Radar detection of wetland ecosystems: a review

Abstract

Periodically, reviews of our knowledge of radar–wetland relationships and detection parameters have been provided by various authors. Since the publication of these works, additional research has been completed. Five major remote sensing journals spanning the years 1965–2007 formed the basis of this review. The vast majority of significant material found its way into these mainstream journals in one aspect or another. A short history of Synthetic Aperture Radar (SAR)–wetlands discovery based on earlier reviews is followed by an update on radar‐related wetland research. Although some trends emerged with regard to which wavelengths or polarizations to use, there was variation in optimum season/time of year and selection of multitemporal imagery. What is evident throughout the recent literature is that multidimensional radar data sets are attaining an accepted role in operational situations needing information on wetland presence, extent and conditions.

1. Introduction

Wetlands are a vital ecological component and core element in resource management programmes ranging from local to global ecosystems. Wetlands perform a variety of functions beneficial to society. Wetlands help to control flooding, ameliorate droughts, provide habitat for a myriad of flora and fauna, maintain and improve water quality, provide storage for water, stabilize water supply, mitigate erosion, reduce hurricane‐related damage and offer recreation possibilities (Novitzki et al. 1996). Globally, wetlands, as methane generators and carbon sinks, are recognized as important contributors to weather modification.

According to the Millennium Ecosystem Assessment (2005), more specific contributions made by wetlands include: being the principle supply of renewable fresh drinking water for an estimated 1.5–3 billion people; reducing nitrate concentrations by up to 80%; contributing some US$34 billion to the gross world product annually through wetland‐related fisheries; and significantly exceeding the economic value of converted or altered wetlands (e.g. intact mangroves in Thailand were estimated to have a total economic value of US$1000–$36 000 per hectare compared to US$200 per hectare for wetlands converted to shrimp farms).

At the same time, wetland loss and conversion has been extensive, as is evident by the following examples. Between the 1950s and 2005, Mesopotamian marshes in southern Iraq decreased from over 20 000 square kilometres to less than 400 square kilometres. Over 54% of the world's total mangrove area has been lost in the past few decades. Diversion of fresh water from estuaries has increased riverine sediment flow by about one‐fifth, reducing crucial contributions to wetland fisheries and incubators. Over 60% of European and North American marshes have been lost or degraded by clearing and draining for agricultural use (Millennium Ecosystem Assessment 2005).

Detection and monitoring of these valuable ecosystems has assumed increasing importance in light of the above‐mentioned benefits and the constant pressures they sustain from land use change and development. The magnitude of this task coupled with the dynamic nature and inaccessibility of wetland ecosystems has limited the use of on‐ground efforts and encouraged the use of various remote sensing platforms, given their ability to record large areas in comparatively short time periods. Although sensors in the optical range of the electromagnetic spectrum have received the greatest attention, considerable effort has also been focused on the use of imaging radars.

The reasons are many. As radar operates in the microwave portion of the spectrum, it offers complementary and supplementary data to sensors operating in the optical and thermal bands. At the same time it patently provides unique data. Radar backscatter is sensitive to dielectric properties (soil and vegetation moisture content) and geometric (surface roughness) attributes of the imaged surface. In many areas of the world (e.g. cloud‐covered and/or low‐light regions of the Earth) imaging radar is the only sensor that can provide consistent, periodic data in a reliable manner. Optical and thermal systems are also limited by their inability to penetrate vegetation canopies whereas radar systems can, to some degree, provide subcanopy information.

Reports on the capability of radar systems to provide information pertaining to wetland attributes have appeared in the literature since the late 1960s and early 1970s (Waite and MacDonald 1971). Periodically, reviews and overviews of our knowledge of the radar–wetland relationships and detection parameters (including wavelength, polarization, incident angle, temporal, and environment – dielectric and surface roughness) have been provided by various authors. Hess et al. (1990), Kasischke et al. (1997), Schmullius and Evans (1997) and Ramsey (1998) have provided useful status reports to this end.

Since the completion of these works, additional data from Radarsat, ERS and JERS‐1, among others, have been studied and the results reported in the literature. This paper is an update of where we stand as a result of these efforts. Specifically, this paper provides a review of our current knowledge and role of imaging radar with regard to wetland analyses. The next section briefly describes the methodology of the review process. Then, a short history of Synthetic Aperture Radar (SAR)–wetlands discovery, by way of a synopsis of the four well‐known review articles cited above, traces the development and conclusions reached by researchers in the geosciences up to that time. Following that, an update to the present is provided as a retrospect and prospect.

2. Methodology

Five major remote sensing journals spanning the years 1965–2007 (or their date of inception) formed the basis of this review. Although SAR–wetlands results have also been published in non‐remote sensing journals, presentations given at professional meetings and conferences, and reports printed in proceedings and symposia, inclusion of these venues was too broad in scope for this study. In addition, it was felt that the vast majority of significant material that may have been included in such outlets also found its way into the mainstream journals in one aspect or another. The five journals used in this review are the following:

Photogrammetric Engineering and Remote Sensing

Remote Sensing of Environment

Canadian Journal of Remote Sensing

International Journal of Remote Sensing

IEEE Transactions on Geoscience and Remote Sensing.

Each issue of each journal was searched for articles containing reference to wetland detection with SAR. This search reviewed all articles that concentrated on wetlands. It also included articles whose main focus encompassed related landscape features such as land cover, forestry or agriculture but also reported meaningful wetland observations in the abstract and/or conclusions. Only research results based on the analysis of imaging radar are included in this review. Such a process is admittedly subjective but permitted screening and comprised a pragmatic standard.

3. History and development

This section is a summary of the four published reviews provided by Hess et al. (1990), Kasischke et al. (1997), Schmullius and Evans (1997), and Ramsey (1998), presented in that order. The intent is to provide the reader with a brief perspective and synthesis of trends and progress leading up to this retrospective. Typical of most historical overviews, each synopsis contains some overlap with the earlier work. To provide continuity we have attempted to summarize the linkages and discussion inherent therein.

3.1 Hess review

Hess et al. (1990) examined 34 articles on forested wetland detection dating from 1971 to 1990. This suite of articles encapsulated most of what had been discovered over that 19‐year period. They reported that most studies up to that time had concentrated on swamp forests in the southeastern US and used L‐band SARs. Accuracies of up to 93% for flooded forest were reported when SAR was combined with optical imagery of the US Atlantic Coast. Sources of error and confusion with wetlands included agricultural crops and urban areas, but the major problem was confusion between flooded forests and flooded non‐forest (marsh/herbaceous) vegetation.

Incident angle influence varied inconsistently with: (1) forest type, (2) stand structure and (3) canopy composition. Like polarization (HH, VV) was deemed preferable to separate flooded from non‐flooded forest and produced higher contrast between swamps and dry forests than cross‐polarization at Ka‐, X‐ and L‐bands. However, some studies concluded that cross‐polarization may be preferable when attempting to separate swamp from marsh with L‐band imagery.

Radar polarimetry as a field of investigation was just being introduced at this time. The ability to examine the complete scattering matrix of received amplitude and phase of vertical and horizontal polarizations was intriguing. Large phase differences between H and V polarizations were found between swamp and dense stands of upland Maine forest. Large coefficient of variation values for forests and small values for swamp implied that swamps were subject to large changes in backscatter due to polarization shifts (Durden et al. 1989). The reason for such change was thought to be due to a change in the dominant scattering mechanisms: double‐bounce in the swamps, in contrast to randomly polarized backscatter from forest branches. Hess et al. (1990) concluded that such empirical data supported the theory that the increase in backscatter in flooded forests was due to signal double‐bounce between trunks and branches and the water surface beneath.

Not surprisingly, given the predominance and availability of Seasat, SIR‐A and B imagery, the majority of analyses at this time used L‐band imagery. The longer wavelength L‐band imagery also penetrated forest canopy cover and resulted in enhanced backscatter in comparison to shorter wavelength systems. Nevertheless, earlier work with Ka‐band imagery had noted high backscatter from mangrove and cyprus‐tupelo swamps. High backscatter in X‐band imagery was also reported for marsh vegetation (sedges, emergent reeds, and grasses) but flooded forests produced low returns, the opposite of L‐band studies. However, Hess et al. (1990) also noted one study where bare branches of a dead forest produced high backscatter on both X‐ and L‐bands and an instance where Seasat imagery recorded high backscatter from lowland herbaceous vegetation in eastern Maryland and Virginia. P‐band work was insufficient in quantity for any determination of its backscatter response in such environments; however, one study did suggest that the longer P‐band wavelength would produce greater signal penetration and double‐bounce signals in tall and/or very dense forest stands.

Hess et al. (1990) thought that flooded forests provided a simplified setting to understand and model SAR–vegetation interactions. Continuous water cover provided a consistent and known surface roughness and dielectric constant. Many flooded forests were also species poor; the lack of within‐stand diversity simplified model development. Possible effects of topographic relief and local incident angle changes were minimal due to the flat terrain. At the same time the authors cautioned that there were also disadvantages. A lack of accurate data on ground conditions including the extent of flooding at the time of image acquisition as well as the spatial variability of flooded surface areas could prove problematic. The relationship between hydrology and tree species must also be kept in mind to ensure that observed backscatter differences were the result of flooded/non‐flooded conditions and not due to differences between upland and lowland vegetation.

In conclusion, Hess et al. (1990) recommended a steep incident angle to facilitate signal penetration of the flooded forest canopy and the resultant double‐bounce effect. However, they also advised using multiple incident angle imagery to facilitate discrimination of forest structure differences. Flooded forest and non‐flooded forest stands and forest swamp and herbaceous marsh vegetation were generally easily separated (3 dB to 10 dB) with L‐band imagery. Like‐polarized data were preferable to separate swamps from dry forest but cross‐polarized data might be preferred for separating swamp from marsh. Much improvement in detection was anticipated with the use of P‐band data, polarimetric data and multiple wavelengths. At the same time the authors stated that ‘a better understanding of the separate and integrated effects of canopy components is necessary’ (Hess et al. 1990, p. 1322).

3.2 Kasischke review

Ecological applications with imaging radars were reviewed by Kasischke et al. (1997). This report of a National Research Council review by 16 panelists used 110 citations to reach conclusions in four general study areas: (1) detecting and classifying land cover change; (2) estimating woody biomass; (3) monitoring the extent and timing of flooding; and (4) monitoring other temporal and dynamic processes. Their recommendations and conclusions contained much new information regarding wetlands research with SAR.

New radar systems, almost global imagery acquisition, and increased quantitative analysis techniques mark this review period. Among their general findings reported were: (1) vegetation information is enhanced by using multitemporal, multifrequency and/or multipolarization data; (2) land cover classifications of >90% accuracy have been attained with SAR imagery; and (3) because of image speckle, individual landscape units must be several times larger than the spatial resolution of the radar system to be studied effectively; for spaceborne SAR the authors calculated the minimum unit to be at least 200×200 m in size.

The authors also reiterated and reinforced points made by earlier reviews including: (1) the presence of surface water in forested wetlands will result in additional double‐bounce return and an enhanced signal response, but this condition is usually only detected in L‐ or P‐band systems; and (2) there is no signal enhancement at X‐ and C‐band wavelengths except in leaf‐off conditions; otherwise the forest canopy attenuated the signal, resulting in surface scatter. Much of the research during this period focused on distinguishing flooded versus non‐flooded conditions in forests and marshes rather than wetland ecosystems per se. New observations from analysis of small incident angle ERS C‐VV imagery indicated that herbaceous wetland vegetation with surface water will produce additional forward scatter and a lower radar signal return than under wet‐soil or dry conditions.

Another significant addition to our knowledge of the interactions between SAR signals and wetlands centred on research analysing methane sources and sinks. As wetlands are considered to be the world's largest natural source of atmospheric methane, the ability to monitor their spatial and seasonal changes would be a great aid in modelling global methane exchange. Research in the Arctic herbaceous tundra found that ERS C‐VV backscatter data recorded different hydrologic conditions. Whereby dry sites generated low signal backscatter, flooded sites produced a high return and an intermediate return occurred when water was perched on the surface. In addition, it was possible to separate wetland from non‐wetland areas. Backscatter also increased with Leaf Area Index (LAI). C‐band VV Airborne SAR (AIRSAR) imagery was acquired at a second site in the taiga, where upland forest and tall shrubs were separable from bogs and open water. Overall classification accuracy was 89% but varied by category: from 67% for riparian tall shrubs to 100% for coniferous forest, water, flooded, and bare soil (Morrissey et al. 1994).

A companion article by Durden et al. (1996) expanded on this work. Neural networks, multifrequency and polarimetric/non‐polarimetric data comparisons were used to classify four taiga categories (fen, bog, forest, and water) with AIRSAR L‐ and C‐band data. Incident angles ranged from 30° to 50° but the authors found little dependence between SAR system parameters and incident angle change. Fens comprised herbaceous grasses and sedges while bogs consisted of black spruce and shrub bogs with tussocks. Double‐bounce was noted for fens but the radar backscatter values at C‐band were similar for fens, forests, and bogs. At L‐band wavelengths the fen radar backscatter was lower than forest and bogs but higher than water. Polarimetric and non‐polarimetric L‐band data produced similar overall accuracies (78–81%) but both confused bogs with forest. At C‐band, forests were still confused with bogs; in addition, fens were also confused with bogs. Although Morrissey et al. (1994) were able to identify fens with C‐HH data, it was not possible in this instance; L‐band data were required.

Complementing this effort in the Arctic, work in the tropics with C‐ and L‐band data from SIR‐C also reported the ability to separate flooded from non‐flooded forests in the Brazilian Amazon with L‐band imagery (Hess et al. 1995). Tidal cycle‐related flooded/non‐flooded detection with ERS C‐VV imagery was also noted in Florida black needlerush (Juncus roemerianus) marsh areas (Ramsey 1995).

Differences between flooded and non‐flooded sedge were observed on both C‐HH and L‐HH airborne imagery of Kenya. Both image sets could also discriminate between flooded and non‐flooded grasses, but emergent grasses extant in flooded areas could only be detected with L‐band data. Flooded emergent grasses appeared similar to non‐flooded grasses on the C‐HH imagery (Pope et al. 1992).

Herbaceous wetlands of the Yucatan Peninsula were the focus of research using single‐polarized and polarimetric C‐band and L‐band SIR‐C data (Pope et al. 1997). The three marshes were: (1) rush (Eleocharis cellulosa); (2) sawgrass (Cladium jamalcense); and cattail (Typha domingensis). Both C‐ and L‐band polarimetric data could separate dry or partially flooded marsh from flooded marsh, but neither could separate dry marsh from partially flooded marsh. Flooding was detected foremost when occurring in tall emergent vegetation (sawgrass and cattails). C‐PD (phase difference) data were found to be the superior overall polarimetric mode and the most sensitive to marsh flooding and vegetation types. Other results reflected the complexity of SAR–wetlands relationships. C‐HH data provided the highest accuracies for delimiting sawgrass and cattail marshes; however, C‐VV excelled only in detecting cattail marshes and low‐density marshes. L‐HH response showed backscatter increases from high‐density marshes and a decreased response from low‐density marshes. As interesting as these results are, it would have been useful if accuracy and category confusion assessments had been included in these reports.

Single‐ and multiscene ERS C‐VV imagery of Florida was also used to determine the possibility of monitoring seasonal water levels in coastal wetlands (Ramsey 1995). The results supported backscatter models pertaining to biomass and soil moisture conditions. The discrimination of wetland types was better under dry conditions than during wet periods. Although such general statements are useful to some extent, again the incorporation of accuracy assessments and a discussion of confusion aspects would have proved beneficial.

Based on the radar research reviewed, Kasischke et al. (1997) concluded that (1) like‐polarized radars were well‐suited for the detection of flooded vegetation; (2) L‐HH was preferred for wooded vegetation; and (3) C‐HH was preferred for herbaceous wetlands. These conclusions agreed with the earlier review by Hess et al. (1990). Comparing the potential of ERS and Radarsat imagery for wetland analysis, the authors rated ERS as poor unless there were no leaves. Radarsat was considered slightly ‘better’ than ERS but also limited to leaf‐off conditions and instances of low‐density stands. ERS was rated ‘very good’ in monitoring flood conditions in coastal/‘low stature’ wetlands; whereas Radarsat was rated slightly ‘better’ due to a higher repeat frequency of image capture and the acquisition of HH polarization (Kasischke et al. 1997). No mention was made of incident angle differences.

3.3 Schmullius and Evans review

Schmullius and Evans (1997) provided a tabular status of ‘optimal SAR parameters’ and recommendations for several environmental aspects based on a review of research efforts using SIR‐C/X‐SAR data. Eighteen articles in the ‘ecology’ section formed the basis of their wetland evaluation. To monitor flooded forests, they listed C‐band data as non‐mandatory while L‐HH was considered important and L‐VV and L‐cross‐polarization were listed as helpful. By contrast, X‐HH, X‐VV, X‐cross‐polarized, C‐HH, C‐VV, C‐cross‐polarized and L‐HH imagery were all listed as important contributors in herbaceous wetland investigations. L‐VV and L‐cross‐polarized were listed as helpful. This article was not as detailed as the other reviews but did provide an update and broad picture of radar research with SIR‐C/X SAR data and related wavelengths and polarizations. Incident angle data were not included in their evaluation.

3.4 Ramsey review

The book chapter authored by Ramsey (1998) provided the latest, albeit 10 years old, review of SAR sensing of wetlands. His digest of 99 scientific articles from literature dating from 1971 to 1997 included sections on: (1) flood detection under forest canopies; (2) identification of mangrove forests; (3) flood detection under grass canopies; (4) the complexities in flood detection; and (5) determining soil moisture and delimiting open water. A summary of the relationship between radar system characteristics and wetlands completed the review.

The preference for L‐HH imagery for investigations of forested wetlands was reiterated. Cases where incident angle was found insensitive to the detection of flooded forests with L‐HH imagery were contrasted with work where differences were observed among tree species, pointing to the need to explore these relationships through additional research. Work was also cited that reported the ability of C‐, X‐ and even K‐band radars to detect forested wetlands at lower incident angles under leaf‐off conditions. Factors such as canopy density, differences in the morphology of tree species, and moisture content were thought to be responsible for the variation in signal response.

Mangroves were given special attention because they are considered the world's most productive ecosystem based on net productivity (Ramsey 1998). Ramsey observed that flooded mangroves could be separated from non‐flooded mangroves with L‐HH data because the primary response from the former is due to double‐bounce whereas volume scatter was the primary source for the radar signal response from non‐flooded mangrove. Small incident angles were preferred. Confusion in detection linked to the mangrove's unique root systems was mentioned, but Ramsey stated that this should be minimal in flooded conditions. Confusion related to water pooling and non‐contiguous flooding was viewed as more problematic.

Ramsey's consensus of works focusing on flooded marsh grass indicated that L‐band signal response was minimal due to specular reflection; X‐band signals never reached the water surface. One C‐band analysis suggested that flooded marshes could be detected in some cases. In reviewing work in low marshes, Ramsey (1998) reported that detection was possible at X‐ and C‐bands but that this ability lessened with increases in canopy cover and LAI. In taller marsh the radar return in X‐ and C‐bands was from volume scatter whereas double‐bounce return was dominant in shorter marsh conditions. Like polarized radar backscatter was enhanced by flood conditions but not cross‐polarized returns.

The author drew an interesting parallel between the interaction of radar parameters (especially polarization) with vertically oriented flooded trees at L‐band and vertically oriented stalks of flooded herbaceous vegetation at shorter wavelengths. An analysis of ERS C‐VV data in a black needlerush marsh concluded that flooding reduced signal response. However, another study of similar marsh vegetation noted that the response increased with increases in new green biomass. In general, flood detection was deemed an asset to forested wetland detection if the flooding was permanent but a handicap if the flooding was intermittent and/or non‐contiguous unless the periodicity of the flooding was under investigation.

His summary of radar parameters provided continuity between this and past reviews. Like polarizations were preferred at L‐band to separate flooded from non‐flooded forests but cross‐polarized data were preferred to separate wetland forests from marshes. Signal attenuation was found to increase with increases in incident angle and was more severe for vertical than horizontal polarizations at L‐band. Signal attenuation was not dependent on polarization at C‐band. The increase in incident angle produced more signal/canopy interaction and less penetration. However, monitoring regrowth of burned marsh areas was possible with high incident angle X‐band data but not at low incident angles. The relationships among biomass and canopy and wavelength were found to be complex and inconsistent based on research at that time.

In conclusion, Ramsey (1998) stated that radar imagery could provide information on three wetland components: hydrology, soil moisture, and vegetation type. L‐band HH polarized low to moderate incident angle imagery was recommended for forested canopies, but steep incident angle C‐band HH and VV data could provide useful data in cases of sparse canopy closures. Detection of marsh and herbaceous wetlands was deemed limited and inconclusive. In cases where vegetation type could not be ascertained, soil moisture measurement was crucial to wetland detection. Again, longer wavelength imagery was recommended for forested areas whereas shorter wavelength imagery at low incident angles was preferred for herbaceous areas.

4. Current research results

More than 60 articles reporting SAR analysis of wetland ecosystems have appeared in the five major remote sensing journals since the last published review in 1998. A look at their results finds support for some past beliefs, coupled with modifications and differences for others, plus new information, especially with regard to our knowledge of C‐band data.

Considerable research in this period focused on the analysis of South American ecosystems with Radarsat and/or JERS imagery. A selection of Standard Mode Radarsat images (S1≈23.3°, S5≈39°, S6≈43.5°) over the Amazon were used to examine the relationships between incident angle and five macrophyte biophysical variables (wet weight; dry weight; percentage moisture; stand height; and percentage ground cover) (Novo et al. 1998). The best overall results were obtained with the S5 image, which was found to be sensitive to wet and dry biomass and percentage moisture. Highest backscatter occurred on the S1 image, but backscatter decreased with increases in biomass and moisture content on all images. However, a complex relationship between incident angle/backscatter, individual biophysical variables and the three species Scirpus sp., Eleocharis sp. and Typha sp. pointed to the benefits of multi‐incident angle imagery. A false colour composite of the three incident angles was generated to observe the unique information contained in each image.

A later study incorporated JERS‐1 and Radarsat imagery of the Tucuruí Reservoir in Brazil to continue this investigation of the relationship between radar backscatter and macrophyte stand variables (Novo et al. 2002). The L‐HH JERS‐1 data were collected at a 35° incident angle and the C‐HH Radarsat data were collected between 41° and 46° incident angles. Although C‐band was more sensitive to leaf shape, L‐band was preferred for measuring above‐ground biomass and stand height. C‐band did not respond to height changes. Based on scatter diagrams of C‐ and L‐band backscatter coefficients, the authors suggested that the combination of these radar bands could be used to accurately map three macrophyte genera: Eicchornia, Typha and Scirpus. This conclusion is partially supported by their accuracy assessment for Scirpus (83.6%) and Typha (90.5%) but is somewhat problematic given their stated accuracy of only 23.6% for Eicchornia. It should be noted that the low percentage reported for Eicchornia may be incorrect because the overall accuracy is 84.6%. Unless Eicchornia is a small class compared to the other two, it appears to be a mistake in the table. Other inconsistencies in the presentation and rounding of decimals support the potential for problems in this table. The authors stated that because their research was based on a small sample, their findings should not be generalized and only extended into other environments with caution.

Costa et al. (1998) also explored multifrequency and multi‐incident angle data using Radarsat and JERS‐1 imagery of the Amazon. Their research supported results reported earlier by others: L‐band data were recommended for detecting forested wetlands; C‐band was more sensitive to aquatic plants (average height 1 m); and steep incident angles were preferred. However, the L‐band JERS imagery exhibited specular return from aquatic plants unlike the double‐bounce reported from similar vegetation by Pope et al. (1997). The authors suggested that the difference may be that the JERS images were obtained at a less steep incident angle (i.e. 25–26° SIR‐C imagery used by Pope et al. (1997) vs. 39° for JERS imagery). JERS produced a higher accuracy in the five land cover classes examined (water, agriculture, forest, flooded forest, and aquatic plants) than Radarsat, but JERS confused forest with flooded forest. JERS/Radarsat composites increased accuracy by 10% to 66% depending on the cover class. Unfortunately, no S1/S6 Radarsat combination was evaluated to test multi‐incident angle data.

Continuing the regional emphasis on the Amazon and multifrequency data, Costa (2004) and Costa and Telmer (2006) incorporated additional radar data sets to the ones reported above. Costa (2004) expanded into the multitemporal aspect by using five dates of Radarsat and four dates of JERS‐1 acquired in 1996, 1997 and 1999, in conjunction with field data collection. The Radarsat and JERS‐1 data sets used the same polarization scheme and spatial resolution; however, the acquisition dates, wavelengths and incident angles (43° for Radarsat and 35° for JERS‐1) were different. A table of mean, lower and upper bounds for radar backscatter coefficients in dB for five ground cover types and nine radar acquisition dates included two categories directly relevant to this study, Floodplain Forests and Aquatic Plants.

An overall increase in backscatter coefficient sigma‐naught (σ⁰) with water level was observed in the Floodplain Forest. L‐band had a higher dynamic range between water level extremes as well as within a single image than C‐band, suggesting that the longer L‐band radar was more sensitive to variation of size and thickness of the vegetation elements. L‐band backscatter was higher than C‐band because of greater penetration and increased double‐bounce related to the longer wavelength and smaller incident angle with JERS‐1. Regions with seasonal flooding showed large variations in radar backscatter (σ⁰); when inundated, σ⁰ varied with the degree of defoliation of flooded trees. The reduced canopy allowed greater penetration, resulting in an increase in double‐bounce and radar backscatter (Costa 2004). The grass‐like Aquatics demonstrated the largest temporal σ⁰ variation. This was attributed to the interaction of microwave wavelengths with aquatics, resulting in specular scattering, volume scattering and double‐bounce scattering; moreover, the ratio among these scattering types changed as the water level varied (Costa 2004).

An accuracy assessment was conducted based on the comparison of classified maps with ground truth maps from the interpretation of aerial photography (Costa 2004). Accuracies for flooded Floodplain Forest were more than 93% for every case except the November date when it dropped to 58.9%. Because November is the period of lowest water levels in the Amazon, it was presumed that much of the area in question was not flooded at the date of the radar acquisition. By contrast, the highest accuracy assessment value for Aquatic Plants (99.1%) was recorded in November, when aquatic vegetation is beginning its growth cycle. However, the overall accuracy assessment for identification of Aquatic Plants never went below 93% for any of the five data sets. Using flood inundation records and the vegetation map produced from Radarsat and JERS‐1, the author then generated a four‐category vegetation zone map defining three zones based on annual duration of flooding: (1) at least 300 days, (2) approximately 150 days, and (3) approximately 60 days. The first class, flooding for at least 300 days, was further broken down into two pioneer vegetation‐based classes (grass‐like semi‐aquatics and tree‐like semi‐aquatic plants) (Costa 2004).

Costa and Telmer (2006) also used Radarsat and JERS‐1 radar imagery but of the Pantanal in southwest Brazil. They found a strong relationship between radar backscatter and aquatic vegetation assemblages from lakes with different salinity concentrations based on two dates of Radarsat‐1 collected during August 2001 and one JERS‐1 image collected in July 1993. The authors used these relationships to categorize lakes into three salinity classes: brackish, hard and fresh. After collecting geochemical data from 167 lakes, they calculated an accuracy of 91% when identifying two (fresh and brackish lakes) of the three salinity classes. The authors reasoned that because they could identify two of the three classes of lake, then the remaining class (hard) could be estimated. Based on these findings it would appear that the continued collection of radar data would permit the necessary monitoring of the Pantanal ecosystem using multitemporal radar‐derived aquatic vegetation delineation as an indicator of the ecosystem's health.

Although the main focus of a study by Rignot et al. (1997) was deforestation and recovery of tropical forests in the Amazon, flooded forests killed by inundation were also mapped. Using a combination of SIR‐C, JERS‐1, Landsat and SPOT, their findings supported high backscatter observations from flooded forests reported by other investigators. They also found a large phase difference between HH and VV polarizations at both C‐band and L‐band that they felt could be used to help identify and delineate flooded forests.

Mapping flood stage changes in the Amazon at centimetre‐scale accuracy was found to be possible using L‐HH SIR‐C interferometric data (Alsdorf et al. 2001). Flooded forest and inundated shrubs produced high backscatter on this 35° incident angle imagery due to signal double‐bounce. Although the size of the shrubs was not mentioned, earlier research had reported low, specular return from such herbaceous vegetation and moderate incident angles.

Previously, Miranda et al. (1996) had been able to map flooded vegetation in northwestern Brazil with JERS‐1 L‐HH band imagery. However, in more recent work, flooded vegetation could be identified only when it was oriented perpendicular to the JERS‐1 look direction due to double‐bounce return signal from the water surface to tree trunks lining the banks and back to the SAR antenna (Miranda et al. 1998). This effect was not visible when the river edge vegetation was oriented parallel to the SAR look direction. The authors reported a 61% accuracy for flooded vegetation using a semivariogram textural classifier. Findings by Rosenqvist et al. (2002) contributed to this regional puzzle involving JERS data. In this case a time series of JERS‐1 imagery was used successfully to model floodplain inundations and spatial variations at 100 m resolution in the Brazilian Amazon. The floodplain data also provided useful input to modelling regional estimates of CH₄ emissions.

Exploring analysis techniques continued to be a focus of research attention (Noernberg et al. 1999). Here, multiple‐polarization C‐band airborne SAR imagery (incident angles from 38° to 50°) of aquatic vegetation in the Amazon provided the setting to test the use of coefficient of variation (texture) and several biophysical indices (cover structure index; volume scatter index; biomass index) used earlier by Pope et al. (1994). Based on their analysis of several scattergrams of the data, the authors (in contrast to Pope et al. 1994) concluded that any of the indices could map open water, dead tree trunks, forest, vertically oriented genus Scirpus (a leafy foliage with blade‐like leaves 4–10 mm wide and height of 5–20 cm) and a group composed of all other vegetation. HH‐polarization produced the best results. Unfortunately, accuracy and confusion data were not discussed.

Multiple regression analysis of JERS‐1 and Thematic Mapper (TM) imagery of the Amazon was used to quantify the imagery relationships with various land cover types (Shimabukuro et al. 2007). The SAR data gave consistent high backscatter from the flooded vegetation but the response from the TM data varied inconsistently, producing little correlation. A better correlation between SAR/TM data and land cover types was achieved by excluding the flooded vegetation category from the analysis.

A multitemporal, multi‐incident angle investigation was conducted by Kandus et al. (2001) using winter and summer S1, S4 and S6 Radarsat images of Argentina. Here, the signal response from flooded forest plantations (willows and poplars) changed from attenuated to double‐bounce returns, but flooded rushes (Schaenoplectus californicus) produced the opposite, changing from double‐bounce to specular reflection in flooded conditions. Both forests and rushes experienced double‐bounce at steep incident angles under normal water conditions. Increases in incident angle were found uniformly to decrease radar backscatter from vegetation. However, the authors noted, but could not account for, instances of high backscatter from rushes on a shallow incident angle S6 image. They recommended a combination of incident angles, steep angles to detect wetlands and large incident angles to delineate the open water‐land interface, a position that agrees with earlier work by Leconte and Pultz (1991) and Brown et al. (1996).

In a companion piece to Kandus et al. (2001), Parmuchi et al. (2002) used five of the S1 and S6 Radarsat images of the same study area to compare a decision‐based classifier with the ISODATA unsupervised classifier. They concluded that multitemporal data were requisite to map wetlands and that the decision‐based classifier provided flexibility and slightly improved accuracy. However, while the overall accuracy was 85%, the highest error was for marsh wetlands (57% omission error), leading to their observation that it was not possible to map marshes with this data set.

Although focusing on water storage and water levels with JERS‐1 data of the Negro River, Argentina, Frappart et al. (2005) mapped eight classes of vegetation. The four wetland ecosystems were: (1) occasionally flooded forest, (2) occasionally flooded low vegetation, (3) always flooded forest and (4) submerged vegetation. The data were collected during two time periods (September to December 1995 and May to July 1996) and compiled into two continental‐scale 100‐m resolution mosaics. From the data set they calculated and used the mean backscatter coefficient of the two seasons and the ratio of the two seasons to map vegetation classes. The authors stated their findings could not be validated because of the large area mapped, but they concluded that the type of vegetation and the status of inundation could be accurately mapped using JERS data from two seasons (wet and dry) (Frappart et al. 2005).

Two projects carried out by Grings et al. (2005, 2006) examined temporal changes of marsh grasses in Argentina's Paraná River Delta. The two marsh types present different plant geometries: junco‐like are more cylindrical and cortadera‐like are more randomly oriented discs. The first study (Grings et al. 2005) used 13 dates of ERS‐2 from 5 June 1999 to 14 December 2000, one spring date of Radarsat (11 December 2000) and one spring date of Envisat (20 November 2003). The multitemporal, multipolarized data sets were all C‐band radars and their incident angles were all ±2° of 21°. The multitemporal, multipolarized data from ERS‐2 exhibited: (1) a large range in radar backscatter that was not exhibited by backscatter from a second land cover (forest) used as a reference; (2) large differences in HH and VV σ⁰ from junco marshes; (3) a change in HH/VV ratio response as the junco emerged; (4) increased sensitivity to plant density in VV data; and (5) more and better information on junco marshes with multipolarized data than with a single polarization. The authors concluded that the changes in the polarization ratio were related primarily to changes in HH backscatter from the early emergent stage to maturity. When junco first emerged, the HH response was less than VV response, but with growth, junco presented a sufficient cylindrical stalk. The HH signal response subsequently became higher than the VV backscatter due to the cylindrical structural parameters of junco and the polarization differences in double‐bounce return (Grings et al. 2005).

The second study by Grings et al. (2006) analysed multitemporal/multipolarized radar backscatter from the same two types of marshes (junco and cortadera) during tidal events using only Envisat Advanced Synthetic Aperture Radar (ASAR) (C‐band, HH and VV polarization) of the same Argentine study area. As found in the earlier study (Grings et al. 2005), the cylindrical structure of junco increased the occurrence of double‐bounce return during inundation. The authors also reported a significant difference in radar return between junco and cortadera when inundated by tidal waters and concluded that the higher return from junco during a flood event could be used to separate these two types of marshes.

Other studies involved the identification and mapping of mangrove forests in the Amazon Basin, French Guiana and Gabon with radar (Mougin et al. 1999, Simard et al. 2002, Sgrenzaroli et al. 2004, Walfir et al. 2005). One of the few studies using aircraft as a platform made use of polarimetric AIRSAR P‐, L‐ and C‐band data collected in June 1993 to investigate the structural characteristics of mangrove forests in French Guiana (Mougin et al. 1999). The vegetation parameters studied were tree height, tree diameter at breast height (DBH), tree density, basal area, and total above‐ground biomass (recognizing that many of the tree structural parameters were correlated). In general, radar backscatter and forest metrics were found to be strongly related, with P‐HV and L‐HV data exhibiting the best correlations to total biomass (0.94).

Mangrove forests were one of the tropical vegetation classes mapped at a regional scale in the Amazon by Sgrenzaroli et al. (2004). They provided radar backscatter values (σ⁰) for Lowland floodplain forest, Mangrove forest and Flooded forests plus two other ground cover types (Dense canopy forest and Open canopy). Using the Global Rain Forest Mapping (GRFM) data set (L‐band, 100 m resolution), they delineated the cover classes based on radar texture from the canopy structure and high σ⁰ indicative of double‐bounce scattering from flooded conditions. The regional scale map of swamp forest extent was verified using Landsat TM imagery but no accuracy assessment was provided.

Using a combination of two dates of Landsat 5 (December 1991 and August 1985) and one date of fine‐mode Radarsat‐1 collected in September 1998, Walfir et al. (2005) used conventional manual interpretation techniques (tone and texture) for identifying seven wetland cover classes in the Brazilian Amazon: (1) Young intertidal mangrove; (2) Intertidal mangrove; (3) Supratidal mangrove; (4) Outer salt marsh; (5) Inner salt marsh; (6) Degraded mangrove; and (7) Regenerated mangrove. They found that where mangrove had been removed (deforested), the image was smooth textured and dark toned, whereas regenerated mangrove areas exhibited a high backscatter (light toned) due to double‐bounce scattering. They also stated that ‘inner marsh’ was dark when flooded and backscatter increased when not inundated. They determined that radar provided information on vegetation height, water content, plant geometry, and degraded and regenerating mangrove forests. Unfortunately, no accuracy assessment was conducted.

The complementarity of JERS and ERS radar mosaics to map regional flooding and identify tree swamps and high and low mangroves in Gabon was the subject of Simard et al. (2002). Using a decision tree classifier, accuracy increased to 84% with the merged data set compared to 61% for ERS and 66% for JERS alone. Both images were needed to identify high mangrove and ERS was necessary to define grass swamps. Low mangrove, raffia (palms) and flooded forest could be separated as a group using JERS imagery or the ERS‐JERS merge but could not be identified as individual classes on any data set.

Northern boreal forest environments comprise another ecosystem that has continued to receive the attention of several researchers. Adam et al. (1998) examined two Radarsat Standard mode images (S2 and S6) of the Peace River in Alberta at peak flood. Ground cover was primarily leaf‐off willows 2–5 m above water with 20–50% coverage. They found the steeper incident angle S2 image to be better than the S6 image but both images attained better than 90% accuracy in mapping flooded willows, open water and non‐flooded willows. The authors concluded that leaf‐off, reduced speckle imagery was requisite and that incorporation of image texture reduced accuracy.

Land cover of western Ontario containing cattails, Phragmites (common reed) and swamp‐treed wetlands was examined with multidate ERS imagery (Wang et al. 1998). Using multidate imagery, wetland accuracy increased to 85% compared to 51% for the best single‐date image, but the authors also reported that adding more than five images decreased accuracy. While the detection of cattails and swamp increased with multiple dates, the detection of the narrow lines of phragmites remained low (ca. 47%).

A similar environment in southern Ontario was studied by Arzandeh and Wang in 2002 and 2003 using an unspecified fine‐mode Radarsat image. Using the grey‐level co‐occurrence matrix (GLCM) texture‐measuring algorithm, Arzandeh and Wang (2002) delimited eight classes (Cattail marsh, Phragmites, Marsh, Swamp, Tall grass prairie, Upland forest, Built‐up area, Agriculture and Water) with 71% overall accuracy. However, two of the three wetland classes were poorly classified regardless of how many texture features were used to assist in the classification. The highest accuracies for each of the three wetland classes, Cattail, Phragmites and Swamp, were 76%, 19% and 16%, respectively, and occurred with the use of four texture features (or measurements). Except for Cattail, this method is not accurate enough for wetland mapping. When the eight classes were reduced to two classes, wetland and non‐wetland, the overall accuracy increased to 88%, indicating that single‐date, single‐polarized imagery can be used to delineate wetlands from non‐wetlands in this instance with the aid of multiple feature textural analysis.

Following on from their earlier work, Arzandeh and Wang (2003) then attempted to use SAR to fill the cloud cover gaps in optical data to map Phragmites incursion into cattail wetlands. Again, they found that although the addition of the SAR texture analysis slightly improved the overall land cover classification, it did not improve the accuracy of the wetland categories.

Yatabe and Leckie (1995) compared the ability of JERS‐1, ALMAZ and ERS imagery to discriminate clearcuts from other forested land including wetlands in eastern Ontario. Although the wetlands included sites of aquatic vegetation, sedges and various densities of black spruce, wetlands were reported as a single category when compared to clearcuts. Similar to earlier studies, JERS‐1 L‐band imagery produced higher backscatter from flooded forest than the shorter wavelength imagery from ALMAZ and ERS. High backscatter was also noted on the S‐band HH ALMAZ imagery but the contrast with general forest was less than that observed with L‐band. Discrimination was not evident on the ERS imagery. JERS‐1 and ALMAZ imagery produced the greatest separability between wetlands and clearcuts but contrast decreased as clearcut density increased. ERS imagery generated considerable signal overlap and large signal variation within and between wetlands and clearcuts.

Six standard mode Radarsat images of the wetlands in Hudson Bay lowland were used by Murphy et al. (2001) to look at seasonal variations and incident angle differences. Wetland types in the study included coastal marshes, coastal fen, treed fen, shrub fen, and bog. Descriptive statistics for these wetland types were provided for the six dates of radar coverage. However, the accuracy assessment for the unsupervised classification was only reported for wetlands and not the six individual types. Almost no association was found between a changing incident angle and radar backscatter and the May S1 and S7 images produced similar returns. The authors also reported a greater increase in backscatter values between the May S7 image and the June S1 image than could be attributed to incident angle changes. Rather, the differences were thought to be due to snowmelt with accompanying increases in surface roughness and vegetation moisture. Overall, the highest wetland backscatter values (except coastal marsh) were found on spring images; as the seasons progressed, backscatter values declined. Coastal marsh returns, however, were lowest in spring and highest in November. Here, the results were attributed to the presence of snow cover, ice, and destructed vegetation conditions in the spring.

Radarsat (S1, S2 and S7) and SPOT data of Saskatchewan were used to examine variations in incident angle, season‐canopy cover, and SAR/optical merges for flood mapping (Töyrä et al. 2001, Töyrä and Pietroniro 2005). Large incident angle imagery was needed to identify open water because high backscatter from wave action caused confusion on the steep angle imagery, but steep incident angles were preferred in identifying flooded dead willows, dense flooded grasses and sedges. However, detecting grasses and sedges was possible only in the summer when new growth stalks were upright; in the spring, brown vegetation and thatch attenuated the signal, minimizing double‐bounce. Individually, SAR and SPOT imagery produced poor accuracies for mapping three land cover categories (open water, flooded vegetation, and non‐flooded land), but the Radarsat S1 and SPOT combination contributed complementary data that produced significantly better classification results (Töyrä et al. 2001). Using a combination of Radarsat, Landsat or SPOT and Lidar, Töyrä and Pietroniro (2005) generated a flood duration map for 1996 to 2001 and compared it with the satellite based land cover map. As might be expected, this comparison illustrated a relationship between the spatial distribution of vegetation classes and the spatial‐temporal pattern of flooding. Aquatic types of vegetation were primarily found in areas of persistent inundation, suggesting that the use of flood duration maps from radar data and other sources should help to improve the delineation of wetlands.

In earlier work Brown et al. (1996) had advocated large incident angles (Radarsat Standard Modes S4‐S7) to map flooded vegetation in their review of the GlobeSAR Program, but they also advised using multitemporal data. This work and also others cited above using steep incident angle or steep and moderate incident angle data have produced contrary and mixed results and recommendations to date.

Cihlar et al. (1992) had noted earlier that marshes acted like an annual crop in winter and a woody canopy in summer on C‐band HH and VV imagery. They commented that the high summer response was due to a combination of the high biomass of coarse plants (cattails) and double‐bounce from plants standing in water. In winter, the marshes consisted of dead biomass above ice that was covered by snow.

Using a series of 12 ERS‐1 images of Ottawa, Canada, Crevier et al. (1996) added to these observations and understanding in their investigation of hydrologic applications and the effects of environmental conditions on radar backscatter. Backscatter from wetlands (comprising tree swamps and marshes) was found to be more sensitive to frost development in winter than to sudden moisture variations caused by precipitation. Wetland backscatter was stable from July to October and similar to that of forests but increased in November in leaf‐off conditions. Late autumn was proposed as the best time to detect wetlands based on these observations.

Concurrently, Morrissey et al. (1996) reported similar findings, exploring 24 ERS‐1 scenes of herbaceous sedges and grasses on the Alaskan North Slope over two growing seasons. Backscatter differences between wetlands and non‐wetlands were minimal during the growing season. In this area of low LAIs, double‐bounce did not separate wetlands from non‐wetlands. Rather, the maximum backscatter differences occurred when air temperatures were near or below freezing, preceded by non‐freezing conditions. Increased contrast was also observed when non‐wetlands were frozen but wetlands were not. The backscatter from land and wetlands was similar when wetlands were frozen.

More recent studies have also addressed the use of radar in detecting and mapping wetlands in the high latitudes of North America (Li and Chen 2005, Racine et al. 2005). Stating that peatlands occupy over 90% of Canada's wetlands, Racine et al. (2005) evaluated multitemporal fine and standard mode Radarsat data for classifying boreal peatlands and researched the effect of hydrological conditions on σ⁰ from peatland. Using seven image dates, three of which were supported by field data collected during June and August 2003 in concert with Radarsat overpasses, the James Bay, Canada, study area was divided into four classes (Open peatlands, Forested peatlands, Water and Mineral). The authors reported that the standard mode Radarsat data resulted in an improved classification accuracy over the fine mode. This was not expected but explained by a plethora of reasons followed by a statement that they could not recommend standard mode imagery. Their assessment of the influence of soil water on radar backscatter reached similar conclusions. The authors determined that additional research with more radar imagery, polarization schemes, and field data were required.

Knowledge‐based decision rules and a trio of data sets (Radarsat‐1, Landsat 7 ETM+ and DEM data) were used to help map wetlands in eastern Canada into five classes (Li and Chen 2005). Radarsat‐1 was collected for May and August of 1999, 2001 and 2002 at an incident angle of 45°. The wetland classes along with their corresponding accuracies were: Open bog 92%; Treed bog 71–87%; Marsh 84–89%; Swamp 79%; and Open fen 83%. Although the authors concluded that the addition of radar data to the optical data helped to delineate shrub/tree wetlands from shrub/forest, it also may have a negative affect on overall land cover classification accuracy because of the introduction of speckle and noise from the radar data.

A review of most of the recent research tends to support and reinforce our knowledge base acquired and accumulated over the years regarding wetland–SAR parameters. However, Baghdadi et al. (2001) working in Ontario presented a complex mix of findings. Six cover types (forested peat bog; non‐forested peat bog; marsh; open water; clearing; and forests) were mapped using decision tree analysis of C‐band airborne SAR obtained three times over the growing season. Among the findings were: (1) backscatter from marsh was highest on HH images for all three dates; (2) VV consistently produced the highest backscatter from both non‐forested and forested bog; (3) only HV data could separate forested bog, non‐forested bog and forest from each other; (4) marsh was most easily separated from all other land cover on HH imagery; and (5) HV data produced the best overall classification. The authors deduced that single‐date multipolarization data were necessary to separate each land cover from the others. Contrary to many other studies, multidate imagery did not improve the results attained from the best single‐date (October) imagery.

The southern US has served as a third area of continued research concentration although the number of studies has diminished recently in comparison to work in the areas reported previously. Bourgeau‐Chavez et al. (2001) used spring and fall steep incident angle (25°) SIR‐C data to evaluate multiwavelength and multipolarization data of southeastern Virginia wetlands. In contrast to Baghdadi et al. (2001), they found April data better than the October data to identify emergent and forested wetlands. In Virginia there was confusion between agriculture and emergent wetlands on the October imagery, a problem not faced in the Ontario area. Although L‐HH, C‐HH and C‐HV imagery had the highest accuracy, again pointing to the advantage of like‐polarization data, multipolarized L‐ and C‐band data were needed to separate all land cover types. The authors stated that their overall findings were similar to Hess et al. (1995) and Wang et al. (1995) in that all three studies preferred HH steep incident angle imagery. More importantly, the results were found to differ significantly by study area, ranging from 20% to 100% accuracies for some wetland types.

Flooding in south Florida wetlands (Kasischke et al. 2003) was the focus and extension of earlier work by Kasischke and Bourgeau‐Char (1997). Thirteen study sites and three wetland types on 22 ERS images obtained over a 25‐month period were examined to delimit mean flood levels. The MIMICS (MIchigan MIcrowave Canopy Scattering) model prediction of a decrease in backscatter with an increase in water depth was observed in both non‐wooded and low‐density wooded wetland sites. The authors also observed large changes in backscatter as a result of differences in soil moisture but saw no backscatter sensitivity to variation in above‐ground biomass. While the group concluded that ERS imagery could be used to monitor regional‐scale flooding conditions in south Florida on a seasonal basis, they also included a degree of caution. Specifically, the south Florida wetlands consisted of relatively short canopies, low biomass levels, and small stem sizes and other work had produced different results under higher biomass conditions.

Sensitivity to biomass variations was observed in another Florida‐based study. Stages of regrowth in burned black needlerush marshes were measurable using L‐band airborne SAR data (Ramsey et al. 1999). As recovery and vegetation growth increased, initially high VV returns decreased while HH and VH returns rose. Based on this relationship it was possible to monitor post‐burn regrowth with VH data.

Interferometric and polarimetric airborne SAR data analysis of coastal Texas used C‐ and L‐band linear, phase difference, and circular polarizations (Crawford et al. 1999). An array of neural network, Baysian pairwise classifiers was applied to multiresolution and multisource (AIRSAR and TOPSAR) radar data to map low proximal marsh (Spartina alterniflora); high proximal marsh (S. alterniflora and Salicornia virginica); high distal marsh (Spartina patens, S. virginica, and Juncus roemerianus); and brackish marsh, among other land cover categories. Although the incident angles and exact image combinations were not provided, each of the 13 classifiers tested produced impressive mapping accuracies for these classes ranging from 87% to 100%.

Hurricane‐related flooding of the Louisiana coastal zone was considered using three Radarsat‐1 images collected between 23 September 2002 and 3 October 2002 at incident angles between 29° and 38°. The October image was acquired 8 to 9 hours after peak water levels from Hurricane Lili (Kiage et al. 2005). Although the research did not directly address wetland mapping, it did assess the mapping of flooded vegetation. Findings indicated that: (1) differences in radar backscatter patterns were related to slight elevation differences and vegetation changes with elevation; and (2) the difference image generated from two reference images (different dates) was the best method for delineating marsh flooding. From these observations it follows that if a familiarity with vegetation helps to detect flooding, then vegetation mapping can be improved by an increased knowledge of flooding.

Flooded Georgia forest composed mainly of tupelo gum (Nyssa aquatica), black gum (N. biflora) and bald cypress (Taxodium distichum) served as the study area for Wang and Davis (1997). Total power images of AIRSAR polarimetric C‐, L‐ and P‐band imagery were decomposed into three backscatter types: single reflection, double reflection, and cross‐polarized backscatter. Double‐bounce (and canopy penetration) increased from flooded forest as wavelength increased from C‐ to L‐ to P‐band data. Cross‐polarized backscatter was not as strong as the like‐polarized signal, but the authors also noted that the cross‐scatter return decreased as wavelength increased from C‐ to P‐band. Significantly, their polarimetric work was consistent with that of earlier efforts mentioned above that used single‐polarized imagery.

Neural network analysis of AIRSAR imagery separated wetland forests from upland forests on Maryland's Delmarva Peninsula (Augusteijn and Warrender 1998). C‐ L‐ and P‐band data in HH, VV and VH polarizations, along with total power in each band and with incident angles from 21° to 34°, were unable to categorize wetlands into four groups based on wetness but were able to separate ‘least wet’ from ‘most wet’ wetlands and upland. The ‘most wet’ category attained the highest accuracy (79%) while the ‘least wet’ category was confused with the two other wetland categories. The highest backscatter value from the ‘most wet’ class was observed on L‐band Total Power and P‐band HV polarized data. These results again substantiate the use of steep to moderate incident angle, long wavelength imagery to identify forested wetlands. However, in contrast to other works discussed earlier, there was a preference here for total power (TP) and HV polarization for the detection of forested wetlands per se rather than HH polarized data. Although the neural networks performed well in this instance, accuracies (about 72% overall) seem comparable to other studies using other, less complex methods.

Three studies examined North Carolina forested wetlands. Using 11 Radarsat scenes of a North Carolina floodplain, Townsend (2001) found that flood inundation mapping of forests could be performed regardless of season or water level but leaf‐off was the most accurate (98.1%). However, non‐forest areas were masked from the scenes prior to classification, greatly simplifying the land cover complexity and confusion opportunities. Although previous work such as those cited above had stated a preference for a steep incident angle, it made little difference here. All the standard mode images, ranging from S1 (23°) to S2 (27°) to S6 (44°) incident angles, produced satisfactory results albeit with less marked contrast apparent on the S6 image. Radarsat C‐HH data performed well here, but the author cautioned that tropical regions with denser and more complex forests might prove more difficult to map. A later study by Townsend (2002) of the same area to assess the role of forest structure using ERS C‐VV imagery found it incapable of mapping the extent of flooded forests. Townsend concluded that species composition was not a contributor but basal area and height to the bottom of the canopy did have considerable influence in mapping accuracy; the larger the basal area and the taller the forest, the greater the double‐bounce scatter. Inundated successional forests proved particularly difficult to detect.

Wang (2004) used two JERS‐1 images and a decision tree classifier to map North Carolina floodplain inundation into five categories: water, marsh, flooded forest, field, and non‐flooded forest. Although flooded forest could be accurately mapped, the classifier could not satisfactorily identify marsh, water, or field. Classification was also uncertain in transition zones.

In one of the few studies to include urban areas, several fine and standard mode Radarsat images, one ERS image, and a TM image of Long Island, New York, were evaluated to determine the synergy of combining SAR data to optical images for coastal land cover mapping (Henderson et al. 2002). Wetland categories consisted of estuarine emergent, palustrine emergent, palustrine scrub shrub, and palustrine forested. Incident angle was less important here than spatial resolution; the fine mode image merge produced the best results. No merge combination produced acceptable operational (ca. 85%) wetland mapping accuracies although merging SAR with TM data did improve results. The best merge technique also varied among the wetland categories. Two differences between this study and other recent work should be noted. First, the expanding urban/suburban New York setting for this study was much more spatially complex than the simple environments used in most of the other work reported here. Second, the wetland categories mapped were much more detailed than most other work and were generally found in smaller parcels than other study areas.

Investigators have looked at other wetland areas of the world, but the body of research is less extensive. Milne et al. (2000) visually examined freshwater wetlands of northern Australia with Standard (S4) and ScanSAR mode C‐HH Radarsat imagery obtained in wet and dry periods. They concluded that general landscape differences could be best seen in a dry season image. The S4 (25m spatial resolution; 34–40° incident angle) imagery could be used to detect flood conditions under forest canopy and macrophytes. During the wet season the boundary between floodplain and woodland was obscured because both were covered by rainwater and had high surface dielectric constants. General changes in flooding extent were observable as areas went from dry to wet conditions.

Rio and Lozano‐Garcia (2000) compared 18 low‐pass spatial filters to ascertain the preferred method to eliminate speckle and improve detection of brackish marsh in Mexican wetlands. An S1, an S7 and a combined S1–S7 Radarsat image with 8 m pixel spacing were filtered, followed by supervised classifications using five land cover classes (water, pasture, crop, bush and brackish marsh). They found that the Lee Sigma filter with three iterations and three, five and seven square window sizes produced the highest marsh detection accuracy (95%); this compared to 79% accuracy for the unfiltered images.

The utility of SAR to improve waterline mapping was studied by Horritt et al. (2003). Flooded salt marshes on the east coast of the UK exhibited increased backscatter of approximately 1.2 dB on both airborne C‐VV imagery and HH‐VV phase difference L‐band images acquired at 3000 m altitude. The inclusion of SAR‐derived flooding extent under vegetation canopy data improved determination of actual waterline height and location compared to the use of only open water location boundaries but the authors cautioned that the complicated nature of their data and analysis might limit its use in other environments. Although primarily concerned with improving the mapping of the waterline, their results provide insight and information into mapping emergent (flooded) vegetation in a salt marsh environment.

An examination of Germany's Elbe River with various polarization configurations produced recommendations for rapid flood mapping (Henry et al. 2006). ERS‐2 data and single (HH) and alternating polarization modes of Envisat data with similar incident angles were analysed with TM interpretations serving as an accuracy reference. HH data were found superior to HV or VV data in delimiting flood extent but HV did contribute important information in flood detection; the HH/HV combination was recommended for crisis mapping.

Optical examinations of ERS‐1 coastal wetland data in West Bengal indicated that they could be used to identify non‐vegetated from sparsely vegetated wetlands (Dwivedi et al. 1999) and to update wetland maps created 20 years ago from Landsat MSS data (Rao et al. 1999). ERS‐1 data were also used to map the entire Congo River basin into swamp forest or lowland rain forest (DeGrandi et al. 2000). The results were then compared to earlier maps derived from Landsat TM data. Although the SAR map attained a 71% overall accuracy, DeGrandi et al. (2000) suggested that the actual accuracy may be higher because the SAR‐based map provided more detail and recorded land cover change as well. They concluded that such large‐area thematic maps can be completed with one‐pass imagery and provide important ecological information to inaccessible areas. In this suite of studies, satellite SAR imagery occupied an applied role, seemingly used as an accepted analysis technique in an operational mode. SAR imagery collected at a general level of detail can be, and is being, applied to large areas, providing data not available previously.

A comparison of amplitude change detection with interferometric coherence‐based change detection using ERS‐1 imagery of southern France found that both threshold‐based methods tended to over‐ and underestimate the extent of flooding (Nico et al. 2000). By combining data sets, the difference between the extent of flooding noted on the SAR map and the actual extent was 300 m. However, there was no mention of terrain conditions prevalent in these ‘difference’ areas.

A few articles have analysed radar imagery from a different perspective than the articles cited above. Moreau and LeToan (2003) and Kaya et al. (2004) did not focus on SAR–environmental parameter, radar backscatter variations, or classifier development. Rather, they used SAR‐derived information as an accepted applied research tool and resource for geoscience investigations. Moreau and LeToan (2003) analysed 12 ERS scenes of the Bolivian Andes to derive biomass information on wet grasslands. Proper forage and livestock resource management is important in these mountain wetland pastures but essential spatial information on pasture conditions is lacking. Radar backscatter‐derived biomass data provided key input to livestock management efforts at three crucial periods in the annual cycle (fall, spring and winter). It is interesting that calculating biomass in the Andes is seemingly facile while no sensitivity to above‐ground biomass was observed in south Florida (Kasischke et al. 2003). As both studies used ERS data, the difference must be due largely to environmental modulation.

Kaya et al. (2004) developed a method to track malaria risk in Kenya using SAR data. They first used an object‐oriented classifier with multitemporal S7 Radarsat imagery to delimit wet and dry season wetland habitats (mosquito‐disease breeding grounds), settlements, agriculture and forests. A proximity/distance map linking settlement distance to wetlands identified malaria risk areas where health providers could then concentrate disease control efforts. The authors then described how Radarsat‐2 would improve these and other environmental/epidemiological applications.

Applications related to environmental monitoring of surface mining were reported in two studies. Limpitlaw and Gens (2006) used the environmental effects on dambos, a South African term for seasonal grassy wetlands in central South Africa, as an indicator of the pollution resulting from mining in the Zambian Copperbelt. Although both TM and ERS‐2 imagery could identify natural dambos, neither could identify disturbed dambos. A fusion of the two highlighted the unique complementary spectral information of each sensor and produced a reliable environmental indicator of mining disturbance. In the second case, a combination of TM, SPOT and Radarsat imagery formed a useful tool to locate water‐filled artificial lakes produced as an after‐effect of abandoned strip mines along the Turkish coast (Uca et al. 2006).

5. Conclusions and observations

Recent research involving ERS, JERS and Radarsat satellite data sets as well as airborne systems has reinforced earlier trends and patterns but also produced a mixed bag of contradictory results. As we add to data availability, we often find new questions and less consistency and generalization. The conclusions and observations gleaned from an analysis of recent research reports are organized as follows by wavelength, polarization, incident angle, temporal, environment and analysis techniques.

5.1 Wavelength

Although longer wavelengths are preferred for detection of forested wetlands, other methods seem in need of re‐examination. Most research suggests that L‐band is necessary for forest and C‐band for herbaceous wetland detection, as was also noted consistently in the earlier research reviews. However, both L‐ and C‐band data are needed to separate forest from herbaceous wetland as well as from other land cover. C‐band data has produced favourable results in some wooded wetlands but not in others. Current research is supportive of earlier predictions that C‐band would do best in leaf‐off and low biomass conditions. In general, multiple wavelengths seem necessary for mapping consistency and accuracy.

5.2 Polarization

HH polarization is preferred over VV but cross‐polarized data seem to contribute more than was earlier considered in some cases. Total polarimetric polarization seems comparable to HH in accuracy. Some combination of multipolarized imagery consistently outperforms single polarization. Multipolarization imagery is frequently as good as or better than multitemporal imagery.

5.3 Incident angle

The preference for a steep incident angle seems to vary as a function of study area and environment under assessment; a moderate incident angle does seem to provide useful data under certain conditions, especially if it is of a higher spatial resolution. Incorporation of multiple incident angles did not always improve accuracy and should not be expected to with rough targets such as tropical rainforests. However, the closer the target is to becoming a specular reflector, the more variation, and therefore the more information multiple incident angle data should provide.

5.4 Temporal

Although single‐date, single‐polarization imagery is not adequate in most cases, adding dates only increases accuracy to a point beyond which further data sets contribute nothing or diminish accuracies. There is no one set of optimum image acquisition dates. The selection of optimum date/time is probably as important as the number of image requisitions.

5.5 Environment

Type and complexity of environment play an extremely important role. Environments of few cover types with strongly contrasting morphologies in large homogeneous patches are generally readily and accurately mapped. Simple environments or the ability to create simple environments (e.g. masking off all non‐forest land cover) are the easiest to identify and classify. As complexity and land cover mix increase and parcel size decreases, accuracy falls. Mapping wetlands per se is much easier than mapping wetland species. However, identification confusion is often greater among the wetland types themselves than between wetlands and other land cover types.

Optimum time of year for image acquisition varies with study area and environment. Mapping flood inundation and its seasonal variation has also been a focus of recent investigations. The results in this regard, not surprisingly, mirror the results of wetland detection and wavelength. That is, the ability to delimit surface water is a function of the wetland type (forest or herbaceous) and condition of foliage (leaf‐on/leaf‐off), where standing water is present and therefore dictates the optimum wavelength needed (C‐ or L‐band) to detect the water extent.

From a different perspective, inconsistencies in the definitions of land cover classes and the assignment of land cover classes based on field or image data add to consistency and accuracy problems of radar‐based wetland mapping. Flooded forests, perhaps the least complex of wetland categories, can still be a highly variable land cover class. When is a forest that is inundated a flooded forest and not open water? Should it be at 1%, 30% or 60% canopy closure?

5.6 Analysis techniques

This review points to a lack of consistency in recommended analysis techniques. Traditional supervised hard classifiers, neural networks, knowledge‐based and expert systems, scattering indices, and hierarchical classifiers have all been used and compared with one another. Images have been analysed using digital numbers as well as σ⁰ dB backscatter coefficients. The results and recommendations vary between studies as well as between different environments. The success of the analysis technique appears to be more dependent on (1) sensor specifications (wavelengths and polarizations); (2) project criteria (minimum mapping unit and detail of mapped categories); and (3) environmental characteristics (modulation transfer function) than the technique. The same can be said for the filters used to minimize the effects of speckle and prepare the data for digital or visual analysis. Advanced image analysis techniques may not be critical because simple approaches seem comparable in results, both positive and negative, to more complex methods.

SAR/optical sensor merges always produced better results than either data set alone, but there was no consistent pattern or combination that was always the best.

As shown throughout the article, much has remained consistent and been reinforced over the past 20 to 30 years of research, but there is still much to be determined with regard to the role of SAR in mapping wetlands and wetland ecosystems. Thus, while some trends are beginning to emerge with regard to which wavelengths or polarizations to use, we see more interest in the variation in season/time of year and the incorporation of multitemporal imagery. HH polarization is preferred if using only one polarization scheme but multitemporal sets are then requisite. Some observations seem to be case study and local environment dependent.

New trends are evolving. What is evident throughout the recent literature is that multidimensional radar data sets are attaining an accepted role in operational situations needing information on wetland presence, extent, and conditions. Wetland mapping with radar imagery is no longer confined to research. That trend will undoubtedly accelerate as satellite SAR imagery availability and variety continue to increase. Mapping flood inundation is now almost routine at regional scales. The use of SAR imagery‐derived data for environmental applications (e.g. disease mapping, range management, pollution) has appeared on the horizon and will continue to evolve into an accepted research tool. At the same time, continued research into the system parameters and other applications will only expand.

Radar plays an important part in the scientific research conducted to better understand and protect our fragile wetland environments. The ability to collect data day or night and under almost all weather conditions with spatial resolutions from spacecraft altitudes better than almost any multispectral scanner makes radar systems unique as an earth resource remote sensor. In addition, radar provides information on the physical characteristics from and occasionally below the surface land cover that is not available at other portions of the electromagnetic spectrum. As the collection of radar data with more discrete wavelength bands, more polarizations schemes and at more incident angles progresses, the information value of radar sensors continues to increase.