A wetland permanence classification tool to support prairie wetland conservation and policy implementation

Wetland permanence, the duration and frequency that surface water is present, affects biological communities and whether wetlands are protected under legislation in some jurisdictions. Wetland drainage in the Prairie Pothole Region (PPR) has changed the distribution of wetlands because smaller and more temporary wetlands are more likely to be drained. This change in distribution affects biodiversity and other wetland ecosystem services. In Manitoba, Canada, wetlands are treated differently under the Water Rights Act based on permanence classification and can either be drained with a simplified registration (temporary and ephemeral wetlands), drained with a permit requiring mitigation (seasonal wetlands), or are protected from drainage (semipermanent and permanent wetlands). To facilitate implementing a conservation program targeting the most vulnerable wetlands, we built a classification model using LiDAR and Sentinel‐2 data (1312 training observations). Our random forest model had 73% accuracy on 563 test observations and is applicable across the agricultural region of southwestern Manitoba. We predicted the wetland permanence class of 365,499 wetlands and built an online tool to help practitioners implement a conservation program that pays producers to conserve temporary and ephemeral wetlands. Our approach is applicable elsewhere in the PPR and other regions with variation in wetland permanence.

standing water persists above the surface and it affects many aspects of ecosystem functioning and associated ecosystem service values (Marble, 1991). Our definition of wetland permanence is restricted to effects of surface water in a typical year and does not refer to whether an area transitions from a wetland to upland. Prairie wetlands are typically fed by spring snowmelt and by precipitation (LaBaugh et al., 1998). They range from ephemeral basins that hold water only after snowmelt or major precipitation events to permanent waterbodies that have standing water from spring melt until ice forms (Euliss et al., 1999;Stewart & Kantrud, 1971). Stewart and Kantrud (1971) developed a classification system with five main wetland permanence classes for natural (i.e., undrained), non-fluvial prairie wetlands based on the distribution and composition of wetland vegetation communities ( Figure 1). The classification system has additional wetland classes for fens, alkali ponds, and tilled ponds that we did not include in our study. The foundation of their classification system is that vegetation communities reflect long-term wetland hydroperiod, soil characteristics, and water chemistry and thus are good indicators of wetland permanence. Permanence affects both the biogeochemistry and biological communities of the ecosystem. For example, permanence influences vegetation community composition and structure, invertebrate community composition, and habitat suitability for amphibians and waterbirds (Brandolin & Blendinger, 2016;Euliss et al., 1999;Pearl et al., 2005;Stewart & Kantrud, 1972).
The variety of wetland permanence classes within a landscape supports the region's biodiversity, with many organisms using different wetland types, depending on the species, life stage, season, or hydrological (drought) conditions (Elliott et al., 2020;Gabrielsen et al., 2022;Murkin et al., 1997;Snodgrass et al., 2000). Wetland permanence class is somewhat related to size, and small basins have been disproportionately drained (mean drained wetland area = 0.2 ha; Watmough & Schmoll, 2007). The disproportionate draining of small and temporary wetlands has resulted in a homogenization of wetland sizes (Van Meter & Basu, 2015). Thus, wetland drainage threatens biodiversity through the direct loss of habitat area, but also via the homogenization of wetland sizes and permanence classes.
Legislative protection for prairie wetlands varies greatly among jurisdictions. Historically, prairie wetlands have been afforded little legislative protection compared with other types of aquatic ecosystems owing to their isolation from navigable waters and their natural fishless state (Marton et al., 2015). However, growing recognition of the hydrological and ecological importance of these systems has led to policy changes in both the USA and Canada. In the USA, prairie wetland protection under the Clean Water Act was entrenched in 2015 when the Clean Water Rule regulation was published (Alexander, 2015). Since then, that rule has been rescinded and reinstated by subsequent administrations. The Clean Water Rule recognizes prairie pothole wetlands for their role in protecting drinking water and is built on a mitigation hierarchy to avoid loss, minimize impacts, and restore. The Clean Water Rule does not offer blanket protection of wetlands but has a goal of no-net-loss of wetlands. Unlike in the USA, wetland policies in Canada are generally the responsibility of provincial governments.
In the Canadian province of Manitoba, the Water Rights Act provides a licensing system to manage water, including wetlands, for use by agriculture, industry, and municipalities (Government of Manitoba, 2019). In October 2019, the province of Manitoba introduced a Regulation under the Water Rights Act, which simplifies F I G U R E 1 Wetland permanence classes for prairie pothole wetlands according to the Stewart and Kantrud (1971) wetland classification system and the drainage rules associated with each class within the Manitoba Water Rights Act and associated Regulation. guidance for drainage and water retention works. The Regulation's guidance on prairie wetlands follows a mitigation hierarchy with a goal of no-net-loss in wetland benefits and the treatment of wetlands depends on their permanence class using the Stewart and Kantrud (1971) classification system. The Manitoba Water Rights Regulation simplifies drainage permitting for ephemeral and temporary wetlands (Class I and II), allows drainage of seasonal wetlands (Class III) through permitting with an offset or fee, and prohibits draining of semipermanent (Class IV) and permanent (Class V) wetlands.
In 2019 and 2020, the Manitoba government invested $CAD 50 million in a trust to support the conservation of temporary and ephemeral wetlands through the Wetlands "Growing Outcomes in Watersheds" (GROW) Trust in recognition that temporary and ephemeral wetlands are susceptible to drainage and are unprotected by regulation. The fund was designed for program delivery by watershed districts, to ensure that programs matched local and provincial priorities. Four Manitoba watershed districts obtained funding to provide incentive payments to landowners to keep ephemeral and temporary wetlands intact for 10 years. However, there is no widely available data on the permanence class of wetlands in Manitoba and program delivery to conserve ephemeral and temporary wetlands is hindered by this lack of information. Further, delivery of the wetland retention program would be more effective with a tool that also identifies areas with a high density of eligible wetland basins for program enrollment. Thus, there was a need for an accurate and scalable tool to target watershed district activities in areas of high abundance of temporary wetlands, and to provide delivery staff with the necessary information to rapidly assess candidate lands for program eligibility.
There have been previous efforts to classify prairie wetland permanence or hydroperiod (Daniel et al., 2022;Montgomery et al., 2018), but none met the needs of the watershed districts. The approach developed by Montgomery et al. (2018) classified pixels within a wetland as belonging to different wetland permanence classes. However, the Manitoba Wetland GROW program enrolls entire wetlands and the imagery requirements are not viable for large geographic areas, so this approach did not meet the program's needs. Daniel et al. (2022) built a classification model to predict wetland permanence class of Alberta prairie potholes but reported high error rates in test data (48%-61%). Specifically, the watershed districts delivering the Manitoba Wetland GROW program required a tool to predict wetland permanence class and uncertainty for wetlands in southwestern Manitoba.
Our objective was to characterize wetland permanence class of prairie wetlands in southwestern Manitoba using remotely sensed data to support the implementation of programs conserving ephemeral (Class I) and temporary (Class II) wetlands ( Figure 2). We present a novel remote sensing method to characterize wetland vegetation and topography that, coupled with a classification model, can be used to accurately predict wetland permanence class. We worked with the Manitoba Association of Watersheds to package our approach in a user interface to identify wetlands most at risk of drainage based F I G U R E 2 Conceptual model of data collection, classification models, and identification of eligible wetlands to enroll in conservation programs.
on permanence class so that wetlands can be enrolled in conservation programs.

| Study area
Our study area was southwestern Manitoba, which is part of the PPR and contains a high density of depressional wetlands. The dominant land use is agriculture with a mix of annual cropland and pasture. Our study area is part of the Aspen Parklands Ecoregion, which is a mosaic of aspen forest, grasslands, and depressional wetlands.

| Canadian wetland inventory
We used Canadian wetland inventory (CWI) wetlands as the base layer for building our models and classifying wetland permanence class (Fournier et al., 2007). The CWI uses a standard definition of wetland following the Canadian Wetland Classification System (National Wetlands Working Group, 1997) and strives to delineate wetlands from the wet-meadow zone into the deepest portion of the basin. The Prairie Ecozone was targeted for high-resolution data in the CWI to adequately represent the numerous small wetland features that characterize the region. Wetland mapping in agricultural Manitoba follows the high-resolution approach with features captured from stereo pairs following a common interpretation guideline and quality assurance protocols (Canadian Wetland Inventory Technical Committee, 2016). All wetland data were collected to a standard minimum mapping unit of 0.2 ha and classified to the mandatory attribution of the CWI data model (Canadian Wetland Inventory Technical Committee, 2016). The CWI covers 33,563 km 2 (45%) of the PPR region of Manitoba, with mapping efforts historically targeting wetland-dense areas and using imagery from 2007 to 2009. CWI wetland polygons were dissolved to generate contiguous wetland basins. We excluded riverine and constructed wetlands because the Stewart and Kantrud (1971) system was not designed to classify those types of wetlands.

| Field data collection
We used field observations of wetland permanence class to train classifier models. Field staff sampled 2000 wetlands using eight roadside transects in August 2018 (n = 411 wetlands on two transects) and August 2019 (n = 1589 on six transects). August is an ideal time for field assessments because most species of dominant wetland vegetation can be readily identified. Field assessments during this time reflect vegetation responses to long-term average hydrological conditions. We used these observations to build classification models of wetland permanence class. Transects were chosen to capture a range of wetland permanence, wetland size, and surrounding agricultural land use (e.g., pasture and various annual crop types). Field staff assigned wetland class to all visible wetlands within 75 m of the road transect according to the vegetation community occupying the deepest portion of the basin, as in Stewart and Kantrud (1971), using common indicator species. Of the 2000 sampled wetlands, we used 1875 observations that were within the CWI and had complete data for further analyses.
Field staff from the Watershed Districts collected wetland permanence class data from an additional 331 wetlands in fall 2021. During field operations, staff assessed wetland permanence class using indicator vegetation species. Wetlands were chosen by staff opportunistically based on existing GROW referrals that were being assessed in the field. We used these observations as independent data to test the accuracy of our wetland permanence class model.

| Remote sensing process
The goal of the remote sensing process was to use a scalable approach to measure vegetation zone and elevation characteristics of wetlands that may predict wetland permanence classes. Sentinel-2 is a European wide-swath moderate-resolution, multispectral imaging mission comprised of sun-synchronous orbiting satellites with a revisit time of 5 days (Drusch et al., 2012). It has sensors with resolution ranging from 10 to 60 m, depending on the spectral band. We chose Sentinel-2 imagery to derive predictor variables of wetland permanence class because of its high availability of free content, moderate spatial resolution, and high spectral resolution. We selected spring and mid-summer imagery from 2017 to 2018, timed to capture spring melt and runoff, and mid-summer highvigor in-season wetland vegetation growth and to reduce the potential for temporary spring ponding to influence the classification. The field data collection (2018-2019) and Sentinel-2 imagery (2017-2018) did not completely overlap in time. However, the Stewart and Kantrud (1971) classification system assumes vegetation zones reflect long-term hydrological patterns. Further, if our model can predict wetland permanence class assessed in a different year as the remotely sensed data used in the model, then the approach is applicable in a wider range of scenarios where limited imagery data are available.
We downloaded Sentinel-2 imagery from the European space agency (ESA) portal as Level 1-C granules in standard archive format for Europe format. Each scene was processed with ESA's Sen2cor module to convert the bands from top of atmosphere to bottom of atmosphere reflectance. A 10 m resolution composite image was generated by combining the 10 m bands (B2, B3, B4, and B8) with resampled bands (B5, B6, B7, B8a, B11, and B12) in PCI Geomatica 2018. We generated principal components from the 10 m optical bands and added PCA1 and PCA2 to the composite. The cloud masks generated from Sen2cor were inspected and confirmed that no cloud obscured the study areas. We used light detection and ranging (LiDAR) to measure elevation and related physical characteristics of wetlands because elevation affects how much water a basin collects and its depth. We used 1 m Bare Earth LiDAR for the study area collected by the Province of Manitoba as part of the High Resolution Digital Elevation Model CanElevation Series (Natural Resources Canada, 2019) to generate terrain derivatives that were predictor variables for vegetation zones and wetland permanence class. All terrain derivatives were generated using Python scripts in ArcGIS 10.6.1 (ESRI, 2018). As a precursor, we mosaiced the 1 m tiles into seamless coverage of our study area and then resampled to 5 m resolution to facilitate efficient processing. We identified a depression mask by thresholding the stochastic probability of depression data (≥ 0.6) with a cut-fill surface that identifies depressions from a filled 5 m digital elevation model (DEM).
We used the Bare Earth LiDAR data to estimate contributing area (ha) of a wetland, depth-to-spill elevation for all 5 m cells within a wetland, and slope for all 5 m cells within a wetland. We used the Flow Direction function to determine DEM flow paths and the Watershed function to determine the upstream cells contributing to a wetland for the entire study area to estimate the contributing area of each wetland. We estimated a depth to spill elevation metric to characterize wetland depth for each wetland basin. The process assumes the lowest elevation along the polygon boundary represents the basin spill elevation of each wetland. For every 5 m cell within a wetland, we measured the difference between the wetland spill elevation and the cell elevation. For a wetland, we summarized the mean spill elevation as a metric of wetland surface water capacity. We generated slope metrics for all basin contributing areas with the intent of characterizing the upstream topography contributing to each wetland basin. We applied the Slope function to the 5 m DEM to generate a slope surface in degrees and used the maximum slope for the contributing area of each wetland as an input in our classification model.

| Vegetation zone classification
We built a classification model to identify vegetation zones by grouping similar pixels together because vegetation zone is a key input for assessing wetland permanence class. To classify vegetation zones, we used Digital Globe 50 cm RGBI imagery during the 2017-2018 growing season. We acquired high-resolution imagery from six 25 km 2 sites that overlapped road transects of wetlands in the Assiniboine West, Souris River, and Central Assiniboine watershed districts. We segmented the 4-band imagery in Geomatica Banff Object Analyst (scale 50, shape 0.5, compactness 0.5) to group similar pixel values into contiguous objects to assign labels. We labeled 2236 segments using aerial photo interpretation by vegetation zone according to Stewart and Kantrud (1971), including deep marsh, shallow marsh, open water, and wet meadow. We also included three upland vegetation categories: cropland, grassland, and woodland. We did not include the low prairie vegetation zone because CWI protocols focus on boundaries from the wet meadow to open water vegetation zones. The imagery area overlapped the field collection area but included vegetation segments outside of the field sampled wetlands.
We used 1274 training observations (58.4% of 2180 observations) of vegetation zones, the Sentinel-2 optical stack, and terrain derivatives based on LiDAR data to build a vegetation zone classifier. In ArcGIS Pro 2.9.2 (ArcGIS Pro. Environmental Systems Research Institute, Redlands, CA, 2021), we built a random forest classifier using Spatial Analyst's Random Tree Classifier with 150 trees, a maximum tree depth of 30, and a maximum of 1000 samples per class. We tested the accuracy of the vegetation random forest using 906 test observations (41.6% of observations). We used the classified surface within CWI wetlands to describe the proportion of a wetland with different vegetation zones. We grouped upland habitats together as the proportion of the wetland polygon that contained grassland, cropland, and woodland. The ability to distinguish vegetation communities in small basins or within narrow margins was constrained by the 10 m resolution of Sentinel-2 imagery.

| Wetland permanence classification analyses
We built classification models to predict wetland permanence class using remotely sensed predictor variables and field assessments of wetland class using 1875 wetlands with no missing data. We randomly split data into training (1312 wetlands; 70%) and test data (563 wetlands; 30%).
The Stewart and Kantrud (1971) classification system has seven classes, but the Manitoba Water Rights Act and associated Regulation treats wetlands in three groups (Class I/II, Class III, Class IV/V) differently. Our field data had 31 Class T (tilled) wetlands that we excluded and 0 Class VI (alkali pond) or VII (fen pond) wetlands. Thus, the maximum number of groups we tried to classify was five, corresponding to Class I (ephemeral) through Class V (permanent pond). We built classification models with two groups (Class I/II vs. Class III/IV/V), three groups (Class I/II vs Class III vs. Class IV/V), and five groups (Class I vs. Class II vs. Class III vs. Class IV vs. Class V).
The distribution of wetland classes in our training data was imbalanced and classification models are sensitive to class imbalances (Ganganwar, 2012). Of 1312 training data wetlands, 17 were Class I, 208 were Class II, 673 were Class III, 259 were Class IV, and 155 were Class V. In very imbalanced datasets, classification becomes biased toward the most common group (Ganganwar, 2012). To balance training data sets, we used synthetic minority oversampling techniques (SMOTE; Chawla et al., 2002), which generates new synthetic cases based on the nearest-K neighbor. We used SMOTE to increase the number of the minority case until the classes were approximately balanced within one factor of each other.
Using the training data sets balanced with SMOTE and 14 predictor variables, we fit random forests with the ran-domForest package (Liaw & Wiener, 2002) in R (R Core Team, 2021). The wetland permanence class was the response variable and the predictor variables were the vegetation (percent deep marsh, percent shallow marsh, percent wet meadow, percent open water, percent upland, the dominant vegetation zone, and the deepest vegetation zone), elevation (mean depth to spill elevation, wetland mean depth, the maximum slope, and contributing area), and wetland characteristics (surface area, ratio of contributing area to surface area, and ratio of perimeter to surface area). Random forests are an ensemble machine learning classification method that classifies cases with a series of decision trees that use random selections of cases and features (Breiman, 2001). Each tree gets a vote on classifying a case and the majority vote is the predicted classification. We used 500 trees and 10 variables to randomly sample for each tree split. We repeated the random forest procedure for our different classification levels (two-, three-, or five-groups of wetland permanence classes).
We assessed model accuracy using Cohen's Kappa (Cohen, 1960), the percent of test cases correctly predicted, and the percent of wetland visits in fall 2021 that correctly predicted wetland permanence class. Cohen's Kappa is a measure of agreement between true (i.e., observed) and predicted wetland class that takes into account the probability of agreement by chance (Cohen, 1960). A value of 0 indicates that model predictions are no better than what would be expected by chance, and a value of 1 indicates perfect agreement between predictions and observations. Cohen's Kappa <0.4 is generally considered poor model agreement (Landis & Koch, 1977). Producer accuracy is the percent of observed features that are correctly classified, whereas user accuracy refers to the percent of classified features that have been correctly identified (Congalton, 1991). We also considered feedback from Manitoba Watershed District staff, the needs of the Wetlands GROW program, and implementing the Water Rights Regulation when comparing models by valuing models that can separate more wetland permanence classes higher than models with fewer classes. To compare the contribution of individual variables to classification models, we examined the relative drop in accuracy (%) and in Gini Index from removing that variable. The Gini Index is a measure of class impurity in a sample (Breiman, 2001;Strobl et al., 2007) commonly used to build classification trees.
Our study included remotely sensed data (2017)(2018), field data (2018-2019), and additional field data for testing (2021) collected from years that varied in the amount of precipitation. During the study, the area underwent significant shifts in the Canada Drought Monitor index (Agriculture and Agri-Food Canada, 2023). In August of each study year, our study area varied from abnormally dry to severe drought during remote sensing data collection, normal conditions to severe drought during field data collection, and severe drought to exceptional drought during the collection of 2021 test field data. Thus, our training and testing data were collected across a wide range of hydrological conditions.

| Wetland permanence classification user tool
We used our wetland permanence classification model to build a web application to assist wetland conservation program delivery by identifying areas with a high density of predicted ephemeral and temporary (Class I and Class II) wetlands and by providing wetland-specific predictions of wetland permanence class. First, we scaled our wetland permanence classification model to all CWI wetlands in our study area (n = 365,499). Second, we used kernel density estimators to measure the variation in the density of ephemeral and temporary wetlands in our study area. We packaged the results of our predictions in an online application for Watershed District staff delivering wetland conservation programs targeting ephemeral and temporary wetlands.

| Field data
Along the transects, field teams classified 1875 basins according to the wetland permanence classification system of Stewart and Kantrud (1971). Class I (n = 23) and II (n = 302) wetlands were relatively rare, collectively making up 17.3% of total basins and 5.3% of total wetland area (Table 1). Seasonal (Class III) wetlands were the most numerous (50.4% of basins), and semipermanent (Class IV) and permanent (Class V) wetlands made up 32.2% of basins, collectively.

| Vegetation zone classification
The random forest model for vegetation zone had an overall accuracy of 74.7%. Producer accuracy was highest for the shallow marsh zone (85.8%) and user accuracy was highest for the open water zone (91.3%; Table 2). The model underpredicted deep marsh zones (producer accuracy 58.4%), which were confused with shallow marsh zones (78.7% of misclassified deep marsh zones). We did not include the low prairie vegetation zone because it could not be readily distinguished in imagery and because the boundary was outside of the mapped wetland boundary according to CWI protocols.

| Wetland permanence classification analyses
The three-group random forest model had the highest Kappa (0.56) and the two-group random forest model had the highest accuracy predicting test data (85% correctly classified; Figure 3). We chose the three-group random forest model to scale to the study area because of the utility of separating seasonal (Class III) wetlands for the Manitoba Water Rights Regulation, the relatively high accuracy (73% on test data), and ability to T A B L E 1 Distribution of 1875 wetland basins among the five Stewart and Kantrud wetland permanence classes in southwestern Manitoba, Canada.

Class
Basins ( discriminate groups (Kappa = 0.56). The confusion matrix indicated that predicting ephemeral and temporary wetlands had the lowest accuracy (59% user accuracy, 65% producer accuracy; Table 3). The predictor variables that had the largest effects on accuracy and Gini-index were the proportion of upland vegetation (cropland, woodland, and grassland) within the wetland, the proportion of open water within the wetland, the proportion of deep marsh within the wetland, the size of the wetland, and the proportion of shallow marsh within the wetland (Figure 4). The terrain metrics, including maximum slope, mean depth to spill, and mean wetland depth were less influential on the wetland permanence classification.
As an additional test of the model's validity, we looked at the prediction accuracy with field data collected in fall 2021. The fall 2021 field data had a similar distribution of wetland permanence classes as the broader field data set with seasonal wetlands being the most frequent class (n = 84), and fewer ephemeral (n = 51), temporary (n = 41), semipermanent (n = 28), and permanent (n = 29) wetlands. We excluded 1 alkali wetland (Class VI) and 97 tilled wetlands (Class T) from our accuracy assessment because those wetland classes are outside the scope of our classification tool. The threegroup random forest model had an overall accuracy of 63.5% on 233 wetlands, similar to the accuracy on the original test data collected along road transects. With this independent test data set, our model was able to accurately predict temporary and ephemeral wetlands (user accuracy of class = 63.0%, producer accuracy of class = 80.6%).
F I G U R E 3 Model comparison of accuracy on test observations (%; n = 563 wetlands) and Cohen's Kappa for wetland permanence classification models with two-group, three-group, and five-groups using random forests on observations of wetlands in southwestern Manitoba, Canada.
T A B L E 3 Confusion matrix for the three-group random forest model predicting wetland permanence class using 563 testing data wetlands from road transects in southwestern Manitoba.

| Wetland permanence classification user tool
Our model predicted that 139,556 of 365,499 wetlands (38%) are ephemeral or temporary and potentially eligible for the Wetland GROW program. Our online application included spatial predictions of temporary and ephemeral wetlands. Density of ephemeral and temporary wetlands were particularly high in the northern and western parts of the Rural Municipality of Two Borders and the southern part of the Rural Municipality of Oakview ( Figure 5).

| DISCUSSION
Our wetland permanence classification model contributes to wetland conservation program implementation, including the Wetland GROW program, because it provides predictions of wetland permanence class and identifies hot spots of ephemeral and temporary (Class I and Class II) wetlands in southwestern Manitoba. No-net-loss wetland policies, such as the Manitoba Water Rights Act, that rely on wetland permanence classifications require scalable tools that reduce field costs for assessing wetlands that are unlikely to qualify for specific programs. Hydrologic permanence affects wetland evaluation and qualification under other legislation, including the U.S. Clean Water Act (Hough & Robertson, 2009 wetlands is important for targeting conservation of those habitats. The most important variables for predicting wetland permanence class were proportion of upland vegetation, proportion of open water, proportion of deep marsh, wetland size, and proportion of shallow marsh. It was unsurprising that wetland vegetation community zones, including open water, deep marsh, and shallow marsh were strong predictors of wetland permanence class because these zones are the basis of the Stewart and Kantrud wetland classification system. However, it was surprising that upland vegetation communities of grassland, cropland, and woodland, had such a strong effect on wetland permanence class. All three vegetation communities were rare (median values of wetland basin area for each individual vegetation community = 0%), but together upland vegetation communities differed between Class I and II wetlands (47 ± 1.9%), Class III wetlands (23.1 ± 0.8%), and Class IV and V wetlands (9.0 ± 0.6%). We posit the association between upland vegetation and wetland permanence is driven by small wetlands being more likely to support upland vegetation because of only intermittent flooding and because smaller wetlands are more likely to be affected by surrounding land use.
Our wetland permanence classification model performed well but had lower accuracy predicting the target group of ephemeral and temporary wetlands (user accuracy: 59.0%, producer accuracy: 64.8%) due to several limiting factors. First, less permanent wetland basins are smaller on average, so they have fewer pixels to estimate the proportion of different wetland vegetation communities. This both increases the number of wetland pixels misclassified as upland vegetation and increases uncertainty because estimates may be based on fewer pixels (e.g., 20 pixels in a wetland that was the minimal mapping unit of 0.2 ha). Further, we used the boundary of wetlands determined by stereoscope interpretation, as per the CWI protocols, but this may limit inclusion of some vegetation communities important for distinguishing wetland permanence based on vegetation, such as the low prairie zone. Annual variation in climatic conditions may have lowered the accuracy of our classification tool because remote sensing and field data were from different years. Despite these challenges, our classification model performed well on imbalanced test data from different years and is a useful tool for predicting wetland permanence class across a large geographic area.
Although our model was designed to meet the needs of the Manitoba Watershed Districts for delivery of conservation programs, model design and accuracy are suitable for numerous other applications, especially if it were scaled to broader geographies within the PPR. For example, waterbody permanence affects waterbody governance in policies in other Canadian jurisdictions (e.g., Alberta's Public Lands Act). Beyond policy applications, permanence affects the biogeochemistry and ecology of wetlands. Thus, information on permanence class is essential for numerous research and conservation applications. For example, permanence influences greenhouse gas emissions from prairie wetlands (Badiou et al., 2011;Bansal et al., 2016) and information on permanence is useful for accurate modeling of wetlands as natural climate solutions (sensu Drever et al., 2021). Wetland permanence also shapes wetland vegetation and invertebrate community composition (Euliss et al., 1999), and influences habitat suitability and use by numerous vertebrate species (Brandolin & Blendinger, 2016;Gabrielsen et al., 2022). At broader scales, the relative abundance of wetland permanence classes can influence duck vital rates (e.g., duckling survival; Howerter et al., 2014). Broader availability of wetland permanence data in the PPR would allow the scaling of existing models for conservation planning and refinement of models where lack of permanence information has inhibited incorporation of known habitat preferences (e.g., waterbird abundance and occupancy models used by the Prairie Habitat Joint Venture; Prairie Habitat Joint Venture, 2021).

| CONCLUSION
Classifying wetland permanence class is challenging because of geographic variation in wetland form, precipitation and temperature patterns, high annual variation in hydroperiod, and limited data availability. Our remote sensing and classification approach can predict wetland permanence class across southwestern Manitoba with acceptable accuracy to help conservation practitioners identify and enroll wetlands in conservation programs. Our approach of measuring vegetation zones and elevation traits affecting wetland permanence can be applied to other parts of the PPR with wetland inventory and LiDAR data. In addition, predicting wetland permanence class will be helpful for linking biodiversity patterns to wetland characteristics, identifying drainage patterns based on hydroperiod, and measuring geographic differences in the dominant wetland permanence class.
AUTHOR CONTRIBUTIONS Lauren E. Bortolotti and L. Boychuk designed the study and secured funds. L. Boychuk led the remote sensing analyses. Lauren E. Bortolotti and James E. Paterson built the classification models. James E. Paterson created the figures. James E. Paterson, Lauren E. Bortolotti, and L. Boychuk wrote the manuscript and contributed editorial input.