2. Wetland classification using SAR data

SAR is an active imaging system, capable of recoding the electromagnetic spectrum at much longer wavelengths compared to optical sensors. Unlike optical sensors, which collect ground target information at the cellular and molecular level, SAR sensors are responsive to physical (e.g., water content and size) and structural (e.g., roughness) characteristics of ground targets [43]. Over the past two decades, synthetic aperture radar (SAR) sensors have provided valuable data for wetland vegetation mapping. In particular, they are of great use when the efficiency of optical sensors is hampered by cloud cover and day/night conditions. Furthermore, SAR signal penetration depth through vegetation and soil offers additional information unavailable from optical remote sensing data [44, 45]. This is of great importance for monitoring the flooding status of vegetation due to enhanced double bounce scattering effects. Notably, the primary characteristics of SAR signals, such as wavelength, polarization, and incidence angle, with regard to key specifications of the ground targets, such as dielectric constant, roughness, and structure, determine the amount of SAR backscattered energy detected by SAR sensors [43]. Despite these benefits, SAR images are affected by speckle noise that degrades the radiometric quality of image, imposing challenges for several subsequent SAR processing tasks [46, 47]. Fortunately, Mahdianpari et al. [41] demonstrated the effect of applying an efficient despeckling method on the accuracy of wetland classification.

2.3. Wetland mapping using PolSAR data

Although single-polarized SAR data have been less useful for wetland classification, they have demonstrated great promise for monitoring open water surfaces in different applications, such as water body extraction and flood mapping [57]. This is because of the side-looking data acquisition geometry of SAR sensors. In particular, a large portion of the microwave signals transmitted to calm open water are scattered away from the SAR sensor, and therefore, open water appears dark in a SAR image, making it distinguishable from surrounding land [45]. Unlike singlepolarized SAR data, polarimetric SAR (PolSAR) imagery was found to be extremely useful for wetland vegetation mapping. This is because a full polarimetric SAR sensor (e.g., RADARSAT-2) collects the full scattering matrix, providing comprehensive information about ground targets for each imaging pixel [58]. Furthermore, PolSAR data allow the employment of polarimetric decomposition techniques to identify the different backscattering mechanisms of the ground targets and, accordingly, regions of flooded vegetation [45, 49, 59]. Unlike coherent decompositions (e.g., Krogager decomposition), which are only useful for man-made structures with deterministic targets, incoherent decompositions determine the relative contributions from different scattering mechanisms. Thus, they may be more efficient for obtaining information from natural scatterers, such as wetland ecosystems [59–61]. Cloude-Pottier, Freeman-Durden, Yamaguchi, Van Zyl, and Touzi decompositions are among the well-known incoherent decomposition tech-

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

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

115

Despite the efficiency of the polarimetric decomposition technique to characterize different scattering mechanisms of ground targets that correspond to different wetland classes, the accuracy of wetland classification could be improved. This is attributed to both the highly dynamic nature of wetland ecosystems and the similarity of different wetland classes. The former of which can be alleviated by using multi-temporal SAR data to accurately characterize wetland dynamics during growing seasons [41, 51, 61, 62]. Furthermore, some studies employed a large number of input features to tackle the problem of similarity between different wetland classes [63]. Despite the promising results obtained from such an approach to date, it may not necessarily be optimal approach due to both computational complexity and redundant information within a large number of input data. Furthermore, some wetland classes can be easily distinguished using a minimal of input features. For example, the shallow water class can be easily separated using a SAR backscattering analysis and employing a threshold. However, this similarity is more pronounced among herbaceous wetlands, indicating the necessity of incorporating a larger number of input data [45]. As such, a hierarchical classification scheme can be useful to optimize the number of input features according to the similarity of wetland classes, which should be distinguished at each classification level. Some recent studies also noted that the discrimination of wetland classes can be further increased by applying a feature weighting approach using the Fisher Linear Discriminant Analysis technique [61, 64]. Such an efficient approach eliminates the

The information content within SAR data increases given the polarization hierarchy, starting from single polarization to dual polarization and reaching both compact and full polarimetric

niques useful for wetland mapping using PolSAR data [45, 49, 61, 62].

necessity for the inclusion of large number of input data.

2.4. Wetland mapping using compact polarimetry data

### 2.1. SAR wavelength

SAR wavelength is another influential factor for wetland vegetation mapping. To date, most SAR satellites have operated in three microwave bands, including X-, C-, and L-bands with wavelength of 3.1, 5.6, and 23.6 cm, respectively. Each wavelength has its own advantages and disadvantages. The selection of an appropriate SAR wavelength depends on the wetland classes since the interaction of SAR wavelengths varies widely with different vegetation types depending on their size. For example, longer wavelengths (L-band) can pass through the vegetation canopy and detect water beneath the flooded trees and/or dense vegetation. Accordingly, several studies reported the superior capability of L-band relative to the shorter wavelengths (e.g., C- and X-band) for monitoring woody wetlands (e.g., swamp), since the incident SAR signal interacts with larger trunk and branch components [48, 49]. In particular, L-band holds great promise in discriminating between forested wetland (e.g., swamp) and dry forest [45, 50]. However, shorter wavelengths are preferred for monitoring herbaceous vegetation because SAR wavelength and vegetation canopies (e.g., leaf) are relatively the same size [51].

Observations from SEASAT L-band data were among the first applications of SAR data for mapping the flooding status of vegetation [52, 53]. Later studies confirmed the suitability of Lband observations for mapping inundation in forested wetlands using JERS-1 and ALOS PALSAR-1 [50, 54]. Following the successful launch of C-band satellites, such as ERS1/2 and RADARSAT-1, several studies have also examined the capacity of C-band observation for wetland mapping. Most of those early studies reported the superior capability of L-band for mapping forested wetlands relative to C-band [44, 55].

#### 2.2. SAR polarization

Overall, the HH polarized signal has been the most efficient for monitoring the flooding status of vegetation, since it is more sensitive to double bounce scattering associated with tree trunks in swamp forest and stems in freshwater marshes [54, 56]. VV polarization can also be useful when plants have begun to grow in terms of height but have a less developed canopy [51]. This is because in the middle of the growing season the vertically oriented structure of vegetation enhances the attenuation of VV polarization signals and, as such, the radar signal cannot penetrate to the water surface below the vegetation [48]. Cross polarization observation (HV and VH) has also been characterized as being highly sensitive to differences in biomass [57].

### 2.3. Wetland mapping using PolSAR data

unavailable from optical remote sensing data [44, 45]. This is of great importance for monitoring the flooding status of vegetation due to enhanced double bounce scattering effects. Notably, the primary characteristics of SAR signals, such as wavelength, polarization, and incidence angle, with regard to key specifications of the ground targets, such as dielectric constant, roughness, and structure, determine the amount of SAR backscattered energy detected by SAR sensors [43]. Despite these benefits, SAR images are affected by speckle noise that degrades the radiometric quality of image, imposing challenges for several subsequent SAR processing tasks [46, 47]. Fortunately, Mahdianpari et al. [41] demonstrated the effect of

SAR wavelength is another influential factor for wetland vegetation mapping. To date, most SAR satellites have operated in three microwave bands, including X-, C-, and L-bands with wavelength of 3.1, 5.6, and 23.6 cm, respectively. Each wavelength has its own advantages and disadvantages. The selection of an appropriate SAR wavelength depends on the wetland classes since the interaction of SAR wavelengths varies widely with different vegetation types depending on their size. For example, longer wavelengths (L-band) can pass through the vegetation canopy and detect water beneath the flooded trees and/or dense vegetation. Accordingly, several studies reported the superior capability of L-band relative to the shorter wavelengths (e.g., C- and X-band) for monitoring woody wetlands (e.g., swamp), since the incident SAR signal interacts with larger trunk and branch components [48, 49]. In particular, L-band holds great promise in discriminating between forested wetland (e.g., swamp) and dry forest [45, 50]. However, shorter wavelengths are preferred for monitoring herbaceous vegetation because SAR

Observations from SEASAT L-band data were among the first applications of SAR data for mapping the flooding status of vegetation [52, 53]. Later studies confirmed the suitability of Lband observations for mapping inundation in forested wetlands using JERS-1 and ALOS PALSAR-1 [50, 54]. Following the successful launch of C-band satellites, such as ERS1/2 and RADARSAT-1, several studies have also examined the capacity of C-band observation for wetland mapping. Most of those early studies reported the superior capability of L-band for

Overall, the HH polarized signal has been the most efficient for monitoring the flooding status of vegetation, since it is more sensitive to double bounce scattering associated with tree trunks in swamp forest and stems in freshwater marshes [54, 56]. VV polarization can also be useful when plants have begun to grow in terms of height but have a less developed canopy [51]. This is because in the middle of the growing season the vertically oriented structure of vegetation enhances the attenuation of VV polarization signals and, as such, the radar signal cannot penetrate to the water surface below the vegetation [48]. Cross polarization observation (HV and VH) has also been characterized as being highly sensitive to differences in biomass [57].

applying an efficient despeckling method on the accuracy of wetland classification.

wavelength and vegetation canopies (e.g., leaf) are relatively the same size [51].

mapping forested wetlands relative to C-band [44, 55].

2.1. SAR wavelength

114 Wetlands Management - Assessing Risk and Sustainable Solutions

2.2. SAR polarization

Although single-polarized SAR data have been less useful for wetland classification, they have demonstrated great promise for monitoring open water surfaces in different applications, such as water body extraction and flood mapping [57]. This is because of the side-looking data acquisition geometry of SAR sensors. In particular, a large portion of the microwave signals transmitted to calm open water are scattered away from the SAR sensor, and therefore, open water appears dark in a SAR image, making it distinguishable from surrounding land [45]. Unlike singlepolarized SAR data, polarimetric SAR (PolSAR) imagery was found to be extremely useful for wetland vegetation mapping. This is because a full polarimetric SAR sensor (e.g., RADARSAT-2) collects the full scattering matrix, providing comprehensive information about ground targets for each imaging pixel [58]. Furthermore, PolSAR data allow the employment of polarimetric decomposition techniques to identify the different backscattering mechanisms of the ground targets and, accordingly, regions of flooded vegetation [45, 49, 59]. Unlike coherent decompositions (e.g., Krogager decomposition), which are only useful for man-made structures with deterministic targets, incoherent decompositions determine the relative contributions from different scattering mechanisms. Thus, they may be more efficient for obtaining information from natural scatterers, such as wetland ecosystems [59–61]. Cloude-Pottier, Freeman-Durden, Yamaguchi, Van Zyl, and Touzi decompositions are among the well-known incoherent decomposition techniques useful for wetland mapping using PolSAR data [45, 49, 61, 62].

Despite the efficiency of the polarimetric decomposition technique to characterize different scattering mechanisms of ground targets that correspond to different wetland classes, the accuracy of wetland classification could be improved. This is attributed to both the highly dynamic nature of wetland ecosystems and the similarity of different wetland classes. The former of which can be alleviated by using multi-temporal SAR data to accurately characterize wetland dynamics during growing seasons [41, 51, 61, 62]. Furthermore, some studies employed a large number of input features to tackle the problem of similarity between different wetland classes [63]. Despite the promising results obtained from such an approach to date, it may not necessarily be optimal approach due to both computational complexity and redundant information within a large number of input data. Furthermore, some wetland classes can be easily distinguished using a minimal of input features. For example, the shallow water class can be easily separated using a SAR backscattering analysis and employing a threshold. However, this similarity is more pronounced among herbaceous wetlands, indicating the necessity of incorporating a larger number of input data [45]. As such, a hierarchical classification scheme can be useful to optimize the number of input features according to the similarity of wetland classes, which should be distinguished at each classification level. Some recent studies also noted that the discrimination of wetland classes can be further increased by applying a feature weighting approach using the Fisher Linear Discriminant Analysis technique [61, 64]. Such an efficient approach eliminates the necessity for the inclusion of large number of input data.

#### 2.4. Wetland mapping using compact polarimetry data

The information content within SAR data increases given the polarization hierarchy, starting from single polarization to dual polarization and reaching both compact and full polarimetric data [65]. Specifically, fully polarimetric data are of great importance for land cover and, in particular, wetland mapping. Such a SAR sensor is constructed based on the standard linear basis (i.e., horizontal [H] and vertical [V]), wherein the sensor interleaves pulse with H and V polarization toward the ground targets and record both received polarizations simultaneously and coherently [65]. As such, the first disadvantage of full polarimetric SAR sensors is a time constraint because two orthogonal polarizations are transmitted alternately. Furthermore, such a configuration implies complexity due to doubled pulse repetition frequency, as well as an increase in the data rate by a factor of four relative to a single-polarized SAR system [65]. Accordingly, the image swath width of FP SAR images is halved, resulting in reduced coverage and an increase in satellite revisit time [66]. Finally, this configuration allows a limited range of incidence angles compared to that of single/dual polarization modes [67].

which quantifies the degree of similarity of the same pixel in the time interval between two SAR

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

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

117

Interferometric coherence is a quality indicator of InSAR observations. The variation of coherence in wetlands is a function of the complex mixture of several factors that contribute to coherence maintenance. The temporal baseline is one of the main parameters that hampers the application of InSAR for wetland monitoring [83]. Herbaceous vegetation, one of the most substantive components of wetland ecosystems, may easily lose coherence within a day or week. In the case of using shorter wavelengths (e.g., C- and X-band), interferometric coherence may be lost due to the shallow penetration depths of the shorter wavelengths. In contrast, longer wavelengths have deeper penetration depth but have been previously associated with longer temporal baselines (46 and 44 days for ALOS PALSAR-1 and JERS-1, respectively), which could cause a loss of coherence. However, this drawback has been addressed in the currently operating L-band SAR sensor (i.e., ALOS-2), wherein the temporal baseline is 14 days. Thus, ALOS-2 repeat-pass SAR images offer a promising source of data for wetland InSAR applications. Geometric decorrelation caused by different satellite look angles, volumetric decorrelation caused by vegetation volume scattering [83, 84], the Doppler centroid effect, and co-registration error during interferometric processing [65, 85] are other sources of

Despite these limitations, several studies reported the feasibility of InSAR for wetland water level monitoring. In particular, when the vegetation within or adjacent to standing water is able to backscatter the radar pulse toward satellite sensor, water level changes are observable from the phase data [86, 87]. Also, vegetation should not be too dense for the penetration of microwave energy [65]. The efficiency of the InSAR technique for wetland monitoring has been initially investigated in the Amazon floodplain [77]. Subsequent investigations have been carried out for a number of other wetland sites such as Florida Everglades [49, 77, 87, 88], the

In addition to hydrological monitoring of wetlands using InSAR, the interferometric coherence can be used for other wetland applications, such as change detection and classification [51, 91]. This is because coherence has a diagnostic function and can be used along with SAR backscatter and polarimetric decomposition techniques for classification of different wetlands. Each feature has specific characteristics and, accordingly, plays a different role for discriminating wetland classes. For example, SAR intensity depends on the electromagnetic structure of the targets, while the interferometric coherence reflects their mechanical and dielectric stability. Thus, an integration of different input feature augments land cover information and improves

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

Wetlands are complex landscapes and ecologically share similar characteristics. However, each wetland type contains its own specifications, which can be effectively investigated using

Louisiana Coastal wetland [56, 89], and China wetlands [89, 90].

classification accuracy of wetland types [51].

optical and SAR data

acquisitions, cannot be maintained [51].

decorrelation over wetlands.

An attractive alternative, which addresses the limitations of full polarimetric SAR sensors, is a compact polarimetry (CP) SAR configuration. The CP SAR image is expected to maintain polarimetric information as close as possible to that of full polarimetric SAR mode imagery while alleviating its primary limitations [68]. In particular, CP sensors collect a greater amount of scattering information compared to single- and dual-polarization modes while covering twice the swath width of full polarization SAR systems [69]. Thus, CP SAR configurations decrease the complexity, cost, mass, and data rate of a SAR system while preserving several advantages of a full polarimetric SAR system [70]. m-delta [71], m-chi [72], and m-alpha [73] are common decomposition techniques of compact polarimetry data. Importantly, the upcoming RADARSAT Constellation Mission (RCM), which will operate in the Circular Transmitting Linear Receiving (CTLR) mode, offers improved operational capabilities (e.g., ecosystem monitoring) along with a much shorter satellite revisit period. Specifically, RCM provides daily coverage over Canada with 350-km imaging swaths [74]. This is of great significance for highly dynamic phenomenon such as wetland complexes. Some recent studies reported the efficiency of simulated compact polarimetric data for wetland mapping [68, 75].

#### 2.5. Wetland monitoring using InSAR

Hydrological monitoring of wetlands is another subject of interest, since they are waterdependent ecosystems. SAR images have shown to be useful for wetland hydrological monitoring using both SAR backscattering responses [76] and a more detailed and sophisticated technique, Interferometric SAR (InSAR) [77]. This is because the flooded and non-flooded statuses of vegetation in wetland environments have distinct differences in radar backscattering responses that play an important role in the hydrological monitoring of wetlands. Specifically, a time series analysis of SAR backscatter signatures has offered information of seasonal patterns of flooding in wetland ecosystems, and the enhanced SAR backscatter signature of flooded vegetation has been examined in a number of studies [76, 78–81].

Although several studies reported the potential of InSAR for wetland water level monitoring, its application in wetlands presents challenges. This is primarily due to the substantial altering of reflectance and energy backscatter of wetland environments, even within hours or days [82], and the low backscatter of the water surface. Under these conditions, interferometric coherence,

which quantifies the degree of similarity of the same pixel in the time interval between two SAR acquisitions, cannot be maintained [51].

data [65]. Specifically, fully polarimetric data are of great importance for land cover and, in particular, wetland mapping. Such a SAR sensor is constructed based on the standard linear basis (i.e., horizontal [H] and vertical [V]), wherein the sensor interleaves pulse with H and V polarization toward the ground targets and record both received polarizations simultaneously and coherently [65]. As such, the first disadvantage of full polarimetric SAR sensors is a time constraint because two orthogonal polarizations are transmitted alternately. Furthermore, such a configuration implies complexity due to doubled pulse repetition frequency, as well as an increase in the data rate by a factor of four relative to a single-polarized SAR system [65]. Accordingly, the image swath width of FP SAR images is halved, resulting in reduced coverage and an increase in satellite revisit time [66]. Finally, this configuration allows a limited range of incidence angles compared to that of single/dual polarization

An attractive alternative, which addresses the limitations of full polarimetric SAR sensors, is a compact polarimetry (CP) SAR configuration. The CP SAR image is expected to maintain polarimetric information as close as possible to that of full polarimetric SAR mode imagery while alleviating its primary limitations [68]. In particular, CP sensors collect a greater amount of scattering information compared to single- and dual-polarization modes while covering twice the swath width of full polarization SAR systems [69]. Thus, CP SAR configurations decrease the complexity, cost, mass, and data rate of a SAR system while preserving several advantages of a full polarimetric SAR system [70]. m-delta [71], m-chi [72], and m-alpha [73] are common decomposition techniques of compact polarimetry data. Importantly, the upcoming RADARSAT Constellation Mission (RCM), which will operate in the Circular Transmitting Linear Receiving (CTLR) mode, offers improved operational capabilities (e.g., ecosystem monitoring) along with a much shorter satellite revisit period. Specifically, RCM provides daily coverage over Canada with 350-km imaging swaths [74]. This is of great significance for highly dynamic phenomenon such as wetland complexes. Some recent studies reported the efficiency

Hydrological monitoring of wetlands is another subject of interest, since they are waterdependent ecosystems. SAR images have shown to be useful for wetland hydrological monitoring using both SAR backscattering responses [76] and a more detailed and sophisticated technique, Interferometric SAR (InSAR) [77]. This is because the flooded and non-flooded statuses of vegetation in wetland environments have distinct differences in radar backscattering responses that play an important role in the hydrological monitoring of wetlands. Specifically, a time series analysis of SAR backscatter signatures has offered information of seasonal patterns of flooding in wetland ecosystems, and the enhanced SAR backscatter signature of

Although several studies reported the potential of InSAR for wetland water level monitoring, its application in wetlands presents challenges. This is primarily due to the substantial altering of reflectance and energy backscatter of wetland environments, even within hours or days [82], and the low backscatter of the water surface. Under these conditions, interferometric coherence,

of simulated compact polarimetric data for wetland mapping [68, 75].

flooded vegetation has been examined in a number of studies [76, 78–81].

2.5. Wetland monitoring using InSAR

116 Wetlands Management - Assessing Risk and Sustainable Solutions

modes [67].

Interferometric coherence is a quality indicator of InSAR observations. The variation of coherence in wetlands is a function of the complex mixture of several factors that contribute to coherence maintenance. The temporal baseline is one of the main parameters that hampers the application of InSAR for wetland monitoring [83]. Herbaceous vegetation, one of the most substantive components of wetland ecosystems, may easily lose coherence within a day or week. In the case of using shorter wavelengths (e.g., C- and X-band), interferometric coherence may be lost due to the shallow penetration depths of the shorter wavelengths. In contrast, longer wavelengths have deeper penetration depth but have been previously associated with longer temporal baselines (46 and 44 days for ALOS PALSAR-1 and JERS-1, respectively), which could cause a loss of coherence. However, this drawback has been addressed in the currently operating L-band SAR sensor (i.e., ALOS-2), wherein the temporal baseline is 14 days. Thus, ALOS-2 repeat-pass SAR images offer a promising source of data for wetland InSAR applications. Geometric decorrelation caused by different satellite look angles, volumetric decorrelation caused by vegetation volume scattering [83, 84], the Doppler centroid effect, and co-registration error during interferometric processing [65, 85] are other sources of decorrelation over wetlands.

Despite these limitations, several studies reported the feasibility of InSAR for wetland water level monitoring. In particular, when the vegetation within or adjacent to standing water is able to backscatter the radar pulse toward satellite sensor, water level changes are observable from the phase data [86, 87]. Also, vegetation should not be too dense for the penetration of microwave energy [65]. The efficiency of the InSAR technique for wetland monitoring has been initially investigated in the Amazon floodplain [77]. Subsequent investigations have been carried out for a number of other wetland sites such as Florida Everglades [49, 77, 87, 88], the Louisiana Coastal wetland [56, 89], and China wetlands [89, 90].

In addition to hydrological monitoring of wetlands using InSAR, the interferometric coherence can be used for other wetland applications, such as change detection and classification [51, 91]. This is because coherence has a diagnostic function and can be used along with SAR backscatter and polarimetric decomposition techniques for classification of different wetlands. Each feature has specific characteristics and, accordingly, plays a different role for discriminating wetland classes. For example, SAR intensity depends on the electromagnetic structure of the targets, while the interferometric coherence reflects their mechanical and dielectric stability. Thus, an integration of different input feature augments land cover information and improves classification accuracy of wetland types [51].
