High-latitude wetlands are important for understanding climate change risks
because these environments sink carbon dioxide and emit methane. However,
fine-scale heterogeneity of wetland landscapes poses a serious challenge
when generating regional-scale estimates of greenhouse gas fluxes from point
observations. In order to reduce uncertainties at the regional scale, we
mapped wetlands and water bodies in the taiga zone of The West Siberia
Lowland (WSL) on a scene-by-scene basis using a supervised classification of
Landsat imagery. Training data consist of high-resolution images and
extensive field data collected at 28 test areas. The classification scheme
aims at supporting methane inventory applications and includes seven wetland
ecosystem types comprising nine wetland complexes distinguishable at the
Landsat resolution. To merge typologies, mean relative areas of wetland
ecosystems within each wetland complex type were estimated using
high-resolution images. Accuracy assessment based on 1082 validation
polygons of 10
High-latitude wetlands are important for understanding climate change mechanism as they provide long-term storage of carbon and emit a significant amount of methane. The West Siberia Lowland (WSL) is the world's largest high-latitude wetland system and experiences an accelerated rate of climate change (Solomon et al., 2007).
Poorly constrained estimates of wetland and lake area constitute a major
uncertainty in estimating current and future greenhouse gas emissions
(Melton et al., 2013; Turetsky et al., 2014; Petrescu et al., 2010).
Although wetland extent in WSL has been reasonably well captured by global
products based on topographic maps (Lehner and Döll, 2004;
Matthews and Fung, 1987), mapping of fine-scale heterogeneity of WSL's
wetland landscapes (Bohn et al., 2007) requires adding fine
scale information on ecosystem functioning as made in wetland CH
Modelers simulating methane emission are in need for high-resolution wetland maps that do not only delineate wetlands but also identify the major sub-types to which different environmental parameters could potentially be applied (Bohn et al., 2015). Several wetland maps have been used to define the wetland extent in WSL, however their application to net primary production (NPP) and methane emission inventories was accompanied by difficulties due to crude classification scheme, limited ground truth data and low spatial resolution. One peatland typology map that distinguishes several vegetation and microtopography classes and their mixtures was developed at the State Hydrological Institute (SHI) by Romanova et al. (1977). Peregon et al. (2005) digitized and complemented this map by estimating the fractional coverage of wetland structural components using Landsat images and aerial photographs for five test sites. However, the limited amount of fractional coverage data and coarse resolution still result in large uncertainties in upscaling methane fluxes (Kleptsova et al., 2012).
Our goal was to develop a multi-scale approach for mapping wetlands using Landsat imagery with a resolution of 30 m so the results could better meet the needs of land process modelling and other applications concerning methane emission from peatlands. In this study, the WSL taiga zone was chosen as the primary target for the land cover classification due to wetland abundance. The objectives were the following: first, to develop a consistent land cover of wetland classes and its structural components; second, to provide the foundation for environmental parameter upscaling (greenhouse gas inventories, carbon balance, NPP, net ecosystem exchange, biomass, etc) and validation of the process models.
The West Siberian Lowland is a geographical region of Russia bordered by the
Ural Mountains in the west and the Yenisey River in the east; the region
covers 275 Mha within 62–89
No single classification algorithm can be considered as optimal methodology for improving vegetation mapping; hence, the use of advanced classifier algorithms must be based on their suitability for achieving certain objectives in specific applications (Adam et al., 2009). Because mapping over large areas typically involves many satellite scenes, multi-scene mosaicking is often used to group scenes into a single file set for further classification. This approach optimizes both the classification process and edge matching. However, large multi-scene mosaicking has essential drawback when applying to highly heterogeneous WSL wetlands. It creates a variety of spectral gradients within the file (Homer and Gallant, 2001), especially when the number of the appropriate scenes is limited. It results in spectral discrepancy that is difficult to overcome. In this study, the advantages of consistency in class definition of the scene-by-scene classification approach were considered to outweigh the inherent disadvantages of edge matching and processing labor. Thus, our entire analysis was performed on a scene-by-scene basis, similar to the efforts by Giri et al. (2011) and Gong et al. (2013).
Wetland complexes (I – Pine bog or ryam, II – Ridge-hollow complex or RHC, III – Ridge-hollow-lake complex or RHLC, IV – Lakes and rivers, V – Open fens, VI – Patterned fens, VII – Swamps, VIII – Palsa complexes) and ecosystems in WSL (1 – Open water, 2 – Waterlogged hollows, 3 – Oligotrophic hollows, 4 – Ridges, 5 – Ryam).
For land cover consistency, data of the same year and season, preferably of the growing season peak (July) are required. However, the main complication was the low availability of good quality cloudless images of WSL during those periods. Scenes collected earlier than the 2000s were very few, so they were used as substitutes for places where no other suitable imagery could be found. Landsat-7 images received after 2003 were not used due to data gaps, while Landsat-8 was launched after starting our mapping procedure. Finally, we collected 70 suitable scenes during the peak of the growing seasons in different years. Majority of the images were Landsat 5 TM scenes from July 2007. The scene selection procedure was facilitated by the ability of smoothing the slight inconsistencies between images by specifying training sites in overlapping areas.
Wetland ecosystem types.
The overall work flow involves data pre-processing, preparation of the
training and test sample collections, image classification on a
scene-by-scene basis, regrouping of the derived classes into nine wetland
complexes, the estimation of wetland ecosystem fractional coverage and
accuracy assessment. Atmospheric correction was not applied because this
process is unnecessary as long as the training data are derived from the
image being classified (Song et al., 2001). All of the images were
re-projected onto the Albers projection. Because the WSL vegetation includes
various types of forests, meadows, burned areas, agricultural fields, etc.,
wetland environments were first separated from other landscapes to avoid
misclassification. We used thresholds of the Green–Red Vegetation Index
(Motohka et al., 2010) to separate majority of wetlands and forests.
Surface water detection was performed using thresholds applied to Landsat's
band 5 (1.55–1.75
Training data play a critical role in the supervised classification technique. Representative data preparation is the most time-consuming and labor-intensive process in regional scale mapping efforts (Gong et al., 2013). As a primary source of information, we used the extensive data set of botanical descriptions, photos, pH and electrical conductivity data from 28 test sites in WSL (Glagolev et al., 2011). Due to vast expanse and remoteness of WSL, we still had a lack of the ground truth information, which hampered training data set construction. As a result, we had to rely mostly on the high-resolution images available from Google Earth. They came from several satellites (QuickBird, WorldView, GeoEye, IKONOS) with different sensor characteristics; multispectral images were reduced to visible bands (blue, green, red) and had spatial resolution of 1–3 m. The processing started with mapping scenes where ground truth data and high-resolution images are extensively available, so the classification results could be checked for quality assurance; mapping continued through adjacent images and ended at the less explored scenes with poor ground truth data coverage.
To collect training data most efficiently, we used criteria similar to those
used by (Gong et al., 2013) for training sample selection, (i) the
training samples must be homogeneous; mixed land-cover and heterogeneous
areas are avoided; and (ii) all of the samples must be at least 10 pixels in
size with an average sample area of approximately 100–200 pixels. The
Bhattacharyya distance was used as a class separability measure. However,
the classifier was designed using training samples and then evaluated by
classifying input data. The percentage of misclassified samples was taken as
an optimistic predication of classification performance (Jain et al.,
2000). When accuracy of more than 80 % across the training set was
attained with no fields showing unreasonable or unexplainable errors, the
classification process was started. Classification mismatch between scenes
was minimized by placing training samples in overlapping areas. Combining
the classified images and area calculations were made using GRASS module in
Quantum GIS. Noise filter was applied to eliminate objects smaller than
2
As a starting point for the mapping procedure, a proper classification
scheme is required. Congalton et al. (2014) showed that the
classification scheme alone may result in largest error contribution and
thus deserves highest implementation priority. Its development should rely
on the study purposes and the class separability of the input variables. In
our case, wetland mapping was initially conceived as a technique to improve
the estimate of the regional CH
The wetland ecosystems generally have sizes from a few to hundreds of meters and cannot be directly distinguished using Landsat imagery with 30 m resolutions. Therefore, we developed a second wetland typology that involves 9 mixed “wetland complexes” composing wetland ecosystems in different proportions (Fig. 1; Table 2). The classification was adapted from numerous national studies (Katz and Neishtadt, 1963; Romanova, 1985; Liss et al., 2001; Lapshina, 2004; Solomeshch, 2005; Usova, 2009; Masing et al., 2010) and encompassed wooded, patterned, open wetlands and water bodies. The criteria for assigning wetland complexes were the following: (i) separability on Landsat images, and (ii) abundance in the WSL taiga zone. Each wetland complex represents integral class containing several subtypes differing in vegetation composition and structure. Subtypes were mapped using Landsat images and then generalized into final nine wetland complexes based on ecosystem similarity and spectral separability.
Wetland types and fractional coverage of wetland ecosystems (Open water – W, Waterlogged hollows – WH, Oligotrophic hollows – OH, Ridges – R, Ryams – Ry, Fens – F, Palsa hillocks – P).
To merge typologies, we estimated relative areas of wetland ecosystems
within each wetland complex of the final map. Depending on heterogeneity, 8
to 27 test sites of 0.1–1 km
During wetland typology development, we made several assumptions; (i) the
wetland complexes were considered as individual objects, while they actually
occupy a continuum with no clustering into discrete units. (ii) All of the
wetland water bodies originated during wetland development have sizes less
than 2
The concept of wetland ecosystems has merits on CH
Wetland map
Based on Landsat imagery, we developed a high-resolution wetland inventory of the WSL taiga zone (Fig. 2). The total area of wetlands and water bodies was estimated to be 52.4 Mha. West Siberian taiga wetlands are noticeable even from a global perspective. The global total of inundated areas and peatlands was estimated to cover from 430 (Cogley, 1994) to 1170 Mha (Lehner and Döll, 2004) as summarized by Melton et al. (2013); therefore, taiga wetlands in WSL account for approximately from 4 to 12 % of the global wetland area. Their area is larger than the wetland areas of 32.4, 32, and 41 Mha in China (Niu et al., 2012), Hudson Bay Lowland (Cowell, 1982) and Alaska (Whitcomb et al., 2009), respectively. The extent of West Siberia's wetlands exceeds the tropical wetland area of 43.9 Mha (Page et al., 2011) emphasizing the considerable ecological role of the study region.
As summarized by Sheng et al. (2004), the majority of earlier Russian studies estimated the extent of the entire WS's mires to be considerably lower. These studies probably inherited the drawbacks of the original Russian Federation Geological Survey database, which was used as the basis for the existing WSL peatland inventories (Ivanov and Novikov, 1976). This database suffered from lack of field survey data in remote regions, a high generalization level and economically valuable peatlands with peat layers deeper than 50 cm were only considered.
Our peatland coverage is similar to the estimate of 51.5 Mha (Peregon et al., 2009) by SHI map (Romanova et al., 1977). However, a direct comparison between the peatland maps shows that the SHI map is missing important details on the wetland distribution (Fig. 3). SHI map was based on aerial photography, which was not technically viable for full and continuous mapping of a whole region because it is not cost effective and time-consuming to process (Adam et al., 2009).
Latitudinal distribution of wetland ecosystem types.
Comparison of wetland classifications:
Distribution of wetland ecosystem areas have changed significantly in
comparison to SHI map (Peregon et al., 2009); in particular, we obtained a
105 % increase in the spatial extent of CH
WS has a large variety of wetlands that developed under different climatic and geomorphologic conditions. Concerning the wetland complex typology (excluding “Lakes and rivers” class), ridge-hollow complexes (RHC) prevail in WS's taiga, accounting for 32.2 % of the total wetland area, followed by pine bogs (23 %), RHLCs (16.4 %), open fens (8.4 %), palsa complexes (7.6 %), open bogs (4.8 %), patterned fens (3.9 %) and swamps (3.7 %). Various bogs are dominant among the wetland ecosystems (Table 3), while fens cover only 14.3 % of the wetlands. Waterlogged hollows and open water occupy 7 % of the region, which is similar to the estimate by Watts et al. (2014), who found that 5 % of the boreal-Arctic domain was inundated during summer season.
The individual wetland environments have a pronounced latitudinal zonality
within the study region. Zonal borders stretch closely along latitude lines,
subdividing the taiga domain into the southern, middle, and northern taiga
subzones (Fig. 2, black lines). To visualize the regularities of the wetland
distribution, we divided the entire area into 0.1
Wetland ecosystem areas for 0.1
Mire coverage of WSL's northern taiga (62–65
RHCs are the most abundant in the middle taiga (59–62
The wetland extent reaches 28 % in WS's southern taiga area
(56–59
Confusion matrix of West Siberian wetland map validation (additional 11 floodplain and 33 mixed class polygons classified as wetlands are not presented).
The map accuracy assessment was based on 1082 validation polygons of
10
Wetland complexes within large wetland systems had the highest
classification accuracies while the uncertainties are particularly high for
small objects. The southern part of the domain is significant with highly
heterogeneous agricultural landscapes neighbor upon numerous individual
wetlands of 100–1000 ha area. Therefore, several vegetation indices were
tested to map them; however, the best threshold was achieved by using
Landsat thermal band. In addition, many errors occurred along the tundra
boundary due to the lack of ground truth data and high landscape
heterogeneity. However, those small areas mainly correspond to palsa
complexes and have a negligibly small impact on CH
Misclassifications usually occurred between similar classes introducing only a minor distortion in map applications. Patterned fens and open bogs were classified with the lowest producer's accuracy (PA) of 62 %. Patterned fens include substantial treeless areas, so they were often misclassified as open fens. They were also confused with RHCs due to the similar “ridge-hollow” structure. Some open bogs have tussock shrub cover with sparsely distributed pine trees leading to misclassification as RHCs and pine bogs. Open fens have higher user's accuracy (UA) and PA; however, visible channels of high-resolution images poorly reflect trophic state, which underrates classification errors between open bogs and open fens. Swamps and palsa complexes have very high PA and low UA, which is related to their inaccurate identification in non-wetland areas. Palsa complexes were spectrally close to open woodlands with lichen layer, which covers wide areas of WSL north taiga. During dry period, swamps were often confused with forests, whereas in the field they can be easily identified through the presence of peat layers and a characteristic microrelief. In both cases, more accurate wetland masks would lead to substantially higher accuracy levels. Lakes and rivers were well classified due to its high spectral separability. They can be confused with RHLCs represented by a series of small lakes or waterlogged hollows alternating with narrow isthmuses. Floodplains after snowmelt can also be classified as lakes (11 polygons). RHCs and pine bogs were accurately identified due to their abundance in the study region and high spectral separability.
The contrast between vast wetland systems and the surrounding forests is so distinct in WSL that wetlands can be adequately identified by the summer season images (Sheng et al., 2004). On the contrary, correct mapping of wetland with pronounced seasonal variations remains one of the largest challenges. Wetlands become the most inundated after snowmelt or rainy periods resulting in partial transformation of oligotrophic hollows and fens into waterlogged hollows (see hollows with brown Sphagnum cover at Fig. 1). Image features of swamps after drought periods become similar to forests. Interannual variability of water table level in WSL wetlands (Schroeder et al., 2010; Watts et al., 2014) also makes impact on mapping results.
New methodologies and protocols are needed to improve our ability to monitor water levels (Kim et al., 2009). Observations of soil moisture and wetland dynamic using radar data such as PALSAR (Chapman et al., 2015; Clewley et al., 2015) and Global Navigation Satellite Signals Reflectometry are promising (Chew et al., 2016; Zuffada et al., 2015). In addition, advanced classification techniques such as fuzzy logic can be applied for mapping fine-scale heterogeneity (Adam et al., 2009). Recent innovations in wetland mapping were described by Tiner et al. (2015).
Water table fluctuations are particularly important for upscaling CH
Although the synergistic combination of active and passive microwave sensor data is useful for accurately characterizing open water (Schroeder et al., 2010) and wetlands; the remote sensing of water regimes is successful only when in situ data are available for calibration. We still lack in situ measurements of the water table dynamics within WSL wetlands. Limited monitoring has been made at the Bakchar field station (Krasnov et al., 2013, 2015) and Mukhrino field station (Bleuten and Filippov, 2008); however, the vast majority of obtained data have not yet been analyzed and published. These measurements are of special importance for the northern taiga and tundra, where shallow thermokarst lakes with fluctuating water regimes cover huge areas.
The scarcity of reliable reference data and subsequent lack of consistency also limit the accuracy of maps (Homer and Gallant, 2001). The use of ancillary data can largely improve it (Congalton et al., 2014); however, more reliable classification accuracy is attainable with detailed field data. Further improvement in mapping is possible with the acquisition of more ground truth data for the poorly classified wetland types and remote regions.
Boreal peatlands play a major role in carbon storage, methane emissions,
water cycling and other global environmental processes, but better
understanding of this role is constrained by the inconsistent representation
of peatlands on (or even complete omission from) many global land cover maps
(Krankina et al., 2008). In this study we developed a map representing
the state of the taiga wetlands in WSL during the peak of the growing
season. The efforts reported here can be considered as an initial attempt at
mapping boreal wetlands using Landsat imagery, with the general goal to
support the monitoring of wetland resources and upscaling the methane
emissions from wetlands and inland waters. The resulting quantitative
definitions of wetland complexes combined with a new wetland map can be used
for the estimation and spatial extrapolation of many ecosystem functions
from site-level observations to the regional scale. In the case study of
WS's middle taiga, we found that applying the new wetland map led to a
130 % increase in the CH
We estimate a map accuracy of 79 % for this large and remote area. Further improvement in the mapping quality will depend on the acquisition of ground truth data from the least discernible wetland landscapes and remote regions. Moreover, distinguishing wetland complexes with strong seasonal variability in their water regimes remains one of the largest challenges. This difficulty can be resolved by installing water level gauge network and usage of both remote sensing data and advanced classification techniques.
Our new Landsat-based map of WS's taiga wetlands can be used as a benchmark
data set for validation of coarse-resolution global land cover products and
for assessment of global model performance in high latitudes. Although
classification scheme was directed towards improving CH
We thank Amber Soja and anonymous reviewers for assisting in improving the initial version of the manuscript. This study (research grant no 8.1.94.2015) was supported by The Tomsk State University Academic D.I. Mendeleev Fund Program in 2014–2015. The study was also supported by the GRENE-Arctic project by MEXT Japan. Edited by: P. Stoy Reviewed by: three anonymous referees