Status and Trends of Wetland Studies in Canada Using Remote Sensing Technology with a Focus on Wetland Classiﬁcation: A Bibliographic Analysis

: A large portion of Canada is covered by wetlands; mapping and monitoring them is of great importance for various applications. In this regard, Remote Sensing (RS) technology has been widely employed for wetland studies in Canada over the past 45 years. This study evaluates meta-data to investigate the status and trends of wetland studies in Canada using RS technology by reviewing the scientiﬁc papers published between 1976 and the end of 2020 (300 papers in total). Initially, a meta-analysis was conducted to analyze the status of RS-based wetland studies in terms of the wetland classiﬁcation systems, methods, classes, RS data usage, publication details (e.g., authors, keywords, citations, and publications time), geographic information, and level of classiﬁcation accuracies. The deep systematic review of 128 peer-reviewed articles illustrated the rising trend in using multi-source RS datasets along with advanced machine learning algorithms for wetland mapping in Canada. It was also observed that most of the studies were implemented over the province of Ontario. Pixel-based supervised classiﬁers were the most popular wetland classiﬁcation algorithms. This review summarizes different RS systems and methodologies for wetland mapping in Canada to outline how RS has been utilized for the generation of wetland inventories. The results of this review paper provide the current state-of-the-art methods and datasets for wetland studies in Canada and will provide direction for future wetland mapping research.


Introduction
Wetlands are ecosystems where terrestrial and aquatic regions meet and share some characteristics. Wetlands also contain water for some periods of a year and are characterized by the presence of water, hydric soil, and specific vegetation adapted to a wet environment [1,2]. Wetlands are invaluable natural resources that provide exceptional benefits to humans and the surrounding environment [2]. Due to numerous environmental services of wetlands, including carbon sequestration [3], water purification [4], sediment filtration [5], soil conservation [6], and other critical services, wetlands have been called the "kidneys" of nature [7]. Additionally, from an economic perspective, wetlands are important due to their extensive applications for recreational activities [8], fish and shellfish aquacultures [9], flood mitigation [10], and providing diverse wildlife habitat [11,12]. Despite their numerous benefits, wetlands have been threatened by climate Swamps exist in both mineral and peatland wetland types [80]. Swam found in contact with other hydrological systems; hence they are difficult to of the distinguishing characteristics of swamps is that the woody vegetati shrubs) dominates the swamp environment (30% up to 100%) [82]. Addition soil in swamps is composed of well-decomposed wooded species rather tha Swamps exist in both mineral and peatland wetland types [80]. Swamps are often found in contact with other hydrological systems; hence they are difficult to identify. One of the distinguishing characteristics of swamps is that the woody vegetation (trees and Remote Sens. 2021, 13, 4025 5 of 43 shrubs) dominates the swamp environment (30% up to 100%) [82]. Additionally, the peat soil in swamps is composed of well-decomposed wooded species rather than the organic sphagnum or sedge-dominated peat in bogs and fens. The water table fluctuates in swamps, and they are not permanently saturated or inundated like bogs and fens; thus, the soil layer can be well-aerated [80].
Marsh is a type of mineral wetland class that experiences a high temporally periodical (seasonal/annual) rate of inundation. The hydrology inputs are from numerous sources, such as tides, water flow, groundwater, surface runoff, precipitation, and flooding. The variety of dissolved mineral inputs and freshening ventilation lead to high productivity and a diverse range of vegetation species [80]. Marsh vegetation communities are often comprised of emergent aquatic types, such as rushes, sedges, grasses, broad-leaved emergent, floating, and submerged aquatic plants [83].
Shallow water is a semi-permanent to permanent water body with a water depth of fewer than 2 m during mid-summer [39]. However, mudflats might be exposed in occasions of water drawdown. Submerged aquatic and floating vegetation with the capability of adaptation to constant inundation are present in shallow waters [82]. Figure 2 shows the workflow of preparing the documents for content analyses in this review. The bibliographic database of the present review was attained through performing a title, abstract, and keywords systematic literature search of relevant articles in the two well-known scientific databases of Clarivate Analytics Web of Science and Elsevier Scopus. To this end, all the combinations of search words (see Table 1) were applied to select English language journal papers, conference papers, and review papers between 1975 and the end of 2020. For instance, the combination of "wetland", "Canada", and "remote sensing" was the first combination for the literature search. It is worth mentioning that "Canada" was separately considered to include those research papers conducted over the entire country of Canada, as well as papers in which the province name was not stated in the title, abstract, or keywords. Afterwards, the Preferred Reporting Item for Systematic Review and Meta-Analysis (PRISMA) checklist was applied to organize the collected documents [84]. First, all the results for each combination were separately examined, and then the duplicate documents were removed. Subsequently, the remaining corpus of documents was combined to generate a consolidated database that encompassed 686 papers. A title filtering was then performed to identify the duplicate documents obtained by different combinations, which led to 473 documents. Later, those documents that their source file was not found, along with irrelevant documents, were also excluded, which finally resulted in 300 remaining papers. Following this filtering step, the title, abstract, and keywords sections of the remaining papers were screened to distinguish papers associated with wetland mapping from other wetland studies. This was performed because further attributes (see Figure 2) were derived from wetland mapping-related papers as the primary focus of this review. Finally, all 300 papers were fully inspected to extract different attributes (see Table 2), such as year, study area, classification method, and data type, for further analyses (see Table A2).

Results and Discussion
Several statistical analyses were first conducted in the following subsections based on the procedure defined in the method section. In addition to demonstrating the general characteristics of 300 RS-based wetland studies in Canada (e.g., publication details, geographical information, and RS datasets), a comprehensive survey and discussion of the meta-analysis status and trends were provided to present a comprehensive overview of 128 mapping studies. Policymakers can gain advantages from this overview in wetland mapping over Canada using RS technology.  Figure 3 also includes those journals that have published more than one paper in each time interval. It is worth noting that for the period 2006-2020, those journals that have published more than three papers are only provided. According to Figure 3, several clear-up conclusions can be drawn and summarized as follows. Over time, the number of published papers increased. As such, the distribution of articles shows a major positive trend in publications of wetland studies in Canada. A total of 9 (3%), 14 (4.7%), 10 (3.4%), 37 (12.4%), 43 (14.4%), 62 (20.7%), and 124 (41.5%) papers were, respectively, published in 1976-1985, 1986-1995, 1996-2000, 2001-2005, 2006-2010, 2011-2015, and 2016-2020. These results show that the published articles gradually increased about 50% in the period 1976-2020. After evaluating the time-level publication rates, we examined the number of publications for each year according to the study area. To this end, 300 articles were divided into 12 categories based on the Canadian provinces and territories, including BC, QC, SK, NU, MB, YT, NS, NL, AB, NT, NB, and ON. Figure 4 summarizes yearly trends in Canada's wetland publications according to the study area. Based on the results, there were no studies published from 1983 to 1987. It must be kept in mind that in this period, articles were presented in printed mode. Although many of them have been scanned into searchable formats and made available online, there may have been some other articles that were not scanned. Additionally, our extensive search of online resources indicates no studies published before 1987, and a small number of papers were published in early 2000 as well as in 2004. Almost 75% of the total 300 papers were published after 2004. The year 2020, with a total of 15 papers, had the most published articles since 1976. Moreover, the years 2019, 2018, and 2017, with a total of 24 published articles, were the second years with the most papers published about RS-based wetland studies in Canada.  After 2000, a wide range of studies has been conducted in different provinces of Canada so that the study on the YT and NB were started in 2011 and 2016, respectively. As such, the wetlands of all 12 Canadian provinces/territories were considered as the study area. After 2017, in most years, a large number of studies were developed in NL, ON, and AB (see Figure 4). We found that 22% (23 out of 103 articles), 18% (19 out of 103 articles),  Figure 5 a illustrates the word cloud generated from the keywords' frequencies. The size of each keyword is related to the frequency that a keyword has been used in all 300 papers. Considering the combination for the literature search, the biggest keywords were "Wetland" and "Remote Sensing". Since the reviewed papers came from different journals with various formats, the keywords were not consistent. Therefore, we preprocessed the words before feeding them into the word-cloud generator. For this purpose, all plural keywords were converted to their singular form. Lower-and upper-case words were justified, and all the first letters were capitalized. For instance, "Remote sensing", "landsat", and "wetland" were changed to "Remote Sensing", "Landsat", and "Wetland", respectively. With this substitution, the word cloud generator algorithm considered such words the same (e.g., "landsat", "Landsat", "LandSat", and "LANDSAT" were considered as one keyword of Landsat). The acronyms and their expanded versions were justified; then, acronyms were used in the word cloud. Finally, due to the similar meaning of some words, such as UAV and Unmanned Aerial System (UAS), they were merged and one of them was used.

Journal and Conference Analyses
In total, the 300 papers were published in 68 journals and 13 well-known international conferences. The journal publishers, as well as journals and conference papers, which published more than three times, are illustrated in Figure 6. The Canadian Journal of Remote Sensing and Remote Sensing (MDPI) with 46 and 40 papers were the top two journals, respectively. Moreover, Hydrological Processes, Remote Sensing of Environment, and International Journal of Remote Sensing with 19, 14, and 12 papers, respectively, were the other journals of interest in this field. In terms of the publisher center, most of the journal papers were published by Taylor & Francis followed by Wiley, MDPI, and Elsevier. Less than 8% of the journal papers were published by the SPIE, IEEE, and Springer. Among the conference papers, the IEEE IGARSS with 16 papers is the top conference for publishing papers on wetland studies in Canada, where the ISPRS Archive was the second one with 10 papers. To find out how many times a keyword has been used throughout the years, Figure 5b scatters the keywords per year. Each dot shows that the keyword has been mentioned in papers published in the corresponding year, and its size represents its frequency. Colors were selected arbitrarily for better visualization. The vertical axis representing the publication year was limited to 2000-2020 and the publications before 2000 were not displayed in this figure. As is clear, "wetland(s)" and "remote sensing" were the most frequently used keywords followed by "synthetic aperture radar (SAR)", "Landsat", and "RADARSAT-2."

Journal and Conference Analyses
In total, the 300 papers were published in 68 journals and 13 well-known international conferences. The journal publishers, as well as journals and conference papers, which

First and Co-Authors Analysis
This section summarizes the number of authors and co-authors in word-cloud, respectively. All the 300 papers were written by 210 unique first authors, and there were 943 co-authorships by 614 unique co-authors.

Affiliation Analysis
In Table 3, the top universities and institutions and their contribution are summarized. For this analysis, only institutions with three or more publications were considered. Memorial University of Newfoundland has the highest number of publications in wetland classification. However, multiple institutions from ON (e.g., Canada Centre for Remote Sensing and National Wildlife Research Centre) also have a significant contribution with a total number of 61 papers.

Citation Analysis
Citation analysis helps to ascertain prominent documents that significantly influence the corresponding field [85]. Furthermore, it also reflects the objectivity and quality of a paper by manifesting the number of attracted scholars to cite such a paper. Therefore, the citation number of all considered papers until the end of 2020 was extracted from Google Scholar to identify the high-contributing papers. It should be noted that earlier papers may have more citations than the recently published articles due to a more extended availability to the scientific community. Thus, the average citation per year was also calculated along with the total number of citations to reduce the effect of the elapsed time since publication. Table 4 presents the ten most cited papers devoted to wetland mapping in Canada. Based on Table 4, Ref. [27] was recognized as the most influential paper in the wetland studies conducted in Canada, in which the authors examined the applicability of various deep Convolutional Neural Networks (CNNs) for wetland mapping using high-resolution RS imagery. As mentioned, 128 out of the 300 papers were about wetland classification in Canada. These 128 papers were analyzed based on the number of wetland classes they included (see Figure 8). Almost all the papers (i.e., 114 papers) used five or fewer wetland classes. In total, 40 articles focused on five wetland classes (i.e., based on CWCS). Then, the second highest amount (29) belongs to papers covering one wetland class. The number of papers considering two, three, and four wetland classes were 14, 20, and 12, respectively. A few studies considered more than five classes. For example, four papers mapped six and seven classes, and two papers considered eight classes. There were only three papers discussing a large number of wetland classes, including 11, 12, and 17 classes.

Province-and Territories-Based Analysis
The percentage of the papers based on the number of mapped wetland classes in each Canadian province/territory are illustrated in Figure 9. Note that articles that covered large regions and nationwide study areas were not considered in this analysis.
Since almost 90 percent of the papers considered five or fewer wetland types, the classes in Figure 9 were decided to be from one to five, and others were considered as having six or more classes. Furthermore, an extra category of CWCS was also considered to depict the percentage of papers that followed the CWCS specifications. The NL province had the highest number of published papers (86.4%) based on CWCS specifications, followed by NS, BC, and YT (~50%). ON had the highest number of papers overall (36); however, none of them used CWCS. In addition, NB and SK were not studied in any CWCS-structured paper. Finally, the only paper studying wetlands in NU considered only one wetland class.  Figure 10 schematically illustrates a breakdown of RS-based wetland mapping studies in Canada by provinces/territories. This figure shows the spatial pattern of wetland mapping in Canada using RS data. Lighter and darker green hues indicate the lower and higher number of studies, respectively. The white hue depicts no study in the corresponding province/territory of Canada. It should be noted that some papers cover multiple study areas (i.e., multiple provinces, ecoregions, and entire Canada), and as a result, each corresponding province/territory was included in the count, separately. In Figure 10, those papers categories in Canada-wide studies contain all provinces. Based on a Figure 10, a large proportion of the studies were developed and assessed for only a few provinces, especially ON and NL. The literature search revealed that more than 40% of the individual case studies were focused on areas in ON and NL (darker green hues in Figure 10). Figure 10 also illustrates that NT with a total of 12 papers and AB and MB with 11 papers are other provinces where the wetlands were considered as the case studies. Very few published wetland research studies were identified from several provinces of Canada, including NU, YT, NB, NS, and BC. Overall, each of YT, NB, and NS accounted for two (less than 1.5% of published articles) studies.
As mentioned before, from all papers counted for all provinces in Figure 10, some papers contained multiple case studies and included several provinces. For example, ref. [92] expanded their study into more than one province, including MB and ON. On the other hand, ref. [54] applied their method on wetland mapping in AB and QC. As such, ref. [93] contained multiple case studies, including MB, NL, QC, and SK.

Geographical Distribution Based on the Extent of the Study Area
We also examined the number of publications according to the extent of the study area (see Table 5). As such, the 128 wetland classification studies were divided into five categories based on the extent of the study area: very small (less than 100 km 2 ), local (between 100 km 2 and 3000 km 2 ), regional (more than 3000 km 2 and less than a provincial scale), provincial, and national (Canada-wide) scales. Table 5. The number of publications focusing on various scales of the study area (small, local, regional, provincial, and national) in each Canadian province/territory.

Scale ON NL SK NT NS MB QC AB YT NU NB BC Canada Total Percentage
Very small 8 According to Table 5, the regional scale with a total of 50 papers, including 2 conference papers and 48 journal papers, had the highest number of published articles since 1976. The local and very small scales with a total of 36 and 32 publications were, respectively, the second and third scales. Each of the provincial and national study areas accounted for five (about 4% of published articles) articles. Table A1 and Figure 11 summarize the information about wetland classification methods used in Canada. Different types of unsupervised and supervised classifiers have been used for wetland mapping in Canada. In total, 16 classification methods were employed across the 128 Canadian wetland classification studies. The RF [94][95][96], ML [97,98], Decision Tree (DT) [38,[99][100][101][102], SVM [46][47][48], Multiple Classifier System (MCS) [11,103], Iterative Self-Organizing Data Analysis Technique (ISODATA) [104,105], CNN [21,27,54], k-Nearest Neighbors (k-NN) [106,107], and Artificial Neural Network (ANN) [30,[108][109][110] were the most commonly used algorithms. The Linear Discriminant Analysis (LDA) [83,111,112], Fuzzy Rule-Based Classification Systems (FRBCSs) [11,19], Markov Random Fields (MRF)-based method [113,114], k-means, and classification methods based on polarization target decomposition [115,116] were used once or less than three times and, here, were categorized as the "Other" group. and, here, were categorized as the "Other" group. Figure 11 shows that researchers have tended to use supervised methods (87%) in studies related to wetland classification in Canada rather than unsupervised approaches (13%). This is mainly because the unsupervised methods typically deal with the untagged data, which require further analysis for mapping classes, and they usually have lower accuracies than supervised methods. Moreover, the RF classifier (27.86%) was the most widely used algorithm, followed by ML (25.71%) and DT (10.34%) classifiers. The ANN (2.86%), k-NN (2.86%), CNN (3.57%), and MCS (3.57%) were rarely employed in the studies. SVM and ISODATA were also used in more than five studies. Finally, 11.43% of the studies used other classifiers for Canadian wetland mapping. The performance of the machine learning algorithms depends on several factors, including the complexity of the study area, type of RS data, quality of training samples, input features, classification algorithm, and tuning parameter settings [2]. Several metrics like overall accuracy, Kappa coefficient, producer's accuracy, and user's accuracy are typically used for classification performance evaluation. The wetland classification review studies rarely reported a complete confusion matrix to express wetland map errors (omission and commission errors), whereas they commonly stated the overall accuracy. Accordingly, the overall accuracy is here considered as a metric for comparing the accuracy of wetland mapping from different points of view.

Classification Methods
The boxplots of the overall accuracy obtained from different algorithms are displayed in Figure 12 to evaluate their performance in wetland mapping in Canada. As shown in Figure 12 all classifiers had more than 80% median overall accuracy, except the "Other" group with the lowest median overall accuracy by 76%. Among them, RF (88%), CNN  Figure 11 shows that researchers have tended to use supervised methods (87%) in studies related to wetland classification in Canada rather than unsupervised approaches (13%). This is mainly because the unsupervised methods typically deal with the untagged data, which require further analysis for mapping classes, and they usually have lower accuracies than supervised methods. Moreover, the RF classifier (27.86%) was the most widely used algorithm, followed by ML (25.71%) and DT (10.34%) classifiers. The ANN (2.86%), k-NN (2.86%), CNN (3.57%), and MCS (3.57%) were rarely employed in the studies. SVM and ISODATA were also used in more than five studies. Finally, 11.43% of the studies used other classifiers for Canadian wetland mapping.
The performance of the machine learning algorithms depends on several factors, including the complexity of the study area, type of RS data, quality of training samples, input features, classification algorithm, and tuning parameter settings [2]. Several metrics like overall accuracy, Kappa coefficient, producer's accuracy, and user's accuracy are typically used for classification performance evaluation. The wetland classification review studies rarely reported a complete confusion matrix to express wetland map errors (omission and commission errors), whereas they commonly stated the overall accuracy. Accordingly, the overall accuracy is here considered as a metric for comparing the accuracy of wetland mapping from different points of view.
The boxplots of the overall accuracy obtained from different algorithms are displayed in Figure 12 to evaluate their performance in wetland mapping in Canada. As shown in Figure 12 all classifiers had more than 80% median overall accuracy, except the "Other" group with the lowest median overall accuracy by 76%. Among them, RF (88%), CNN (86.6%), and MCS (85.75%) had higher median overall accuracies than the others. As expected, the "Other" group had the greatest range of overall accuracy results this group included dissimilar classification methods with different performances. ML, SVM, k-NN, DT, NN, and ISODATA with the median overall accuracies between 83% and 85% were the mid-range classifiers. The best (97.67%) and worst (62.40%) overall accuracies were achieved by RF [117] and Other [118] classifiers, respectively. There are different wetland classification strategies. For instance, analysis of pixel information (i.e., pixel-based methods) has been emphasized in some studies. However, recent studies have frequently argued the higher potential of object-based methods for accurate wetland mapping [2]. The pixel-based methods utilize the spectral information of individual image pixels for classification [2,119]. In contrast, homogeneous information (e.g., geometrical or textural information) in images is considered through object-based methods [17,119]. The pixel-based classification methods were preferred to the object-based approaches in most of the wetland classification studies of Canada. This could be mainly due to the simplicity and comprehensibility of the pixel-based methods compared to object-based approaches. However, our investigations showed that object-based methods had been extensively utilized in recent wetland mapping studies [7,68,73,103,120] due to their higher performance than pixel-based methods. The highest median overall accuracy (87.2%) was achieved by the object-based methods indicating their higher potential in generating accurate wetland maps in Canada. Finally, the pixel-based methods involved a wider range of overall accuracies and had the lowest overall accuracy.

RS Data Used in Wetland Studies of Canada
RS datasets with diverse characteristics (e.g., different spatial, spectral, temporal, and radiometric resolutions) have been widely used for wetland mapping in Canada. In situ data and aerial imagery were the main data resources for wetland mapping in Canada before advancing spaceborne RS systems in the last four decades. Spaceborne RS systems provide a wide variety of datasets with different sensors and, these are great resources for wetland studies at different scales. Additionally, much of the spaceborne RS data is free [121], leading to high utilization in wetland studies. Moreover, with the advent of UAV technology in recent years, images with very high spatial and temporal resolutions have been provided for wetland studies. In general, with the availability of RS datasets acquired by diverse spaceborne/airborne sensors, researchers have more options to produce highly accurate wetland maps. For example, multi-spectral passive optical satellite/aerial images have been frequently employed for wetland studies due to their straightforward interpretation and rich spectral information. However, such datasets are susceptible to clouds, resulting in their inefficiency in the cloudy regions [2,121]. Moreover, due to their short wavelength, optical signals cannot penetrate into the vegetation canopy [18]. In contrast, SAR signals are less affected by climate conditions (e.g., clouds and rain) [2,121,122]. SAR signals also have a high capability to penetrate into vegetation canopies, making them more beneficial than optical sensors to obtain information about wetland characteristics like structure, surface roughness, and moisture content [2,18]. Furthermore, modern SAR missions (e.g., RADARSAT-2, RADARSAT Constellation Mission (RCM)) acquire data in any combination of linear (horizontal and vertical) or circular (right or left) polarizations, which are very helpful for mapping treed and herbaceous wetlands [18,123].
Many wetland studies have combined optical and SAR data to achieve more accurate results. Additionally, a combination of optical, SAR, and elevation data has been extensively used for wetland studies in Canada (see Figure 13) and has usually provided the highest classification accuracies. As shown in Figure 13, single optical data (95 studies) is the most common data for wetland studies in Canada. Moreover, SAR data (57 studies) or dual combinations of SAR and optical data (53 studies) were often used. Single elevation data type (22 studies) was mostly employed to produce different topographic features, which can be accommodated for 3D analysis of wetland species and wetland mapping. Dual combinations of optical and elevation data (19 studies), and triple combination of optical, SAR, and elevation data (24 studies) were moderately considered as input data for wetland studies in Canada. The combination of elevation data with SAR data were the least utilized data types (only six studies). A total of 12 studies employed other data types, such as data derived from satellite telemetry, radiometers, satellite transmitters and ground penetrating radar for wetland studies in Canada.
The studies typically conducted on RS data acquired by different platforms, such as airborne, spaceborne or a combination of them. Most of the studies (~67%) were based on the spaceborne RS systems. This is probably due to the high capability and cost-effectiveness of spaceborne RS datasets for wetland mapping and monitoring over large areas in Canada. The airborne RS datasets were used in 13% of studies, where its combination with spaceborne RS datasets has been utilized in 20% of wetland studies. Recently, the use of Unmanned Aerial Vehicles (UAVs) equipped with RS sensors has become popular in wetland studies. In fact, the provided drone datasets could be a paradigm shift as they can be easily customized according to wetland studies specifications in contrast to spaceborne and piloted airborne RS datasets. Figure 14 provides the frequently used optical and SAR sensors in wetland studies in Canada. Landsat, Sentinel-2, and RapidEye were the most common medium resolution spaceborne optical systems, while IKONOS and WorldView-2 were the most widely used high-resolution spaceborne optical sensors in wetland studies in Canada. Among them, Landsat 4/5 images were often employed in studies due to their affordable spatial/temporal resolutions and rich archive datasets. Moreover, ERS-1 and -2, Sentinel 1, and RADARSAT-1 and -2 are the most popular C-band SAR systems, where ALOS-1 and -2 and TRASAR-X were widely used L-band and X-band SAR system in RS-based wetland studies in Canada. RADARSAT-2 images were frequently employed for wetland studies among SAR sensors because it is a Canadian SAR system, and it provides full/dualpolarization data with suitable azimuth and slant range resolutions. Finally, Compact Airborne Spectrographic Imager (CASI) hyperspectral system was the most popular sensor among airborne sensors.  For a closer look, the overall accuracy reported in wetland classification studies for various data types is shown in Figure 15a. Based on Figure 15a, the median overall accuracy of the various data types and their combinations is more than 80%. LiDAR/DEM data obtained the highest median overall accuracy (92%), resulted from only 3 papers out of 22 LiDAR/DEM papers that reported accuracy. The lowest median overall accuracy (82.4%) was achieved based on the SAR data type. However, a combination of SAR by another data type (e.g., optical or DEM) resulted in a better median overall accuracy. The median overall accuracy obtained by the optical data improved by combining with LiDAR/DEM data. Given the large number of studies conducted based on optical data, a wide range of overall accuracy (between 62.40% and 96.17%) was observed by this data, which was. Finally, the best overall accuracy (97.6%) was achieved by a triple combination of SAR, optical, and elevation data. Depending on the selected spatial resolution, wetland classification studies in Canada can also be categorized into three groups of high-resolution (<4 m), medium-resolution (4-30 m), and low-resolution (>30 m). Accordingly, the median overall accuracy achieved by reviewed papers using high, medium, and low spatial resolutions are illustrated in Figure 15b. The median of overall accuracy for all the spatial resolutions was more than 80%. The best median overall accuracy was achieved for studies that used mediumresolution datasets for wetland mapping, closely followed by the high-resolution datasets. Moreover, a great range of overall accuracies was reported in various studies using medium resolution images. As expected, the weakest results belonged to studies that used low spatial resolution data. The highest (97.67%) and lowest (62.40%) overall accuracies were obtained using a high-resolution and medium resolution data, respectively.
The results showed that 18 types of RS systems were used more than three times in 128 wetland classification studies, which are depicted in Figure 16. Airborne platforms, followed by RADARSAT-2 and Landsat 4-5, were the most frequently utilized sensors in Canada for wetland mapping using RS data. Among the Landsat series, Landsat 7 was less used, which was probably due to the failure of the Scan Line Corrector (SLC) on its board. Sentinel-1/2, Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Quickbird, ERS-1 and -2, and ALOS-2 were also among the sensors which were used in combination with other sensors. However, Quickbird, ASTER, GeoEye, and ERS-1 and -2 were the least common sensors with five or less uses.

Level of Classification Accuracy
For a comprehensive investigation of the RS-based Canadian wetland studies, the reported overall accuracies were assessed and compared with various parameters, including the year of publication, the extent of the study area, and the number of classes considered in the classification method (see Figure 17). Figure 17a presents the histogram of the overall classification accuracies reported in 128 papers. Note that a wide range of studies (39 papers) did not report the overall accuracy of their classification methods (black column in Figure 17a). According to Figure 17a, almost 80% (46 papers) of the studies have an overall accuracy between 84% and 93%; while only 33 papers have an overall accuracy of less than 84% (between 62% and 83%).
Based on Figure 17b, there is not a clear relationship between the overall classification accuracy and the year of publication. Two articles that were published in 1976-1995 have close overall accuracy to each other and the medium overall accuracy of 86%. Two articles that were published in 1996-2000 have achieved different accuracies. The medium overall accuracy of those articles is 71%. In another time-interval, there is a greater number of publications that have a wide range of overall accuracies between 63% and 96%.
Based on Figure 17c, wetland classification methods applied to the provincial scales have the highest median overall accuracies, followed by very small and local study areas. On the other hand, the papers on national scales have the lowest median overall accuracies. Based on Figure 17d, more than 90% of the investigated articles used a few classes (between two and six). In these papers, the overall accuracies vary between 62% and 96%. However, the median overall accuracies of these papers are 87% for 1-3 classes and 86% for 4-6 classes. In the case of 7-9 classes, the total number of papers decreases to four papers. The median overall accuracy of these four papers is 89%. Moreover, those articles that considered a greater number of classes have higher median overall accuracies. We also found two papers that considered 10-18 classes for classifying wetlands and achieved the median overall accuracies of 94%. As seen, a higher number of classes seem to be more accurate for the wetland classification method. We expect higher accuracies for a lower number of classes. Therefore, due to the significant discrepancy in the number of papers, it is impossible to provide a solid conclusion about the relationship between the overall accuracy of classification method and the number of classes.

Conclusions
This review paper demonstrated the trends of RS-based wetlands studies in Canada by investigating 300 articles published from 1976 to 2020. In total, twelve subfields were summarized, including classification methods and their overall accuracies, RS datasets, journals, number of wetland classes, authors/co-authors contributions and affiliations, publications per year, geographical distributions, scale of the study areas, citation, and keywords. Eventually, a deeper meta-analysis was carried out to discuss the utilization of RS systems in these subfields over Canada particularly, which differentiates our survey from previous reviews. Consequently, this paper addresses the status of wetland studies in Canada using RS data and highlights opportunities and limitations for generating and updating Canadian wetland inventories, as well as classification protocols improvements. In summary, the meta-analysis of 300 wetland studies, 128 of which were related to wetland classification, presented the following outcomes: Pixel-based and supervised classification methods were the most popular techniques to map wetlands in Canada due to the simplicity and higher accuracies of these strategies compared to the object-based and unsupervised approaches, respectively. However, the median accuracy of object-based methods was more than pixel-based techniques and, therefore, they have been more frequently used in recent studies.

•
Optical imagery and the combinations of optical and SAR datasets have been the most commonly used RS datasets to map wetlands in Canada. Availability, fulfilled archive, the high capability, and cost-effectiveness of optical and SAR imageries have attracted numerous focuses to utilize them. LiDAR/DEM data also resulted in the highest classification accuracies over small regions.

•
Most (but not all) of the reviewed studies did not present the full confusion matrix and only reported the overall accuracy to evaluate the results which were easily affected by the stratification of samples between dry and wet classes. Additionally, accuracy statistics often depend on the different factors, such as the geographic extent of the study area, type of RS data, the degrees of wetland species, the quality of training and tests samples, and classification algorithm and its tuning parameter settings. Therefore, it would be required to increase the number of wetland studies that try to actually quantify wetland classification errors in different aspects.

•
Approximately 30% of the studies considered the five CWCS wetland classes, and around 54% provides wetland maps using a lower number of classes. • Frequencies of "SAR" and "RADARSAT (1/2)" displayed the importance of SAR data for wetland mapping in Canada because of the capability of SAR data to acquire images in any weather conditions considering the dominant cloudy and snowy climate of Canada.
This review paper highlights the efficiency of RS technology for accurate and continuous mapping of wetlands in Canada. The results can effectively help in selecting the optimum RS data and method for future wetland studies in Canada. In summary, implementation an object-based RF method along with a combination of optical and SAR images can be the optimum workflow to achieve a reasonable accuracy for wetland mapping at various scales in Canada.

Data Availability Statement:
The data presented in this study can be available on request from the author. Table A1. Characteristics of the mostly used classifiers for wetland classification in Canada using RS data.

ISODATA
It is a modified version of k-means clustering in which k is allowed to range over an interval. It includes the merging and splitting of clusters during the iterative process.

ML
It is a parametric algorithm based on Bayesian theory, assuming data of each class follow the normal distribution. Accordingly, a pixel with the maximum probability is assigned to the corresponding class.

k-NN
It is a non-parametric algorithm that classifies a pixel by a variety vote of its neighbors, with the pixel being allocated to the class most common among its k nearest neighbors.

SVM
It is a type of non-parametric algorithm that defines a hyperplane/set of hyperplanes in feature spaces used for maximizing the distance between training samples of classes space and classify other pixels.

DT
It is a non-parametric algorithm belonging to the category of classification and regression trees (CART). It employs a tree structure model of decisions for assigning a label to each pixel.

RF
It is an improved version of DT, which includes an ensemble of decision trees, in which each tree is formed by a subset of training samples with replacements.

ANN
It is a multi-stage classifier that typically includes the neurons arranged in the input, hidden, and output layers. It is able to learn a non-linear/linear function approximator for the classification scheme.

CNN
It is a class of multilayered neural networks/deep neural networks, with a remarkable architecture to detect and classify complex features in an image.

MCS
It advantages from performances of dissimilar classifiers on a specific LULC to achieve accurate classification of the image. Supervised Table A2. List of 300 studies and main characteristics.