. This should be carried out on both

> X N

xi � <sup>μ</sup> � �<sup>2</sup> (1)

(2)

i¼1

<sup>F</sup> <sup>¼</sup> VarB VarW

in which, xi indicates the value of a field sample; μ is the mean value of samples; N is the number of field samples in a feature; F indicates the Fisher-test; and VarB and VarW indicate the between and within variance values in each class pair, respectively. These two variance analyses are more important in the case of wetlands because they are complex environments, and thus, the field samples collected for a wetland class can contain high variance in satellite imagery, especially those acquired by the SAR systems. Figure 2 illustrates an optical spectral band and a SAR feature, for which the variations of field samples are high, and consequently,

So far, different separability measures have been developed, which can generally be classified into two categories: parametric and non-parametric. Unlike parametric methods (e.g. t-test), non-parametric techniques, such as Mann-Whitney U-test, do not assume a normal distribution of the samples and evaluate the separability of samples by their ranks [92]. Considering

Figure 2. Spectral and Backscattering values for field samples for two types of wetlands: (a) Fen, and (b) Shallow water.

individual classes and class pairs. For this, Eqs. (1) and (2) can be used, respectively.

N � 1

Var <sup>¼</sup> <sup>1</sup>

they should be removed before separability analyses as noisy and poor features.

employed for differentiating wetland classes.

118 Wetlands Management - Assessing Risk and Sustainable Solutions


Table 3 summarizes the results of separability analyses performed by U-test on five wetland classes (bog, fen, marsh, swamp, and shallow water) using multi-source optical (RapidEye,

Table 3. The most important optical (provided in the lower left half of the table) and SAR (provided in the upper right half of the table) features for delineating each pair of wetland class in June and August, respectively (the features are ordered based on their separability measures).

Landsat-8, Sentinel-2A, and ASTER) and SAR (Sentinel-1, RADARSAT-2, and ALOS-2) data in NL, Canada. As is clear from this table, the ratio features provided the highest separability measures. NIR Brightness and Red Edge Brightness ratios are most efficient regarding the optical data, and the ratios of HH/HV and HH/TP obtained from RADARSAT-2 full-polarimetric data are the most important SAR features for separating wetlands.

accuracy for individual classes is significantly challenging in these cases. This becomes even more serious when only SAR data are applied for discriminating wetlands. There are several studies which developed new MCSs to improve the classification accuracy of similar

A Collection of Novel Algorithms for Wetland Classification with SAR and Optical Data

http://dx.doi.org/10.5772/intechopen.80688

121

Figure 3. Proposed multiple classifier system by Amani et al. [38] to improve the classification accuracy of the complex

environments.

Comparing the optical spectral bands, the NIR and Red Edge bands are most effective for discriminating wetland classes. Two main characteristics of wetlands are vegetation and water, which can be efficiently studied by these two bands. This demonstrates that it is more efficient to use the optical satellites, in which both NIR and Red Edge bands are included (e.g. Sentinel-2A and RapidEye). In this regard, Sentinel-2A, which provides free imagery, is superior for employment in operational wetland mapping and monitoring. The red band is also helpful in separating wetlands, especially discrimination between bog and other wetlands, because of bogs' red appearance. Additionally, there is a high overlap between the spectral signatures of wetlands in the green, SWIR, and TIR bands, and thus, there is a difficulty in using these bands for wetland studies. Finally, the blue band is not very useful in most of the cases.

Comparing various decomposition methods, including Freeman-Durden, Cloude-Pottier, Touzi, Van Zyl, Yamaguchi, and Krogager, it is observed that coherent decomposition techniques, such Krogager, are not recommended for wetland classification. The reason is that the coherent decompositions are mostly applicable for detecting man-made features in urban areas and less useful for naturally distributed targets such as wetland classes [93]. In addition, the Cloude-Pottier and Freeman-Durden methods are most optimum for separating wetland species. In this regard, the volume scattering component of Freeman-Durden and Anisotropy element of Cloude-Pottier are generally the best. Moreover, some SAR features extracted from the eigenvalue/eigenvector of the coherency matrix demonstrated a high potential for separating wetland class pairs and all wetland classes. In this regard, the serd, normalized serd and normalized derd, introduced by [94], are frequently selected for wetlands separation.
