AI-Driven Wetland Mapping Across Diverse Natural Regions of Alberta, Canada, Using Combined Airborne and Satellite Remote Sensing Data

Highlights

What are the main findings?

• AI models integrating multi-sensor satellite data and LiDAR met or exceeded Alberta’s provincial wetland mapping standards across four pilot regions.
• Deep learning achieved the highest overall accuracies, while machine learning captured finer details and more effectively detected rare wetland types.

What are the implications of the main findings?

• AI-driven approaches offer a scalable and efficient pathway to modernize wetland inventory mapping and support operational updates across Alberta.
• Future provincial mapping initiatives should carefully consider the cost–benefit of integrating different technologies, including LiDAR and high-resolution satellite imagery, which offer high impact and scalability but can be costly to acquire.

Abstract

This study evaluates the performance of artificial intelligence (AI) technologies for wetland classification in the province of Alberta, Canada, using integrated remote sensing inputs, including airborne light detection and ranging (LiDAR), orthophotography, and multi-sensor satellite imagery (Sentinel-1, Sentinel-2, PlanetScope). Our primary objective was to assess whether AI-driven modelling approaches, specifically machine learning (ML) and deep learning (DL), can meet Alberta’s provincial wetland mapping standards. We hypothesized that integrating high-resolution LiDAR with multi-seasonal optical and radar data composites into advanced AI algorithms would achieve the required classification accuracy, detail, and minimum mapping unit targets. We tested several methodologies in four ecologically distinct pilot areas representing Alberta’s Boreal, Grassland, and Parkland Natural Regions. AI models included ensemble ML using Extreme Gradient Boosting (XGBoost) and Random Forest, and a DL U-Net convolutional neural network (CNN). AI models were trained on expert-labelled photoplots and validated using in situ field surveys. Our findings demonstrate that both ML and DL models met and, in several cases, exceeded the provincial mapping standards with validation overall accuracies surpassing >70% (form), >80% (class), and >90% (wetland–upland). U-Net CNN models generally produced the highest overall accuracies and most precise wetland extent delineation, but XGBoost offered finer detail and granularity for detailed mapping of rare wetland forms. Integrating LiDAR data and derivatives further enhanced model performance, improving accuracy by as much as 13%. Based on these outcomes, we provide a set of recommendations for scaling up these approaches, focusing on model selection, LiDAR imagery integration, and the continued value of field surveys to support the operational scaling of AI-driven classification approaches for wetland inventory updates across Alberta’s diverse landscapes. However, key challenges remain in scaling up this approach due to the cost of acquiring high-resolution LiDAR and satellite imagery.

1. Introduction

Wetlands are among the most productive ecosystems globally, providing critical ecological services including water quality regulation, flood control, nutrient retention, and essential habitat for wildlife [1]. However, wetlands are experiencing rapid and widespread decline, with an estimated global loss of approximately 3.4 million km² since 1970, a trend that has been particularly pronounced in Europe and North America [2]. This loss affects a wide range of wetland systems, including swamps, marshes, and open water bodies, which play key hydrological roles and support migratory birds and other wildlife, many of which are species at risk [3]. Key drivers of wetland degradation include land use change, industrial development, and climatic warming, including increasing wildfire frequency and severity [4]. The scale and pace of wetland loss underscore the urgent need for improved monitoring, management, and policy interventions, all of which rely on accurate, spatially explicit, and up-to-date wetland inventories that remain unavailable in many regions [5,6]. In Alberta, Canada, wetlands cover over 28% of the land surface [7], with many designated as Ramsar sites—wetlands designated as important under international environmental treaty [8]. Despite efforts to raise awareness of their benefits [9], the loss of wetlands continues at an alarming rate [10]. For example, it is estimated that 90% of wetland area has been lost in the Prairie Pothole Region of Alberta, primarily due to drainage for agricultural expansion [11]. This loss prompted the development of Alberta’s Wetland Policy [12], which prioritizes avoidance as the primary mitigation strategy when development activities may impact wetlands [13,14]. To meet the needs of this policy, wetland data of high quality and consistency are necessary [15]. Rapidly advancing technologies, such as digital remote sensing (RS), are providing encouraging avenues for meeting these information needs while ensuring adherence to established standards [16]. In Alberta, these standards include strict requirements for wetland classification detail, accuracy, and feature size. Alberta’s wetland mapping standards were established to support the development of a provincial-scale wetland inventory, while considering what was achievable using Earth Observation (EO) data.

Wetlands have traditionally been mapped using a combination of aerial photography, field-based reconnaissance, and manual image interpretation [17]. However, recent advances in low-cost RS data and EO satellite systems have enabled more cost-effective, data-driven, and repeatable wetland mapping at broad scales [18]. Using different regions of the electromagnetic spectrum, RS sensors capture information on water, vegetation, and soil properties, with optical imagery from platforms such as Landsat and Sentinel-2 widely used for wetland mapping and change analysis [19,20,21,22], though performance can be limited in cloud-prone and forested environments such as the boreal biome [23]. In contrast, active synthetic aperture radar (SAR) systems penetrate vegetation, operate independently of solar illumination, and are highly sensitive to wetland dielectric properties [24]. Increasingly, the integration of optical and SAR data has been shown to improve wetland classification performance, particularly in boreal regions [25,26,27], and when complemented by topographic RS data such as light detection and ranging (LiDAR), which provides critical information on elevation, hydrology, and vegetation structure relevant to wetland presence and type [28,29,30,31,32,33].

More recently, wetland mapping has greatly benefited from data science and artificial intelligence (AI)—i.e., machine learning (ML) and deep learning (DL)—technologies [34,35,36,37]. Classic ML algorithms such as Random Forest (RF) and support vector machine (SVM) have been widely applied for wetland classification [38,39,40,41]. More sophisticated AI modelling techniques, such as DL convolutional neural networks (CNNs), have gained traction with usage and novel developments, with varying but encouraging results [42,43]. Prominent examples have demonstrated the value of DL CNNs for wetland classification on both medium-resolution [44,45,46] and high-resolution [47,48,49,50] RS data. The success of DL models, such as CNNs, is owed to their ability to extract hierarchical features (i.e., from simple features in early layers, to more complex and abstract features in deeper layers) from high-dimensional RS imagery for complex landscape classification [51]. Moreover, CNNs are effective at capturing spatial context and landscape patterns, which manifest in wetlands in the form of gradients and transitional zones [52].

The primary aim of this study is to evaluate innovative AI-driven mapping methodologies that have the potential to meet or exceed Alberta’s provincial mapping standards and support repeatable operational inventory development, and updates [16]. Further motivation comes from the limitations of Alberta’s current wetland inventory, the Alberta Merged Wetland Inventory (AMWI; [7]), which is derived from various EO sources, from a range of dates, and completed using different methodologies, resulting in variations in quality and accuracy, and is no longer considered up to date. To our knowledge, this is the first direct evaluation of new RS and data science technologies with regard to the recently adopted provincial standards, which are now central to wetland inventory mapping and updates in Alberta. To achieve this, we piloted state-of-the-art technologies specific to Alberta’s ecological zones, in four areas across the province. More specifically, our goals were to

• Test AI-based methodologies, including ML and DL, and new EO data including high-resolution airborne LiDAR and photos and commercial and publicly available satellite imagery.
• Develop wetland inventories for each pilot area and evaluate them against the provincial mapping standards.
• Provide recommendations on the advantages of new methodologies and their potential for scaling across Alberta to update the provincial wetland inventory.

We expect this study to inform the development of scaled-up, accurate, efficient, and cost-effective wetland mapping and monitoring in Alberta and beyond, to similar landscapes with expansive boreal wetlands, complex parkland areas and, prairie pothole wetland regions.

2. Materials and Methods

2.1. Study Areas

Four pilot areas within the Boreal, Foothills, Parkland, and Grassland Natural Regions were selected for this study to develop effective wetland mapping approaches across Alberta, where a key challenge is to effectively map a wide diversity of wetland types and extent, and that are also situated in a diversity of landscape settings (Figure 1). Alberta’s Natural Regions are characterized by their unique climate, physiographic, vegetation, soil, wildlife, and land use attributes [53]. The four pilot areas were selected within Natural Regions for which the Government of Alberta (GOA) have set wetland mapping standards: Boreal/Foothills and Prairie/Parkland zones. These pilots capture the broad diversity of Alberta’s wetlands, ranging from peatland-dominated systems in the northern forested boreal where wetlands cover large portions of the landscape to the warmer Prairie–Parkland regions in the south where wetlands are smaller, more dynamic, and typically dominated by mineral soil marshes and open water (

Table 1. Pilot area characteristics.

Pilot Name	Natural Region	Area (km2)	Description
Boreal-1	Boreal	7613	Peatland-rich boreal landscape with mixedwood forests, diverse wetlands, and moderate forestry and energy footprint.
Boreal-2	Boreal and Foothills	21,214	Area with mixedwood forests, diverse wetlands, and extensive forest harvesting and energy development.
Parkland-1	Parkland	5170	Landscape with wetlands in morainal and glaciolacustrine settings, dominated by marshes and open water, interspersed with some swamp, fen, and woodland and agricultural lands.
Grassland-1	Grassland	5065	Dry mixed grass landscape with small, hydrologically isolated marshes and ponds, largely agricultural with scattered energy and transport features.

). All pilots are supported by extensive validation data and remotely sensed imagery.

2.2. Wetland Definitions and Target Standards

According to the GOA wetland mapping standards [16], and following local nomenclature, thematic classification level requirements are (1) form for the Boreal/Foothills zone (BFZ) and (2) class for the Prairie/Parkland zone (PPZ; Figure 2). Wetland class and form definitions are based on the Alberta Wetland Classification System ([55]; ESRD, 2015). There are five major wetland classes in the AWCS, which are complementary to the Canadian Wetland Classification System (CWCS; [56]), and 13 wetland forms. Wetland forms are subdivisions of the major wetland classes, based on dominant vegetation structure. The AWCS was developed specifically for wetlands in Alberta and includes a suite of key indicators used to classify wetlands.

In the BFZ, minimum achievable classification accuracies are 90% for separating wetlands from uplands, 80% for wetland class, and 70% for wetland form. However, the standards do not specify which accuracy metric is required, so we report several measures with a particular emphasis on overall accuracy. In the PPZ, these minimum accuracies are the same except for wetland form, which is optionally reported. Both PPZ pilots mapped wetland forms since wetland class diversity is very low in these areas, and also included an additional form: tilled marshes. Tilled marshes are wetlands that have been fully or partially drained and then cultivated for agriculture, and are prevalent in the PPZ [57]. The minimum mapping unit (MMU)—the size of the smallest delineated feature—is 0.9 ha for the BFZ and 0.04 ha for the PPZ [16].

2.3. Remote Sensing Data

2.3.1. Boreal/Foothills Satellite Data

Multi-temporal and multi-source satellite data were collected using Google Earth Engine (GEE; [58]) following methods from DeLancey et al. [44] and Merchant et al. [59]. Image sources included Sentinel-2, Sentinel-1, and ALOS World 3D 30 m (AW3D30; [60]). AW3D30 was acquired for comparison against aerial LiDAR as a source of topographical input data, since both moderate- and high-resolution digital elevation models (DEMs) have proven effective in wetland mapping [32]. All BFZ layers were harmonized at a final 10 m spatial resolution prior to AI modelling; datasets not at this resolution were resampled and aligned using bilinear interpolation.

Atmospherically corrected Sentinel-2 Level-2A surface reflectance data were used from years 2020 to 2022. “Early” and “Late” seasonal composites were created using June–July and August–September images. This was performed to capture the dynamic ecohydrological characteristics of boreal wetlands [61]. Images with a cloudy pixel percentage of less than 20% were selected and then cloud masked. Composites were created using the median value from the time-series stack, and for 10 m (i.e., blue (B), green (G), red (R), near-infrared (NIR)) and 20 m (i.e., 4 red edge (RE) and 2 short-wave infrared (SWIR)) bands [62]. Sentinel-2 derived spectral indices included enhanced vegetation index (EVI), normalized burn ratio (NBR), normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and soil adjusted vegetation index (SAVI; [63]).

Sentinel-1 imagery for the same years and seasons as Sentinel-2 was acquired in Level-1 Interferometric Wide (IW) Ground Range Detected (GRD) mode. Polarization channels included VV (vertical transmit, vertical receive) and VH (vertical transmit, horizontal receive). Preprocessing included thermal noise removal, radiometric calibration, terrain correction, and speckle filtering. Composites were created using median aggregation. Radar Vegetation Index (RVI), cross-polarization ratio (VV/VH), and Span variables (total backscatter power) were derived [64]. Open-source global seasonal SAR coherence data were also acquired from Kellndorfer et al. [65] to complement the backscatter data. Coherence measurements, produced using Interferometric SAR (InSAR) techniques, describe the correlation between two or more SAR images acquired for the same location but during different times. High coherence means minimal change, whereas low coherence means little correlation. Coherence InSAR measurements have proven valuable for wetland characterization [66].

AW3D30 preprocessing included resampling to match the Sentinel-1 and -2 spatial resolutions. Topographic indices were derived using SAGA-GIS and included topographic wetness index (TWI), topographic position index (TPI), topographic ruggedness index (TRI), and valley bottom flatness (VBF). A summary of each remote sensing data source used for BFZ modelling is provided in

Table A1. Summary of remote sensing datasets used for BFZ wetland modelling.

Platform	Date Type	Spatial Resolution	Acquisition Years	Seasonal Composites
Sentinel-2	Optical	10–20 m	2020–2022	Early (June–July) and Late (August–September)
Sentinel-1	SAR	10 m	2020–2022	Early (June–July) and Late (August–September)
LiDAR	Topographic	1 m	2021–2022	-
AW3D30	Topographic	30 m	2006–2011	-

 Appendix A.

2.3.2. Boreal/Foothills Airborne Data

High-resolution airborne LiDAR point cloud data stored in LiDAR Aerial Survey (LAS) format were collected from different providers to obtain coverage over the BFZ pilots. There were minor gaps over large lakes that were filled using the open-source AW3D30. Each LiDAR provider conducted quality assurance and control procedures to meet industry standards defined by the American Society for Photogrammetry and Remote Sensing (ASPRS) [67]. Additionally, ground classification (i.e., point filtering) was performed using established filtering approaches (Axellson 2000 TIN refinement algorithm; [68]) and codes (e.g., unclassified, ground, etc.) to further ensure compliance with ASPRS requirements and accuracies. An overview of BFZ LiDAR data specifications, including providers and point densities, can be found in

Table A2. Overview of BFZ LiDAR sources, coverage, and derivatives.

Pilot Name	Provider	Year	Point Density (Points/m2)	Coverage (km2)	Derivatives
Boreal-1	Tolko	2022	16	6583	DEM, CHM, intensity, TWI, TPI, TRI, VBF
	ABMI	2022	12	1050
Boreal-2	Alberta-Pacific	2022	12	9703
	West Fraser	2021	17–30	11,428

 Appendix A. A DEM, canopy height model (CHM), and intensity layers were created from the LiDAR point cloud using Python (version 3.11) and Point Data Abstraction Library (PDAL; version 2.6.3), and then resampled to 10 m using bilinear interpolation [69]. The remaining topographic features were created from the LiDAR-derived DEM using SAGA-GIS (version 8), and matched the AW3D30 derivatives (i.e., TWI, TPI, TRI, VBF). The choice in DEM derivatives was based on several preceding studies that have successfully leveraged topography for boreal wetland classification [70,71].

2.3.3. Prairie/Parkland Satellite Data

Multi-temporal and multi-resolution optical satellite data were collected from Sentinel-2 and PlanetScope. Both imagery sources were included based on prior research in Alberta demonstrating their combined benefits [72]. The 3 m PlanetScope data were imaged 29 April 2023, to target peak hydroperiod and vegetation senescence, and then on 30 July 2023, to capture lower water levels and high vegetation vigour. PlanetScope data were downloaded as Level 3B surface reflectance products, and then mosaicked with auto cutline and colour balancing parameters in PCI Catalyst using bands compatible with Sentinel-2. Sentinel-2 seasonal composites were generated using images queried in USGS Earth Explorer from May, and then from July to August. Surface reflectance data were normalized and resampled (bilinear interpolation) to 3 m and then mosaiced in PCI using similar parameters applied to PlanetScope. EVI2, NDVI, ratio vegetation index (RVI), aerosol-free vegetation index (AFRI), and red edge normalized vegetation index (RENDVI), were calculated.

Sentinel-1 SAR was not assessed in the PPZ pilots, as prior research found it to have little value for wetland mapping in southern Alberta due to the sensor’s lower noise floor [73]. Similarly, satellite-derived AW3D30 was not assessed because accurate depression modelling in prairie pothole landscapes requires finer spatial resolution topographic imagery [33,74]. A summary of each remote sensing data source used for PPZ modelling is provided in

Table A3. Summary of remote sensing datasets used for PPZ wetland modelling.

Platform	Date Type	Spatial Resolution	Acquisition Years	Seasonal Composites
Sentinel-2	Optical	10–20 m	2023	Early (May) and Late (July–August)
PlanetScope	Optical	3 m	2023	Early (April) and Late (July)
Orthoimagery	Optical	0.25 m	2023	Early (May) to Late (September)
LiDAR	Topographic	1 m	2022	-

 Appendix A.

2.3.4. Prairie/Parkland Airborne Data

Airborne LiDAR point cloud collection was conducted by the GOA across both PPZ pilots following GOA LiDAR acquisition guidelines [75], which meet ASPRS requirements and standards. LiDAR was collected in late fall to capture low abundance of surface water in wetlands, post-harvest conditions in cultivated landscapes, and “leaf off” conditions in woodland areas. This helped optimize bare earth products (processed using [68]), by improving the capacity to model the numerous hydrologically isolated depressions that characterize prairie landscapes. DEM, CHM, and several intensity layers were generated from the point cloud using ESRI ArcGIS Pro (version 3.3) utilities, and processed at 3 m resolution. Subsequent topographic variable extraction included TPI, slope, and several more advanced derivatives including LevelSet Depression Hierarchy or “LevelSet”, probability of depression (PDEP; [10]), depression depth from median elevation (DDME), and depression depth to spill elevation (DDSE; see

Table A4. Overview of PPZ LiDAR sources, coverage, and derivatives.

Pilot Name	Provider	Year	Point Density (Points/m2)	Coverage (km2)	Derivatives
Parkland-1	GOA	2022	6	5170	DEM, slope, CHM, intensity, TPI, LevelSet, PDEP, DDME, DDSE
Grassland-1	GOA	2022	6	5065

 Appendix A for PPZ LiDAR specifications). LevelSet was particularly studied in PPZ pilots, since it emulates water levels based on spill elevations and thus can effectively model wetland hydrology (Wu et al., 2019 [33]). DDME and DDSE were calculated using the following equations:

DDME = DTM − ME

(1)

DDSE = DTM − SE

(2)

where DTM is the LiDAR-derived digital terrain model, ME is the median elevation, and SE is the spill elevation which assumes the lowest point along the depression margin is the natural spill elevation.

Very high-resolution airborne orthophotography was also captured across PPZ pilots. Multiple missions were required to fully image each pilot due to Alberta’s extreme 2023 wildfire season [76]. Grassland-1 was flown between May and June 2023, and Parkland-1 between June and September 2023. Data were acquired in four-band infrared at 0.25 m spatial resolution, but resampled to 3 m using bilinear interpolation. Ortho-mosaics were generated in PCI Geomatica software (version 2018) using similar parameters applied to PlanetScope and Sentinel-2 mosaics.

2.4. Reference Data

Reference data used to train and validate AI models were acquired from several sources (

Table 2. Sources of training and validation for AI models.

Pilot	Model Training Data	Model Validation Data
Boreal-1 and Boreal-2	Photogrammetric plots	Ground and helicopter surveys
Grassland-1 and Parkland-1	Photogrammetric plots	Ground and drone surveys

), including both RS- and field-based (see

Table A5. Summary of field survey sites (polygons) collected for each pilot.

Class	Form	Boreal-1	Boreal-2	Parkland-1	Grassland-1
Water	Bare	3	2	56	13
	Aquatic vegetation	0	2	3	4
Marsh	Graminoid	39	38	332	101
	Tilled	0	0	0	57
Fen	Graminoid	4	63	0	0
	Shrubby	22	34	0	0
	Wooded coniferous	19	43	0	0
Bog	Graminoid	0	19	0	0
	Shrubby	0	7	0	0
	Wooded coniferous	32	75	0	0
Swamp	Shrubby	40	58	20	0
	Wooded coniferous	8	33	0	0
	Wooded deciduous	2	12	10	0
	Wooded mixedwood	7	17	0	0
Upland	Upland	22	73	45	27

 Appendix A for a summary count of BFZ and PPZ field survey sites, and Figure A1 Appendix A for a map of their spatial locations). Field protocols were developed to guide the consistent collection of new field data for model validation. With technical input and review by the GOA, the field protocols were carefully developed for the accurate classification of wetland class, form, and type levels according to the AWCS. Reference data in all pilots were quality assessed and controlled for consistent labelling and attribution. New ground-based sites also underwent a blind expert review by several wetland ecologists to resolve any labelling discrepancies.

2.4.1. Boreal/Foothills Field Data

BFZ field data were acquired using both ground-based and helicopter-based approaches. The latter, which included both historical and newly acquired sites, was necessary due to the inaccessibility and remoteness of boreal environments. Historical (~within 10 years) helicopter sites, which included documented photographs and vegetation attributes (species, coverages, and heights), were collected by Ducks Unlimited Canada (DUC) as part of earlier wetland inventory projects intersecting the BFZ pilots. All historical sites were quality-controlled by an experienced wetland ecologist. This involved adjusting polygon boundaries against contemporary imagery, assessing any disturbances, and assigning new labels according to the AWCS. AWCS labels were determined by cross-walking the DUC Wetland Classification System codes [77] originally assigned to the site, using vegetation and hydrology as key determinants.

The BFZ reference data repository was further enhanced with new helicopter and ground-based survey sites, collected in July and August of 2023. Both approaches used similar site selection processes, focusing on areas with high wetland density and diversity. Wetland coverage estimates, based on DUC’s historical inventory, were used to rank and select areas with the most wetlands. A random stratified sampling approach was applied, stratifying sites by natural subregion to capture wetland differences and by fire history to include both burned and unburned areas—this was to account for the effects of wildfire on spectral reflectance, which is known to complicate wetland classification in remotely sensed data [78]. Ground sites also needed to be within 200 m of a road to ensure field crew access success. Open water wetlands were not targeted for field surveying, since these types in the BFZ are easily identified in RS imagery. These samples were instead captured using photo interpretation techniques.

2.4.2. Boreal/Foothills Photo-Interpreted Data

Photogrammetrically produced plot data were acquired from the Alberta Biodiversity Monitoring Institute (ABMI) 3 × 7 km photoplot repository [79], for use solely in BFZ AI model training. Final accuracy reporting was based on field surveys alone. The ABMI photoplots are detailed and spatially explicit vector inventories characterizing moisture, management status, vegetation features, wetlands, land use, infrastructure, and land cover for approximately 5% of Alberta. However, these photoplots were created over a range of dates (roughly over ~10 years), by various analysts, and therefore required a comprehensive expert review and editing process before use in AI modelling. All photoplots intersecting the BFZ pilots were assessed in reference to the AWCS class and form definitions. Vegetation, soil, and hydrology attributes were used to correct known errors in the photoplot labels. This task was supported using high-resolution optical satellite imagery (e.g., WorldView, Pléiades). New polygons were digitized for small wetlands not originally captured in the photoplots, to account for the MMU of the provincial wetland mapping standards. Overall, this resulted in a collection of enhanced 3 × 7 km photoplots reflective of current conditions (see

Table A6. Summary of ABMI 3 × 7 km photoplot review and correction process, comparing the number of polygons prior to review, and after.

	Boreal-1		Boreal-2
Class	Original	After Review	Original	After Review
Water	99	64	491	347
Marsh	37	100	278	432
Fen	367	315	2622	2437
Bog	64	107	613	1149
Swamp	209	439	2423	2544
Upland	1534	1285	8424	7941

 Appendix A for a comparison of photoplot labels prior to editing and after review).

2.4.3. Prairie/Parkland Field Data

PPZ field data were collected using only ground-based access and approaches, as helicopters were not required for access in these areas. Dedicated field surveys were conducted in summer 2023, however validation data from the previous year (2022) collected for other inventory projects was also leveraged. In fall 2022, four land parcels were visited within each of the PPZ pilots. Parcels were selected based on wetland abundance, wetland complexity (e.g., interspersal of ephemeral basins), and wetland rarity (e.g., forested wetlands). Depressions were modelled in each parcel using a LiDAR DTM, which helped guide the field validation process of wetlands and photo collection. Time synchronous, ultra-resolution (3 cm) colour (RGB) imagery was collected over the parcels with a DJI Mavic Pro unmanned aerial vehicle (UAV) drone. Areas were then resurveyed in 2023 using a DJI Mavic 3 to compile additional near-infrared (NIR) and red edge (RE) band data. The UAV imagery was also used to produce DTM and CHM derivatives. Ground photos, field observations, and UAV imagery and derivatives were collectively used to map wetland polygons in the parcels by an experienced photo-interpreter. Wetlands were mapped to the AWCS class and form.

The 2023 field campaigns followed a similar random stratification site selection process described for the BFZ pilots. Quarter sections were randomly selected, filtered to ensure they were within only one natural subregion and within 200 m of a road feature, and then stratified by high wetland count and diversity estimates based on other available inventory sources [7,80]. Sampling points representing potential wetland basins (i.e., depressions) were generated within the selected quarter sections. Wetlands were visited and validated in June and July of 2023 (see Table A5 Appendix A for a summary on PPZ field collected survey sites).

2.4.4. Prairie/Parkland Photo-Interpreted Data

The PPZ pilots utilized a combination of existing photoplot labels (from 2021) alongside additional, new 5 × 5 km photoplots [81]. New photoplots were randomly selected in areas of high wetland abundance and density based on the Canadian Wetland Inventory (CWI). All photoplots were generated from airborne stereo imagery containing four spectral bands (R, G, B, NIR), with 30 cm to 50 cm spatial resolution, and across a range of years (2019–2023) within the summer growing season. All stereo models were quality-controlled prior to photoplot mapping. Photoplot mapping was conducted by an experienced vendor following an interpretation guideline developed and updated by DUC [81]. Photoplot production was conducted on a progressive on-going basis, with all plots reviewed for attribution and quality by DUC as completed. Any deficiencies were flagged and addressed prior to final delivery.

2.5. AI Modelling

Several AI models were evaluated in this study. A generalized workflow is presented in Figure 3; however, inputs and algorithms varied depending on zone and pilot site. The following sections provide a detailed overview of the BFZ and PPZ AI modelling used for wetland mapping.

2.5.1. Boreal/Foothills Machine Learning

Building on recent boreal ML developments [59,82], a pixel-based Extreme Gradient Boosting (XGBoost) algorithm was chosen for BFZ ML modelling [83]. XGBoost is an ensemble learning method that groups together several “weak learner” decision trees to create a “strong learner”. This is performed sequentially, continuously correcting the previous weak learner until a final strong learner is achieved. Three XGBoost scenarios were tested for BFZ mapping. The first scenario included only important optical and SAR features, while the second and third scenarios used important features plus AW3D30 or LiDAR features. Identification and selection of important SAR and optical features followed methods in Merchant and McBlane [84]. This involved using the Shapley additive explanations (SHAP) algorithm, an explainable AI (XAI) method based on cooperative game theory [85]. SHAP values quantify the contribution of variables to a model’s prediction, allowing variables to be ranked by importance. The 20 most important features identified by SHAP were chosen as BFZ XGBoost model inputs, balancing predictive power and parsimony.

XGBoost models were developed using the Classification and Regression Training (caret) package in R. Training samples were randomly created using the enhanced ABMI 3 × 7 km photoplots. A balanced sampling design was used as this has been shown to produce accurate wetland mapping results with ML [86]. Three thousand points per class or form were created with a minimum of 10 m distance between points. This sampling procedure was repeated four times for each BFZ pilot, resulting in four unique training sets. Separate XGBoost models were trained using each training set, resulting in four class and form predictions. Predictions were summarized into a final classification by taking the mode of each prediction. It should be noted that ML form models were trained to map vegetation community (i.e., graminoid, shrubby, wooded, etc.). Therefore, to make a final form prediction, the vegetation community ML prediction was intersected with the preceding ML class prediction (e.g., if the ML class prediction was bog, and the vegetation community prediction was shrubby, then the form prediction would be shrubby bog). All XGBoost models were optimized by tuning key hyperparameters (e.g., max depth, nrounds, eta, and subsample) which control model complexity and behaviour, using an exhaustive grid search method and k-fold cross validation [87] with a total of 5 folds used for tuning. Model predictions were generated at 10 m spatial resolution, which meets the target 0.9 ha MMU.

2.5.2. Boreal/Foothills Deep Learning

BFZ wetland mapping with DL followed similar CNN methods developed by DeLancey et al. [44]. CNN modelling was implemented in Python using the Keras deep learning library. The CNN was based on a U-Net architecture [88], which is an encoder–decoder that performs image segmentation by producing an output with similar dimensions as the input. Input training patches were of size 224 × 224 pixels. Two U-Net scenarios with different inputs were tested for boreal mapping. Each scenario used the same optical and SAR features, but included either LiDAR or AW3D30 features. Only a minor number of features were discarded in each scenario since DL models can automatically learn relevant hierarchical features [89].

During initial stages of CNN model development, several notable observations were made: that different groups of feature inputs did not affect results as much as training regime, that aggressive data augmentation was the biggest factor for improvement, and that training cycles were short. This suggested that the CNN was data-starved. To address this, additional pseudo-plots were generated by producing weak labels from a baseline U-Net trained only on the enhanced 3 × 7 km photoplots. Pseudo-labelling approaches have proven effective in remote sensing studies [90]. The weakly labelled pseudo-plots expanded the training set and helped the U-Net learn general spatial patterns before later being fine-tuned with the expert labelled 3 × 7 km photoplots.

All U-Net models were developed using data from both Boreal-1 and Boreal-2 pilots, rather than performing pilot-specific training. This was due to the limited amount of label data, and the requirements of DL. U-Net base models were first trained using all pseudo-plots and all but four photoplots (i.e., holdouts), which were used for monitoring training progress. Patches were sampled with a window stride of 45, augmented using flipping and 90° rotations, and normalized. Certain wetland classes or forms were oversampled to account for rarity (e.g., bogs and marshes). Base models were trained for 50 epochs (200 iterations per epoch) with a batch size of 16 and an initial learning rate of 0.001, which was lessened to 0.0005 at the 40th epoch. These base U-Net models were then fine-tuned using only patches from photoplots, and not pseudo-plots. For fine-tuning, a tighter window stride of 10 was used with a learning rate of 0.0002, which decayed by a factor of two every six epochs. These fine-tuned models reached their minimum loss at around 20 epochs (100 iterations per epoch). The best fine-tuned U-Nets were chosen according to their loss performance on the holdout photoplots.

A summary of BFZ AI modelling scenarios is found in

Table 3. BFZ AI modelling scenarios. RS inputs and tests were the same for Boreal-1 and Boreal-2 pilots.

Pilot	Scenario	Algorithm	Modelling Inputs
Boreal-1 and Boreal-2	B-ML-S1	XGBoost	XAI selected Sentinel-1 and -2 variables
	B-ML-S2	XGBoost	XAI selected Sentinel-1 and -2 variables, AW3D30
	B-ML-S3	XGBoost	XAI selected Sentinel-1 and -2 variables, LiDAR
	B-DL-S1	U-Net	Select Sentinel-1 and -2 variables, AW3D30
	B-DL-S2	U-Net	Select Sentinel-1 and -2 variables, LiDAR

. These scenarios were tested for both wetland class and form mapping.

2.5.3. Prairie/Parkland Machine Learning

The PPZ pilots implemented object-based ML modelling. This allowed us to compare an existing operational Random Forest (RF) object-based image analysis (OBIA) method that uses a set of optimized RS variables (unpublished by DUC) against a novel XGBoost OBIA method. All segmentation processing was performed using PCI Object Analyst. OBIA models evaluated segmentations derived from both high-resolution (3 m) PlanetScope and very-high-resolution (1 m) orthophotography (i.e., aerial imagery). Following segmentation, numerous ML scenarios with different RS inputs were tested. This included various combinations of predictors, including optimized terrain and optical variables identified in the conventional DUC workflow, optimized PlanetScope variables, Sentinel-2 variables, or LiDAR variables which included either point cloud variables, topographic variables, or with or without the use of LevelSet. The effect of LevelSet was a particular focus in this work since it is capable of effectively mapping wetland depressions [91].

XGBoost OBIA models also underwent several additional, enhanced ML development steps. First, a training data selection method was employed using Gradient Information Optimization (GIO), which refines training data by minimizing divergence from a reference “pure” set while preserving information content [92]. GIO uses gradient descent and k-mean clustering to accomplish this. GIO was implemented to reconcile PPZ training data collected from different years and seasons, which posed a unique mapping challenge since PPZ wetlands are dynamic and vary in extent over time [93]. To collect the “pure” dataset for GIO, 100 OBIA segments were randomly selected and stratified by form. All segments needed to have at least 80% overlap with validated (i.e., truth) photoplot labels, although upland segments required additional screening due to their high landcover variability. This was performed by vetting upland segments against a PlanetScope derived water/non-water mask, in which segments needed to have an intersection over union (IOU) greater than the calculated median IOU. Once the pure dataset was compiled, GIO was implemented using the following parameters: cluster size of k = 1500, learning rate of 0.03, and maximum iteration size of 4000 for each form.

After generating a refined training dataset with GIO, a test dataset was then created by selecting a subset of segments with high IOU values, and that were spatially disjointed (i.e., did not overlap) from the GIO training dataset. Using the train and test datasets, 10 spatially independent k-folds were created with 1 km buffers. Conjointly, 10 stratified, non-spatial k-folds were created using a train–test split of 0.9 to 0.1. Both k-folds were then merged on the fold number, and upsampled to account for class imbalance. For ML scenarios that used orthophotography, segments used for training and k-fold creation needed to have 90% overlap with PlanetScope GIO segments.

All PPZ XGBoost models were tuned using Bayesian optimization and Parzen Tree Estimators (PTEs; [94]). PTE was chosen because of the large number of PPZ samples to train with. Bayesian optimization keeps track of past results to form probability mappings of hyperparameter combinations, making it a very efficient approach. PTE hyperparameter tuning used the Hyplot Python library. All predictor variables were used to establish a baseline, tuned model prior to variable selection.

After PTE tuning, variable selection was performed using a combination of SHAP values and Genetic Algorithms (GAs). GAs excel at handling large datasets and at finding solutions to optimization problems with many possible outcomes [95]. However, they are computationally intensive since each solution needs to be evaluated. To address this challenge, we leveraged SHAP values to simulate the predictions of a model in an efficient way. This was on the basis that SHAP values are distributed across predictors, and sum to the value of prediction across all variables—meaning if you add up all the SHAP values for a given prediction, you obtain the same result as the model’s actual prediction [96]. To verify this phenomenon, an XGBoost model was built with Parkland-1 training data using all RS variables and 150 model configurations. The predictions from each model configuration were compared to the predictions generated by the SHAP values. Comparing the results of each method produced an R2 of 0.83, suggesting good approximation with this technique. This finding allowed us to use GAs to perform variable selection, by using predictions estimated from SHAP values without needing to compute each model. For each model, the fit function was the resulting F1-score along with an L2 regularization-inspired penalization function, which helped stabilize accuracy. Once each model underwent variable selection, another tuning was performed with Hypopt using only the original k-folds.

The last component of PPZ ML modelling involved correcting XGBoost models using active learning. This was because initial model runs were overpredicting tilled marshes. We used the modAI Python package to implement ranked batch-mode active learning (RBMAL). RBMAL iteratively retrains a model by ranking unlabeled samples based on a composite acquisition score that balances model uncertainty and sample diversity. An uncertainty score for each class is recorded, derived from XGBoost class posterior probabilities, along with a diversity score computed to reduce redundancy among selected samples. These are combined to create a single composite score, using a weighted linear scoring function, with equal weighting applied to uncertainty and diversity (0.5), and used to rank all unlabeled points. RBMAL was used to select points out of a pool of incorrectly classified tilled marsh segments. A starting XGBoost model was created with 500 trees, max depth of 5, and learning rate of 0.05. Model retraining was then repeated for 500 iterations, using 10 points added during each iteration. The augmented training set was used as a template for further PPZ XGBoost models.

2.5.4. Prairie/Parkland Deep Learning

PPZ wetland mapping with DL was also based on a U-Net CNN. CNN modelling followed a two-stage training process. The first stage developed a base model U-Net trained on all photoplot labels, for all time periods. The second stage fine-tuned this model using only labels from 2023. This approach was taken because the most recent 2023 labels had considerably more detail. Thus, fine-tuning the PPZ U-Net encouraged the finer detail to be captured. However, it should be noted that 2023 label and validation data did not contain fen samples. As such, the U-Net was not fine-tuned with this class.

The base model U-Net was trained for 30 epochs (300 iterations per epoch) with a learning rate of 0.002 that halved at epoch 25. Input patches were sampled with a window stride of 45, augmented using flipping and 90° rotations, and normalized. The fine-tuned U-Net was trained for 20 epochs (200 iterations per epoch) with a learning rate of 0.0001 which halved each eight epochs. Patches were samples with a tighter 22 pixel stride, augmented using flipping and arbitrary rotations, and normalized. Both the base model and fine-tuned model used input patch sizes of 256 × 256. During training, loss and metrics were monitored after each epoch on a holdout 5 km2 area of training labels. This holdout was used to choose the best U-Net model.

A summary of PPZ AI modelling scenarios is found in

Table 4. PPZ AI modelling scenarios. RS inputs and tests were the same for Parkland-1 and Grassland-1 pilots.

Pilot	Scenario	Algorithm	Segmentation Source	Modelling Inputs
Parkland-1 and Grassland-1	P-ML-S1	RF	PlanetScope (2023)	Optimized PlanetScope and Sentinel-2, LevelSet
	P-ML-S2	RF	PlanetScope (2023)	Optimized PlanetScope, LevelSet
	P-ML-S3	RF	PlanetScope (2023)	Optimized PlanetScope and Sentinel-2, all LiDAR
	P-ML-S4	RF	PlanetScope (2023)	Optimized PlanetScope, all LiDAR
	P-ML-S5	RF	Orthophotography	Orthophotography, Sentinel-2, LevelSet
	P-ML-S6	RF	Orthophotography	Orthophotography, Sentinel-2, all LiDAR
	P-ML-S7	XGBoost	PlanetScope (2023)	XAI selected satellite and LiDAR terrain variables
	P-ML-S8	XGBoost	PlanetScope (2023)	XAI selected satellite and all LiDAR variables
	P-ML-S9	XGBoost	Orthophotography	XAI selected satellite and LiDAR terrain variables
	P-ML-S10	XGBoost	Orthophotography	XAI selected satellite and all LiDAR variables
	P-DL-S1	U-Net	-	All PlanetScope, orthophotography, and LiDAR variables
	P-DL-S2	U-Net	-	Early season PlanetScope, and all orthophotography and LiDAR variables
	P-DL-S3	U-Net	-	PlanetScope, orthophotography, and LiDAR terrain variables
	P-DL-S4	U-Net	-	All LiDAR and orthophotography variables
	P-DL-S5	U-Net	-	PlanetScope and LiDAR terrain variables
	P-DL-S6	U-Net	-	Orthophotography and LiDAR terrain variables

. These scenarios were tested for wetland form mapping, but not class. This is because, unlike the BFZ, wetland diversity in the PPZ is very low. For example, in most cases, only one form existed per class in each pilot (e.g., graminoid marsh, shrubby swamp, bare open water). Therefore, it was more appropriate to model wetlands directly at the form level.

2.6. Accuracy Assessments

All AI models were evaluated using common mapping accuracy metrics [97]. This included overall map accuracy (OA), since this is the benchmark metric outlined by the GOA, as well per-class F1-score calculated from precision and recall:

Overall accuracy = Correctly classified pixels/total pixels

(3)

Precision = TP/TP + FP

(4)

Recall = TP/TP + FN

(5)

F1 = 2 × Precision × Recall/Precision + Recall

(6)

where TP, FP, and FN reflect the number of true positives, false positives, and false negatives, respectively.

2.7. Post Processing

A non-vegetated human footprint layer, pulled from the ABMI’s 2021 Human Footprint Inventory (HFI; [98]), was used to mask out disturbed areas as a post-processing step. This was performed for all pilots using the following HFI sublayers: reservoirs, borrow pits, non-permeable roads, rail lines, canals, mine sites, and industrial sites.

3. Results

3.1. Boreal/Foothills Modelling Results

At the wetland class detail, the benchmark OA target of 80% set by the GOA [16] was met in both BFZ pilots. In Boreal-1, a highest OA of 82.7% was achieved with B-DL-S1, and in Boreal-2, a highest OA of 87.8% with B-ML-S3 (Figure 4; see Table 3 for scenario descriptions). B-DL-S1 was the only AI model scenario to exceed 80% OA at the class detail in Boreal-1, whereas in Boreal-2 all AI model scenarios exceeded 80% OA. At the wetland form detail, the benchmark OA target of 70% was also met in both BFZ pilots. In Boreal-1, a highest OA of 74% was achieved with B-ML-S3, and in Boreal-2, a highest OA of 70.3% was achieved also with B-ML-S3 (Figure 5). B-DL-S2 was the other AI model that exceeded 70% OA in Boreal-1, whereas in Boreal-2 only B-ML-S3 met the target OA.

The per-class F1-scores achieved by the best BFZ AI models are found in

Table 5. BFZ per-class and per-form F1-scores of the best performing AI models.

	F1-Score			F1-Score
Class	Boreal-1 (B-DL-S1)	Boreal-2 (B-ML-S3)	Form	Boreal-1 (B-ML-S3)	Boreal-2 (B-ML-S3)
Water	0.98	0.98	Bare	0.95	0.78
			Aquatic vegetation	-	0.23
Marsh	0.64	0.71	Graminoid	0.65	0.71
Fen	0.82	0.68	Graminoid	0.26	0.42
			Shrubby	0.74	0.19
			Wooded coniferous	0.58	0.46
Bog	0.83	0.83	Graminoid	-	0.16
			Shrubby	-	0.33
			Wooded coniferous	0.90	0.82
Swamp	0.82	0.76	Shrubby	0.69	0.67
			Wooded coniferous	0.65	0.58
			Wooded deciduous	0.47	0.04
			Wooded mixedwood	-	0.19
Upland	0.89	0.96	Upland	0.87	0.96

. F1-scores are reported for both class and form details. F1-scores varied considerably between and within classes, and across BFZ pilots. At the class level, F1-scores varied from 0.64 to 0.98 in Boreal-1, and 0.68 to 0.98 in Boreal-2. Water and upland classes were the highest F1-scores in both BFZ pilots. At the form detail, F1-scores varied from 0.26 to 0.95 in Boreal-1, and 0.04 to 0.96 in Boreal-2. In general, forms that are more rare, such as graminoid bogs or wooded mixedwood swamps, had lower F1-scores.

The performances of the best models for each BFZ pilot were further evaluated at the more general wetland/upland detail. An OA of 98.1% was achieved in both pilots at this binary detail, using the B-DL-S1 AI model in Boreal-2 and B-ML-S3 AI model in Boreal-2. This meets the 90% OA target set by the GOA. Error matrices found in Figure 6 show the confusion for these models, which indicate minimal commission (599, 2533) and omission (2861, 1747) errors between the general wetland and upland differentiation.

3.2. Prairie/Parkland Modelling Results

The OAs of PPZ models were evaluated directly at the form level, since wetland diversity was low and modelling was performed directly at this detail. The benchmark OA target of 70% was met in both PPZ pilots, and accordingly, the 80% OA class target was also met (Figure 7; see Table 4 for scenario descriptions). In Parkland-1, a highest OA of 93% was achieved with P-DL-S1, and in Grassland-1, a highest OA of 94% with P-DL-S3. All ML models based on the RF algorithm performed poorly relative to XGBoost and U-Net, especially in Parkland-1 where none of the RF models exceeded 70% OA. All DL models for both PPZ pilots achieved 89% OA or higher, although with the caveat that these models could not reliably predict fen wetlands in Parkland-1 (F1-scores of 0.00). P-ML-S9 was the most accurate ML model in Grassland-1, and P-ML-S8 in Parkland-1, with the latter capable of identifying fens.

The per-class F1-scores achieved by the best PPZ AI models are found in

Table 6. PPZ per-class and per-form F1-scores of the best performing AI models.

	F1-Score			F1-Score
Class	Boreal-1 (B-DL-S1)	Boreal-2 (B-ML-S3)	Form	Boreal-1 (B-ML-S3)	Boreal-2 (B-ML-S3)
Water	0.79	0.49	Bare	0.79	0.49
Marsh	0.73	0.74	Graminoid	0.73	0.74
			Tilled	0.00	0.00
Fen	0.00	-	Graminoid	0.00	-
			Shrubby	0.00	-
Swamp	0.18	-	Shrubby	0.18	-
Upland	0.97	0.97	Upland	0.97	0.97

. F1-scores did not vary much between class and form, due to the low wetland diversity in each PPZ pilot—these areas are dominated by open water wetlands and marshes. For example, in Parkland-1, F1-scores for swamps and fens were very low, as these classes are scarcely represented. Bare water and graminoid marsh F1-scores ranged from 0.73 to 0.79 in Parkland-1, and 0.49 to 0.74 in Grassland-1. Tilled marshes were not reliably mapped by any DL model, which is notable since some ML models were able to predict this minority form. Uplands were mapped with high accuracy in both pilots, achieving F1-scores of 0.97.

When evaluated at the more general wetland/upland classification level, the best-performing models achieved OAs of 94.3% in Parkland-1 and 94.6% in Grassland-1, exceeding the 90% OA target. The corresponding error matrices, which also highlight very little omission (22,037, 5,696) and commission (11,015, 14,330) errors, are presented in Figure 8.

4. Discussion

4.1. Analysis of Boreal/Foothills AI Experiments

Our experiments indicated that AI-based modelling can meet the GOA BFZ mapping standards, including target classification detail, accuracy, and MMU (see Figure A2 Appendix A for final BFZ wetland inventory maps). ML modelling with XGBoost produced the best results for form-level mapping in Boreal-1 and Boreal-2, using the B-ML-S3 scenario. This ML scenario integrated satellite imagery and LiDAR derivatives, and represented significant OA increases (3–12%) compared to B-ML-S2 which used the AW3D30 DSM topographic data. Similarly, the inclusion of LiDAR also elevated all DL models for form-level mapping (2–13% OA increases). No ML or DL scenario was able to achieve the 70% OA form-level target without the use of bare earth LiDAR. This finding can be attributed to the higher spatial resolution and quality of airborne LiDAR DTM data, which are capable of preserving important wetland features that manifest from topographic variations [99,100]. LiDAR can also capably separate and characterize vegetation layers that define wetland forms and their communities, by deriving ecologically relevant 3D structural information [101]. This high-resolution, LiDAR-derived information has proven valuable for wetland form mapping in Alberta in other studies [102].

The BFZ pilot results also highlight the effectiveness of ML over DL for detailed wetland form mapping, which we attribute to the rarity of several wetland forms, the fine-scale variability in vegetation communities, and the available reference training dataset. DL models typically require large training datasets, which can make high-thematic resolution mapping challenging with limited samples, and even unsuccessful in some tasks or regions [103]. For instance, several BFZ wetland forms, such as wooded mixedwood swamps and graminoid bogs in Boreal-2, were present but very rare. After modelling, these rare forms were entirely missed by B-DL-S2 (F1-scores of 0.00). Despite applying aggressive DL data augmentation to address class imbalance [104], ML remained more effective for form-level mapping in the BFZ pilots. This aligns with prior studies showing that ML can sometimes outperform DL when sample sizes are limited [105]. El Bilali et al. [106], for example, found XGBoost superior to a DL neural network for RS image analysis, and attributed performance differences to dataset limitations which restricted the DL model’s ability to capture non-linearity. Additionally, Southworth et al. [107] provided a comprehensive review of DL versus ML for land cover classification applications, and found that higher accuracies with DL are not always substantial, nor are they always guaranteed. In fact, the authors identify many studies where ML classification accuracies surpassed DL, noting that the choice in AI algorithm should strongly consider factors such as training data size and RS imagery type.

It is known that DL models tend to sacrifice fine granular details in order to achieve higher performance, and instead focus on more prominent features [108]. This is because the architecture of DL CNNs, like U-Net, are designed to learn high-level features and contextual information very well [109]. The convolutional filters and the pooling processes of CNNs, which are lossy [110], act as edge detectors and tend to produce smooth, coherent outputs which lack “salt-and-pepper” type noise that is more common with ML outputs [111]. An example of this is presented in Figure 9, where predictions from the best ML and DL models are visually compared. At the class level, where wetlands display broader spatial patterns with more distinct boundaries and edges, DL generally demonstrated more favourable performance. Evidently, the aggregation and smoothing processes of U-Net are more effective at this wetland class level, where there is less local and fine-scale variability. Delancey et al. [44] found a similar result when comparing U-Net and XGBoost for class level wetland mapping in northern Alberta, whereby the former produced smoother and more accurate predictions. In the current study, the B-DL-S1 scenario achieved the class level OA target (80%) in both Boreal-1 and -2 pilots, whereas no ML scenario met this target in Boreal-1. However, at the generalized wetland–upland level, both ML and DL demonstrated remarkable performance (98.1% OA), with little omission or commission error between wetlands and non-wetlands.

4.2. Analysis of Prairie/Parkland AI Experiments

Similarly to the BFZ pilots, AI-based modelling met the minimum mapping standards in both the Grassland-1 and Parkland-2 PPZ pilots, and even exceeded the required classification detail by mapping to wetland form detail, despite the provincial mandate specifying only class-level detail ([16]; see Figure A2 Appendix A for final PPZ wetland inventory maps). The highest-performing AI models also achieved accuracies well above provincial targets at the wetland (90%), class (80%), and form (70%) details (Figure 10). DL modelling with U-Net produced the best overall accuracy results at the form detail in both PPZ pilots, although with some notable caveats compared to XGBoost ML models, which are discussed later in this section. In both PPZ pilots, OBIA models using RF yielded underwhelming results, consistent with recent wetland mapping studies that show traditional approaches, while easy to implement and computationally efficient, are consistently outperformed by more advanced and contemporary AI-based methods [112]. In Parkland-1, the best-performing scenario was P-DL-S1, which integrated all available PlanetScope, orthophotography, and LiDAR-derived variables. In Grassland-1, P-DL-S3 performed best, which used similar inputs but excluded CHM and LiDAR intensity. These findings underscore the value of high-resolution topographic inputs for wetland characterization in the PPZ [113,114,115]; however, the contribution of specific topographic variables varies with landscape context. For instance, depression probability and position index metrics show consistently strong associations with PPZ wetland occurrence [116,117], while canopy-related variables (e.g., CHM) are less informative in areas like Grassland-1 where wetland form diversity is limited, but become more valuable in regions with denser vegetation (i.e., from swamps, vegetated peatlands, etc.) that obscures underlying hydrology, such as in Parkland-1. Both DL models also benefited from the inclusion of seasonal PlanetScope imagery and high-resolution orthophotography. Seasonal imagery was particularly important given the highly dynamic nature of PPZ wetlands, which often experience spring flooding and fall drying [118,119,120]. These seasonal dynamics make PPZ wetlands challenging to map using single-date satellite or aerial imagery [121]. Moreover, orthophotos enhanced the DL model’s ability to precisely delineate wetland boundaries, particularly for the many small depressional wetlands (<1 ha) that are common across the PPZ [122].

Despite achieving higher overall accuracies than both ML approaches (i.e., XGBoost and RF), DL models were less effective at mapping rare wetland forms such as fens and swamps. Prairie landscapes often contain extensive wetland cover, but this is typically open water and marshes [123]. CNN-based DL architectures, such as U-Net, are well-suited for learning dominant spatial patterns and therefore tend to generalize well on majority classes, often at the expense of underrepresented ones [104]. This poses challenges for real-world applications, such as operational wetland inventory development with highly imbalanced training data. For example, in Parkland-1, which is closer to the prairie–boreal transition zone and includes some peatland coverage (e.g., shrubby and graminoid fens), the best-performing DL model produced poor per-class F1-scores for these minority forms, both during the modelling validation stage and when compared against field-based reference data. In contrast, the best-performing ML model, P-ML-S8, achieved much higher fen F1-scores in the model development validation stage (e.g., >0.8 for both shrub and graminoid fen forms), and its final predictions aligned better when qualitatively assessed against high-resolution imagery. As such, the XGBoost ML approach produced more spatially realistic peatland distribution estimates across Parkland-1. Figure 11 illustrates this contrast, showing how only the XGBoost-based P-ML-S8 model accurately captured the presence of rare wetland types in a diverse training block, including fens and mixed marsh (e.g., graminoid and tilled). As in the BFZ experiments, this reinforces the adaptability and robustness of ML models to class-imbalanced training datasets, which is an area in which DL continues to experience challenges, particularly for minority class detection [124]. Notably, in more homogeneous landscapes such as Grassland-1, where wetland diversity is lower, DL models excelled with strong performances.

4.3. Study Implications and Recommendations for Provincial Wetland Inventory Mapping

Alberta is home to a vast and diverse array of wetlands. Recognizing their ecological and societal value and the high losses of wetlands due to historical and continued population growth and human pressures, the provincial government has adopted the Alberta Wetland Policy and associated Wetland Replacement Program. The implementation of the policy requires accurate, up-to-date, and spatially explicit information on wetland extent and type to support regulatory, conservation, and monitoring applications [6,9,12,13,125]. Motivation for this study stemmed from the need to modernize Alberta’s wetland inventory using scalable, repeatable, and cost-effective methods that can support ongoing regulatory, management, and conservation needs. Traditional inventory approaches, such as manual air photo interpretation and field-based mapping, while accurate at local scales, lack the efficiency and repeatability needed for consistent province-wide wetland mapping and monitoring updates. The results of this study demonstrate that modern, AI-driven RS approaches, particularly those integrating EO imagery, high-resolution LiDAR, and ML or DL can meet and even exceed the provincial standards for wetland classification accuracy, detail, and MMU. These methods offer a practical path toward operationalizing wetland inventory updates at scale.

However, this study also highlights several important nuances. Scaling these approaches across the province will require careful consideration of regional wetland characteristics, the availability of RS data, and the efficacy and relative return on investment of different AI technologies. As such, implementation strategies should weigh both the ecological complexity of the target biome and the cost-effectiveness of the chosen AI modelling approach and its imagery inputs [126].

Based on the results of this study, we provide the following region-specific recommendations to support the scalable implementation of wetland inventory mapping across Alberta, with separate considerations for BFZ and PPZ.

In the BFZ, recommendations stemming from this study include the following:

• Contributions of LiDAR: LiDAR proved effective in meeting wetland form mapping accuracy targets in the BFZ pilots. Terrain and canopy structure variables were essential for capturing wetland features that manifest from topographic variation, consistent with findings from a broad body of the scientific literature [127]. Given that much of the province’s managed boreal forest area already has, or is planned to receive, high-resolution LiDAR coverage (e.g., from industry, non-governmental organizations, etc.), we recommend that future BFZ wetland inventory efforts incorporate LiDAR-derived inputs to support consistent wetland classification at finer thematic detail (i.e., form).
• Reference data: Field reference datasets are strongly recommended to support AI modelling development and optimization, and to account for subtle hydro-ecological distinctions not easily determined by RS alone. Helicopter-based sampling is particularly favourable for remote area access in the boreal forest, but should be complemented by targeted ground-based surveys to capture critical wetland attributes (e.g., peat depth). Supplementing field campaigns with high-resolution photo interpretation and historical datasets (e.g., photoplots) is also a viable approach, but requires expert validation and quality control. A combined strategy that integrates field, historical, and photo-interpreted data offers a practical balance of quality, feasibility, and spatial coverage for training and validation purposes in the BFZ.
• AI modelling: Both ML and DL approaches performed well in the BFZ pilots, producing ecologically meaningful wetland classifications. U-Net’s strength lies in its ability to produce smooth and coherent wetland class boundaries, while XGBoost excelled at mapping more rare wetland classes (i.e., bogs) and capturing finer form-level granularity and vegetation communities. However, U-Net likely holds greater long-term performance potential as more training and enhanced photoplot data become available. As such, both AI technologies may inform future inventory updates, but DL may consistently surpass ML as data quality and coverage increases.

In the PPZ, recommendations stemming from this study include the following:

• Contributions of LiDAR: Topography is a major driver of wetland presence in the PPZ, making the inclusion of LiDAR data a priority for any future inventory updates. While the contribution of specific LiDAR derivatives varied across PPZ pilots, terrain-based metrics (e.g., probability of depression) proved highly effective—these LiDAR variables were consistently retained as key model inputs following feature selection and optimization. Importantly, alternative open-source DEMs, which have only moderate horizontal and vertical accuracy, are known to be inadequate in characterizing PPZ depressions due to the relatively small size and subtle relief of PPZ wetlands [128]. However, high-resolution LiDAR is not widely available in the Prairie and Parkland areas of Alberta, like it is in the boreal, and these acquisitions are expensive, which is a key challenge for scaling up this approach. Further work is needed to develop cost-sharing strategies to fund high-resolution LiDAR acquisitions for wetland mapping.
• Contributions of high-resolution satellite imagery: Satellite imagery at 3 m resolution is recommended for PPZ wetland inventories, as it provides sufficient granularity for AI models while avoiding the operational challenges and high costs of acquiring airborne imagery across large areas. Platforms with frequent repeat cycles, such as PlanetScope, offer a practical and scalable solution for provincial mapping programmemes.
• Reference data: As in the BFZ, quality training and validation data remain critical. However, operational constraints must be considered. Reliance on newly acquired aerial imagery introduces seasonality challenges, while end-to-end photogrammetric data (i.e., 3D generated photoplots) remain a gold standard but is often cost-prohibitive for larger projects. A shift toward 2D labelling using recent orthophotos or high-resolution satellite imagery is recommended to increase flexibility and scalability. Moreover, plot labelling should be cognizant of temporary, ephemeral wetlands. DL models (e.g., U-Net) are generally more tolerant of label variability in large, lower-quality datasets, whereas ML models like XGBoost require higher label accuracy but can more effectively leverage historical data. Regardless of the AI approach, the underrepresentation of minority wetland classes (e.g., fen, swamp) remains a concern. Targeted field sampling in areas of higher minority class occurrence can help address this imbalance and support more strategic, efficient field programme design.
• AI modelling: In the PPZ, DL achieved the highest overall accuracies but struggled to detect and realistically predict minority wetland classes. ML, particularly XGBoost, outperformed DL in mapping less common wetland types such as fens, highlighting its effectiveness under class-imbalanced training conditions. This suggests that ML approaches may be better suited for operational wetland mapping in areas with greater form diversity. Conversely, U-Net demonstrated strong performance in delineating dominant wetland classes (e.g., marshes, open water). Given these trade-offs, a hybrid strategy that combines the precision of DL for majority classes with the sensitivity of ML for minority forms is recommended for optimal wetland inventory development in Alberta’s PPZ.

4.4. Study Limitations and Future Work

Future efforts would benefit from incorporating a hydroperiod-based approach (i.e., wetland inundation patterns), particularly in the PPZ, given the importance of water presence and duration dynamics in characterizing PPZ wetlands [120,129]. Hydroperiod information may help refine classification outcomes, especially for ephemeral water bodies which are difficult to detect using standard mapping methods [130]. High accuracy hydroperiod mapping of permanent and semi-permanent open water wetlands would help identify public-owned water bodies on private land. These waterbodies are not accurately mapped and require extensive field and GIS work to accurately identify and delineate these wetlands currently. However, generating accurate hydroperiod maps poses its own challenges. The high-frequency, high-resolution imagery (e.g., PlanetScope) needed to capture the fine temporal dynamics of small PPZ wetlands can be cost-prohibitive at scale. On the other hand, open-access sources such as Sentinel-1 and Sentinel-2 offer broader accessibility but lack the spatial and temporal resolution required to track small or short-lived wetlands [33]. This trade-off must be carefully considered.

In the BFZ pilots, active wildfire activity during the 2023 summer field season posed logistical constraints on ground-truth data collection. For instance, approximately 50% of Boreal-1 burned in 2023. Consequently, some areas within the boreal study regions had lower densities of newly collected validation data. This uneven sampling may have influenced AI model accuracy assessments, potentially skewing performance metrics in favour of better-sampled zones [131]. To best account for this, field crews adapted sampling plans in real time and supplemented data gaps with historical, archived surveys. However, future validation campaigns, especially those supporting operational inventory development, must account for the increased frequency and extent of wildfires in Canada’s boreal zone [132], with field protocols that are flexible yet spatially representative.

Natural disturbances such as wildfires can impact wetland classification accuracy and reduce mapping confidence [133]. In this study, both ML and DL models qualitatively performed well in mapping BFZ wetlands within burned areas, exhibiting low noise and minimal confusion (Figure 12). This is attributed in part to the integration of high-resolution bare earth LiDAR, which preserved topo-structural and hydrological features despite fire disturbance, and to the inclusion of intersecting photoplot data that provided wetland training samples over burned areas. These elements enabled the AI models to better learn the spectral characteristics of burned wetlands [134]. However, future research should more rigorously assess model performance in burned areas, particularly the influence of burn severity on wetland classification accuracies.

This study focused on method development in the BFZ and PPZ regions of Alberta, where provincial wetland mapping standards are well-established. In contrast, no dedicated standards currently exist for the Boreal Transition Zone (BTZ) or Rocky Mountain Natural Region. These regions are influenced by complex mountain topography or represent ecotonal zones blending prairie and boreal features, making wetlands there particularly challenging to classify. Evaluating existing AI models or developing tailored approaches for these regions, supported by additional pilot areas, would not only improve model sensitivity to diverse landscapes but also enhance training data coverage, facilitating scaling up to a comprehensive provincial wetland inventory.

Lastly, a key limitation of this study is the lack of water depth information (i.e., bathymetry), since only wetland surface extent was mapped. However, depth is a critical variable for differentiating shallow open water wetlands from deeper systems such as lakes. There are new and emerging satellite-based approaches for regional bathymetry mapping [135,136], but their applicability in Alberta remains uncertain. Future research should test and potentially integrate these scalable methods to better characterize and inform open water wetland classification.

5. Conclusions

This study evaluated the potential of AI-based methods for wetland mapping in the BFZ and PPZ regions of Alberta, Canada. The overall objective was to assess whether advanced ML and DL architectures could meet the provincial wetland mapping standards with respect to classification detail, accuracy, and MMU. A particular emphasis was placed on integrating high-resolution LiDAR data into the AI pipelines to evaluate its contribution to model performance. The results demonstrated that the top-performing AI models met and exceeded provincial standards, achieving over 70% overall accuracy in mapping wetland form detail across all pilots. These methods offer several advantages for operational wetland inventory development: they can efficiently process large volumes of RS data, learn complex spatial patterns, continuously improve with the incorporation of additional training data over time, and produce consistent, repeatable classifications. Compared to traditional photogrammetric approaches, AI-based methods are also more scalable, cost-effective, and less reliant on individual interpreter decisions. DL was shown to be highly effective for wetland mapping, yet data hungry, which has been shown in other studies. This study showed that high-resolution LiDAR across the study areas and high-resolution satellite imagery in the PPZ boosted accuracy, but are expensive to acquire, which is a key challenge in scaling up this approach in the PPZ. This study will inform future wetland inventory updates and wetland monitoring across the province of Alberta.