Wetlands are amongst the most productive and biodiverse ecosystems on Earth [1
]. However, they have been lost at prodigious rates across the globe [2
], and those that remain are imperiled. Junk et al. [3
] estimated that 30–90% of global wetlands have been lost, and that climate change and concomitant temperature and sea level rise, along with precipitation pattern changes, will continue to stress the remaining wetlands. Davidson [2
] reviewed 189 reports of wetland area changes and determined that 64–71% of wetlands have been globally lost since approximately 1900 AD. With wetland areal loss comes loss in various ecological and environmental functions at both local and landscape scales. For instance, wetlands are known areas of high biogeochemical cycling (e.g., [4
]), groundwater recharge and stormflow attenuation (e.g., [6
]), and habitat for many biological species (e.g., [8
]). Wetland processes that underlie these functions vary by habitat or vegetation structure. For instance, emergent (or non-woody) wetlands perform denitrification at different rates than forested wetlands [11
]. Similarly, water storage in depressional wetlands—which decouple storm event flows—differs by wetland habitat [13
]. Understanding wetland abundance and typology is therefore important to properly managing the existing wetland resources and their concomitant watershed functions.
Satellite remote sensing provides a useful mechanism to delineate, assess, and monitor wetland habitats [14
]. Sub-meter to coarse-resolution image data have been analyzed to identify wetlands and demarcate wetland-upland boundaries as well as differentiate habitats within extensive wetland systems (see expansive reviews by [14
]). Critical decisions on imagery acquisition include resolution, spectral bands, revisit period, and cost. The benefits and detractions of various platforms have been—and will continue to be—assessed and debated as an increasingly large number of satellite systems are launched (e.g., [17
Analytical approaches also vary, from visual or manual classification to unsupervised assessments to increasingly complex—and powerful—approaches (e.g., object-oriented classification [18
]; random forest [19
]; and artificial neural networks [20
]). Other approaches include the pixel-based unsupervised Iterative Self-Organizing Data Analysis Technique (ISODATA) [21
] and the supervised maximum likelihood (ML [24
]) image classification techniques of change-detection and pattern recognition [24
In contrast with the parametric ML classifier, more recently developed approaches such as random forest (RF) include non-parametric classification algorithms with no assumption of Gaussian distribution of the input/predictor variables. As a powerful remote sensing image classification tool, RF is applicable for both pixel-based or object-based classifications under supervised or unsupervised settings [19
]. Compared to ISODATA and ML, RF also has an advantage in providing the relative importance of the input variables in predicting the response variable by permuting the predictor variable value and measuring the error estimate before and after the permutation [26
Unlike the aforementioned unsupervised and supervised pixel-based techniques (e.g., ISODATA and ML), the object-based image analysis (OBIA) approach considers contextual spatial information such as shape, smoothness, and compactness of geographical features of interest at different spatial scales [30
]. However, OBIA workflow for image classification involves an iterative trial-and-error image segmentation and optimization step. This is followed by a bottom-up merging of image-objects with the spatial and spectral heterogeneity threshold of adjacent landscape objects constrained by a user-defined scale parameter, with a subsequent step of classifying the primitive image-objects at the object-level using training data [30
Numerous studies have used both pixel- and object-based image classification techniques with RF for various natural resource management applications, including wetlands. For instance, Husson et al. [32
] used 5-cm spatial resolution true-color unmanned aircraft systems data for mapping non-submerged aquatic vegetation, classifying water (vs. vegetation), growth form, and dominant taxon using OBIA and RF classifiers, with overall accuracy results obtained for RF ranging from 62–90% for the growth form type to 52–75% for the dominant taxon classifications. Mahdianpari et al. [33
] used synthetic aperture radar (SAR) data in a hierarchical object-based RF approach to discriminate eight herbaceous wetland cover types in the Canadian province of Newfoundland with an overall accuracy of 94% achieved. Dronova et al. [31
] used the 32-m Beijing-1 satellite data and fuzzy supervised classification methods and an OBIA technique to detect changes of the major wetland cover types (i.e., water, mudflat, vegetation, and sand) of Poyang Lake, the largest freshwater lake–wetland system in China, with comparatively higher overall accuracy achieved for vegetation and water (90% and 82%, respectively). Ariel et al. [34
] applied a set of spatial and spectral image-object metrics to classify water and four vegetation types using RF with an overall accuracy of 92%. Tian et al. [35
] found higher overall accuracy using the OBIA coupled with the RF classifier (overall accuracy 93%) when compared to support vector machine and artificial neural network approaches in classifying nine land cover types, including wetlands.
Pixel- and object-based classification of landscape components is frequently further improved through the inclusion of spatial and spectral metrics in the algorithms. That is, in addition to the direct use of the spectral bands of the chosen sensor, numerical band combinations and band ratios may provide additional information [31
]. For instance, the well-known Normalized Difference Vegetation Index (NDVI [36
]) can be used as a proxy variable for indicating the presence and condition of vegetation (e.g., vigor, health, and abundance). This index can vary by vegetation types and habitats, thus providing useful information for improving land cover classification. Other vegetation indices have improved classification of remotely sensed data; their use is frequently dependent on site-specific conditions. For instance, to minimize the atmospheric aerosol scattering effect, the atmospherically resistant vegetation index (ARVI) [37
] has been derived. Similarly, soil brightness can affect vegetation indexes, and this can be compensated by using the soil-adjusted vegetation index (SAVI [38
]). In another example, to simplify or reduce computation time and computer processing power requirements, the infrared percentage vegetation index (IPVI) has been used to replace the NDVI [39
Furthermore, auxiliary input variables such as digital elevation model (DEM) and spatial metrics such as derivatives of the Grey Level Co-occurrence Matrix (GLCM) are also extensively used in various studies for improving land cover classification and prediction accuracies [28
]. Topographic position, through its effect on hydrological processes [41
] such as the prediction of areas of soil saturation in low-lying areas [31
], can influence the distribution of wetland classes in a landscape. Rodriguez-Galiano et al. [40
] have used elevation, slope, and aspect variables derived from a digital terrain model along with Landsat 5 TM spectral data as input to a RF model to classify 14 land cover categories in Spain with overall accuracy of 92% [40
]. Wright and Gallant [42
] have found DEM-derived topographic variables to be relevant in improving wetland mapping by increasing their accuracy in better identifying and differentiating upland areas from wetlands.
These varied random forest and object-based classification approaches and different metrics are useful in assessing and classifying landscape components. However, the methods discussed above also suffer from limitations. It is difficult and resource-intensive to collect sufficiently large amounts of field data for training an object-based (RF) model. Despite widely reported improved performance of object-based over pixel-based image classification approaches, Dronova [31
] indicated results vary by data type, spatial scale, and research objectives when applied in complex wetland systems (e.g., large wetlands with varied wetland vegetation structure and open water).
Wetlands in particular can be challenging landscape elements to classify due to their ecotonal location at the terrestrial–aquatic interface and complex hydroperiod and hydro-patterning which control the vegetation structures found therein. However, their importance in contributing to landscape hydrological, biogeochemical, and habitat functions and massive wetland losses worldwide [2
] make assessing the location and structure of wetland systems a critical research need. With the beguiling and varied approaches to classifying wetland systems, in this study we sought to determine if pixel-based or objected-based approaches performed most satisfactorily in an analysis and classification of the lower Barguzin River Valley, a large wetland study area draining into Lake Baikal, a United Nations Educational, Scientific, and Cultural (“UNESCO”) World Heritage Site located in Siberian Russia. We furthermore compared the efficacy of parametric (ISODATA, ML) and non-parametric (RF) approaches and iteratively analyzed outputs with spatial and spectral classification metrics seeking a parsimonious and effective wetland classification solution. Thus, in addition to accurately classifying the wetland landscape to improve management options, we aimed to provide an assessment and recommendation of methodological approaches for consideration specific to classifying wetland landscapes.
In our literature review, we did not find wetland classification studies assessing and applying highly numerous structural or habitat classes. For instance, an extensive review of OBIA for wetland mapping [31
] indicated few classified wetland systems into more than 10 or 11 classes (e.g., open water, emergent marsh, submergent vegetation, etc.). Our review suggested that overall accuracy in wetland studies frequently fell below an arbitrary benchmark of ~85% when the total number of wetland-specific classes exceeded four.
Thus, in our field-based analyses of a large wetland system in Siberian Russia with nearly 20 different structural and vegetative habitat and wetland classes, we sought analyze the differences in overall accuracy when comparing between three classification methods (pixel-based ML and RF, and object-based RF), constraining our analyses to use the same input field and remote-sensing datasets. The outcome of this study, therefore, assists end-users in selecting (and parameterizing) the proper classifier to analyze the structure of the world’s imperiled and complex wetlands.