Census of Environment: Spatial information products
Land Cover Register: Documentation

Acknowledgments

We would like to extend our gratitude to the various government departments, organizations and people who produced the data used to create the land cover register, particularly Agriculture and Agri-Food Canada, Natural Resources Canada, the Canadian Forest Service, the Canada Centre for Mapping and Earth Observation, and Environment and Climate Change Canada. Without their products, the land cover register would not be possible. We would also like to thank the individuals from these departments, as well as from Parks Canada, who provided valuable feedback on this project and suggested improvements for future versions.

1. Introduction

This document describes the approach used to create a national register of land covers to support ecosystem accounting in Canada. The main purpose of the dataset is to identify land cover and land use assets required for ecosystem accounting under the System of Environmental Economic Accounting (SEEA) Ecosystem Accounting framework.^Note 

This dataset, referred to as the land cover register (LCR), has a 30-metre spatial resolution and represents information circa the year 2020.

While many land cover datasets are available in Canada, these data vary in terms of spatial coverage, spatial resolution, periodicity, consistency, class definitions and accuracy. Through the Census of Environment (CoE)^Note  program, Statistics Canada aims to produce data on land cover, land use and ecosystems that follow the SEEA Ecosystem Accounting framework and Central Framework (CF).^Note  As such, the CoE has a unique purpose for land cover maps—the production of environmental statistics in an accounting style. This requires accurate, up-to-date, spatially explicit land cover data with national coverage.

The contributions of ecosystems to society are often taken for granted. Ecosystem accounting aims to make these contributions more visible, by organizing data about ecosystems and linking it to information on economic and other human activity (United Nations et al., 2021). This is done by using spatially explicit data to build account tables related to ecosystem extent (the area covered by various ecosystem types), condition (the health of ecosystems) and services (the benefits that society derives from ecosystems, in both physical and monetary terms).

The ecosystem extent account serves as a basis for measuring ecosystems, including their condition and services. While land cover and land use types are not equivalent to ecosystems, they are often used as a proxy in the absence of national ecosystem maps.

The LCR was created to serve four main purposes within the CoE program:

1. to serve as an accurate, spatially explicit, pixel-based representation of land cover assets^Note 
2. to support spatial analysis for measuring ecosystem condition, characteristics (such as proximity, fragmentation and homogeneity) and services
3. to serve as a base for estimating land cover extent through design-based area estimation (see Textbox 1)
4. to serve as a base and input for building spatially explicit, object-based representations of ecosystem assets and maps.

2. Classification system used for the land cover register

A statistical classification can be defined as “a set of discrete, exhaustive and mutually exclusive categories which can be assigned to one or more variables” (Hancock, 2013). In land cover mapping, a classification refers to the categories of land cover types to which an area can be assigned.

The classification system for the LCR was developed based on the recommended land cover classes in the SEEA CF, which are themselves based on the Food and Agriculture Organization’s (FAO’s) Land Cover Classification System.^Note  Land cover classes included in the SEEA CF recommended classification that do not occur in Canada, including multiple or layered crops and mangroves, were removed from the LCR classification.

The LCR classification was further refined by taking into consideration the availability and quality of land cover data with national coverage in Canada. For example, the two SEEA recommended classes of “grassland” and “shrub-covered areas” were combined into one LCR class, “grassland and shrubland,” because of the difficulty in accurately distinguishing grassland from shrubland with the current state of mapping based on Earth observation. As mapping technologies improve, grassland and shrubland may eventually be divided into two separate classes.

The “treed area” class recommended by the SEEA CF was divided into two separate classes: “treed (non-wetland)” and “treed wetland.” This distinction was made since the data were available and the two types of treed areas are functionally different from an ecosystem perspective, requiring different condition measures and providing different services. This classification approach gives users flexibility to consider treed wetlands separately in their analysis.

A further deviation from the SEEA CF recommended land cover classification was the inclusion of “treed area disturbance” as a separate class. From a land cover perspective, forest areas that have been harvested or exposed to fires would be classified according to their current state of regrowth, which may lead to them being classified as sparsely vegetated land or grassland and shrubland. From a land use perspective, following the FAO definition of forest,^Note  these areas would be considered as “unstocked forest.” However, since disturbed forests are different in terms of their composition, function and the ecosystem services that they provide (Zhang and Wei, 2013), and because of the availability of high-quality national forest disturbance data produced by the Canadian Forest Service (CFS), “treed area disturbance” was included as a unique class in the classification system.

Table 1 shows the 11 classes used in the LCR, along with a description of each class. The number of classes is relatively small, because it is limited by the ability to map more classes with a reasonable degree of confidence, given the current state of mapping technology. Since Canada is a very large and diverse country, each class includes a broad range of ecological properties or characteristics.

Start of text box 1

Textbox 1 – Thematic map accuracy and estimating land cover extent from mapping products

While satellite-based Earth observation (EO) data are crucial for acquiring national, timely and comparable land cover datasets, particularly for a country as vast as Canada, they are not without limitations. Uncertainty in EO-derived land cover data arises from a variety of factors, including limitations of available satellite imagery (because of atmospheric interference, limited temporal resolution, inadequate spatial resolution of sensors to measure features of interest, and spectral similarities between land cover types), and an insufficient number of well-distributed ground truth sites for training and validation. These factors lead to classification errors and limitations in land cover mapping and need to be carefully considered when evaluating the fitness for use of the dataset.

The practice of calculating land cover areas or extent through pixel counting (i.e., summing the areas of pixels of a land cover class in a dataset) does not account for classification errors and, therefore, often leads to biased and inaccurate estimates (Olofsson et al., 2014). For these reasons, using pixel counting to generate extent estimates for ecosystem accounts is discouraged in favour of design-based area estimation, which mitigates bias and quantifies uncertainty (Venter et al., 2024).

In light of this recommendation, the Census of Environment has developed two complementary products for use in ecosystem accounting: (1) the land cover register (LCR), which is a spatially explicit representation of land covers based on EO data, and (2) tables of land cover area estimates at the ecoprovince and sub-drainage area levels, generated through design-based area estimation (Table 38-10-0177 and 38-10-0178). These two products are linked, because the underlying data used for the thematic map accuracy assessment of the LCR were also used as an input for the design-based area estimation. The design-based approach uses a land cover map (in this case, the LCR) to generate a stratified sample, along with a reference dataset that was generated via manual interpretation of high-resolution EO imagery and used to calculate land cover areas. The areas of the manually delineated reference data can also be used to calculate the thematic accuracy of the LCR.

End of text box 1

3. Datasets used

Over recent decades, advances in the field of Earth observation and mapping, combined with those in computing power and data storage, have enabled the production of medium-resolution land cover maps with coverage of large areas. Various government bodies, including Natural Resources Canada’s (NRCan’s) Canada Centre for Mapping and Earth Observation (CCMEO), the CFS, and Agriculture and Agri-Food Canada (AAFC), have produced high-quality multi-class and class-specific land cover maps. Each of these maps was produced for a certain purpose, using a classification system and spatial coverage to suit that need.

An assessment of available land cover maps revealed that no single map fully met the needs of ecosystem accounting: accurate, nationally consistent, up to date, and with the required classes to estimate land cover and (eventually) ecosystem assets. Therefore, Statistics Canada produced the LCR to take advantage of the best available land cover products in Canada by integrating multiple spatial datasets.

The multi-class land cover datasets used as the base for this integration process were the 2020 Semi-decadal Land Use Time Series produced by AAFC (AAFC, 2023), the 2020 Land Cover of Canada produced by NRCan’s CCMEO (NRCan, Canada Centre for Remote Sensing, 2024) and the 2020 Land Cover version 2 produced by the CFS (Hermosilla et al., 2016). The following acronyms are used for these datasets for the remainder of the paper: AAFC SDLU (2020 Semi-decadal Land Use Time Series), CCMEO LCC (2020 Land Cover of Canada) and CFS LCV2 (2020 Land Cover version 2).

The three multi-class land cover datasets vary in terms of their purpose and spatial coverage. The AAFC SDLU covers all areas south of 60°N latitude, the CCMEO LCC covers all of Canada, and the CFS LCV2 covers the forested ecosystems of Canada (see Figure 1). The AAFC SDLU was produced by combining many spatial datasets developed through a range of methods and techniques, which were then integrated using high-quality evidence and visual interpretation (AAFC, 2023). The CCMEO LCC and CFS LCV2 were both developed from the classification of Landsat image composites complemented by ancillary data to improve classification (for more details, see Latifovic et al., 2017; Hermosilla et al., 2022).

The three main datasets were supplemented with other data products to take advantage of each of their strengths on a region-by-region and class-by-class basis, while also preserving national comparability as much as possible. All datasets that were used had national or large regional coverage. While more accurate regional data often exist at the provincial and territorial level or below, these datasets were not considered in order to maintain as much national comparability as possible. Table 2 lists all datasets used in the LCR and indicates whether they were integrated in the register or used as ancillary data to support decision making.

Description

The title of this figure is “Spatial coverage of the three primary datasets used to build the land cover register.” It provides a visual representation of the area covered by each of the three primary datasets that were used in building in the land cover register: Agriculture and Agri-Food Canada’s Land Use Time Series (AAFC SDLU), Canada Centre for Mapping and Earth Observation’s 2020 Land Cover of Canada (CCMEO LCC) and Canadian Forest Service’s Land Cover version 2 (CFS LCV2).

The figure includes three side-by-side maps of Canada of equal size. The spatial coverage of each dataset is indicated in each map by a different colour, and areas where a dataset does not have coverage are white.

The map on the left shows the spatial coverage of the AAFC SDLU in light brown. This dataset has coverage for all areas of Canada south of 60°N latitude.

The map in the centre shows the spatial coverage of the CCMEO LCC in blue. This dataset has coverage for Canada’s full terrestrial and freshwater extent.

The map on the right shows the spatial coverage of the CFS LCV2 in pink. This dataset has coverage for Canada’s forested ecosystems, which includes all areas of Canada except the Prairie ecozone, southern Ontario and Quebec, and the far north.

Note: Spatial coverage of the Semi-decadal Land Use Time Series (left), Land Cover of Canada (centre) and Land Cover version 2 (right).

Sources: Agriculture and Agri-Food Canada (2021). Land Use Time Series. [Dataset]. Retrieved October 1, 2023; Natural Resources Canada, Canada Centre for Remote Sensing (2024). 2020 Land Cover of Canada. [Dataset]. Retrieved October 1, 2023; Natural Resources Canada, Canadian Forest Service (2024). Land Cover version 2, 2020. [Dataset]. Retrieved October 1, 2023.

Legend Table summary
The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description
Light brown	Coverage area of Agriculture and Agri-Food Canada’s Land Use Time Series
Blue	Coverage area of Canada Centre for Mapping and Earth Observation’s 2020 Land Cover of Canada
Pink	Coverage area of Canadian Forest Service’s Land Cover version 2

4. Methodology

4.1 Software and data preparation

Software used to visualize, analyze, edit, confront and integrate data for the LCR included ArcGIS Pro 3.1.2, PCI Geomatica 2017, the Copernicus Browser web application,^Note  Google Earth Pro and Google Earth Engine. All datasets were projected to a common projection and aligned to a common grid prior to data integration (see section 5 for details).

The three primary datasets (i.e., the AAFC SDLU, CCMEO LCC and CFS LCV2) were reclassified to achieve coherence with the LCR classification (Table 1). The concordance tables used to cross-reference each dataset with the LCR classification are Table 3 through Table 5.

4.2 Evaluation of datasets

Each dataset assessed for use in the LCR was evaluated based on several considerations, including coherence with the LCR classification system and comparisons with other spatial data layers and high-resolution imagery. The tools used to evaluate land cover data included high-resolution true-colour imagery from World Imagery (Esri) and Google Earth Pro, false colour composites and Normalized Difference Vegetation Index (NDVI) imagery from Sentinel-2 and Landsat, and vector data sources (including topographic, hydrographic and transportation data). By confronting multiple sources of data against each other, it was possible to determine the optimal source of land cover data based on randomly selected sites where land cover features were manually delineated using high-resolution imagery and on a tile-by-tile basis, as described below.

4.2.1 Determination of base data source

The AAFC SDLU was chosen as the most suitable dataset to use as a base for the LCR because of the alignment of its class definitions with those targeted for the LCR classification system (Table 1) and because of its overall quality. The results of a thematic accuracy assessment based on the interpretation of high-resolution Earth observation imagery of a collection of randomly selected locations indicated that the AAFC SDLU had higher accuracy than the CCMEO LCC (84% agreement for the AAFC SDLU compared with 74% for the CCMEO LCC). ^Note  This supported the decision.

Since the AAFC SDLU does not have coverage north of 60°N latitude, a different base data source was needed for northern Canada. The CCMEO LCC and CFS LCV2 were considered for this purpose. Given the lower apparent accuracy in the territories, a 250 kilometre by 250 kilometre grid was created, and visual inspection was conducted on a tile-by-tile basis using high-resolution imagery interpretation to determine the most accurate source of base data for each tile (either CCMEO LCC or CFS LCV2). The two main sources of high-resolution imagery used for visual interpretation were the World Imagery base maps available through Esri ArcGIS Pro and imagery from Google Earth Pro. This information was supplemented with Sentinel-2 false colour composites and NDVI available as a web service within ArcGIS Pro and through an in-house Google Earth Engine application. The NDVI time series was useful for differentiating between the barren land, sparsely vegetated land, and grassland and shrubland classes.

In the territories, where both the CCMEO LCC and CFS LCV2 datasets have coverage, thematic accuracy assessment based on random sampling indicated that the CFS LCV2 dataset (70% agreement) performed slightly better than the CCMEO LCC (64% agreement). In the Arctic ecozones, where only the CCMEO LCC dataset has coverage, the agreement with manually delineated sampling units was 66%.

North of 60°N latitude, lower accuracy results (66% to 70%, compared with 84% south of 60°N latitude) indicated that additional care was needed to improve the accuracy of the LCR and make it more comparable with the accuracy in the southern portion of the country.

Figure 2 shows the base data source used for each grid tile.

Description

The title of this figure is “Source of base data for each region.” It provides a visual representation of the source of the base data used for each region in building the land cover register.

The figure includes a map of Canada, overlaid with a grid of 250 kilometre by 250 kilometre squares. The grid lines are grey and 60°N latitude is represented by a red line. There is a legend in the top right corner and a scale bar in the bottom left corner. The map is divided into three regions represented by different colours: green, dark orange and lilac. Green indicates areas where Agriculture and Agri-Food Canada’s Semi-decadal Land Use Time Series was used as the base data, dark orange indicates where the Canada Centre for Mapping and Earth Observation’s Land Cover of Canada was used, and lilac indicates where the Canadian Forest Service’s Land Cover version 2 was used.

Note: In cases where 60°N latitude bisects a grid tile, the Semi-decadal Land Use Time Series was used south of 60°N latitude, while the other datasets were used north of 60°N latitude.

Source: Statistics Canada, Environment Accounts and Statistics Division.

Legend Table summary
The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description
Green	Agriculture and Agri-Food Canada’s Semi-decadal Land Use Time Series used as the base data
Dark orange	Canada Centre for Mapping and Earth Observation’s Land Cover of Canada used as the base data
Lilac	Canadian Forest Service’s Land Cover version 2 used as the base data

4.3 Data integration

4.3.1 Harmonization of primary datasets

After determining the base data source for each grid tile, the three reclassified datasets were mosaicked together according to their coverage areas. Where a particular class did not exist in the base layer, that class was created using supplementary data as described in the following paragraphs.

4.3.1.1 Semi-decadal Land Use Time Series coverage area

Since the AAFC SDLU did not include a sparsely vegetated land class, the sparsely vegetated land pixels from the CCMEO LCC were burned into this coverage area.

4.3.1.2 Land Cover version 2 coverage area

The CFS LCV2 did not include equivalents of the following classes: built-up and artificial surface, cropland, treed area disturbance, or sparsely vegetated land. Sparsely vegetated land pixels from the CCMEO LCC were burned into this coverage area.

The cropland class was created in this region using Environment and Climate Change Canada’s (ECCC’s) 2020 Northern Anthropogenic Disturbance (NORDIS) dataset (ECCC, 2020). This dataset maps linear and polygonal anthropogenic disturbances in the northern regions of Canada based on manual detection and digitizing of features identified using 30-metre resolution Landsat satellite imagery. Agriculture polygons from NORDIS were used to identify areas of cropland. Since these polygons were somewhat generalized, only LCR pixels within these features that were classified as one of the following classes at the time of integration were converted to cropland: grassland and shrubland, wetland (non-treed), sparsely vegetated land, or barren land.

The processes for creating the built-up and artificial surface and treed area disturbance classes in this coverage area are described in section 4.4.1.

4.3.1.3 Land Cover of Canada coverage area

The CCMEO LCC did not include equivalent classes for treed wetland or treed area disturbance. The process for creating these classes is described in section 4.4.1.

After integrating the three primary datasets (the AAFC SDLU, CCMEO LCC and CFS LCV2), visual analysis of the transition zones between coverage areas revealed spatial inconsistencies between classes, often indicative of limitations in the source data. These inconsistencies prompted some of the additional improvements made to the LCR that are described in section 4.4.

4.4 Additional improvements to the land cover register

A series of steps were taken to improve the accuracy of the LCR. These steps generally fell into two categories: integrating additional data sources for certain classes and correcting systematic or broad misclassifications in the base layers. Most of these improvements occurred in northern areas, where the accuracy of Earth observation-based maps is generally lower because of issues such as unevenly distributed or lower-quality training sites; limited or suboptimal imagery caused by cloud cover; and, in some regions, the presence of shadows caused by mountainous terrain.

4.4.1 Integrating secondary data sources

4.4.1.1 Wetlands in the territories

In northern areas, a comparison of the base layers against high resolution satellite imagery, other data sources and published literature indicated that wetlands were underestimated. To improve the representation of wetlands in the territories, several wetland datasets were compared using a similar approach to that used to determine the base data source. At first glance, the amount and spatial distribution of open wetlands from different sources were quite varied, prompting further investigation. The wetland datasets considered included the Canadian Wetland Inventory Map version 3 (Mahdianpari et al., 2021); Global Land Cover and Land Use (Hansen et al., 2022); CanVec Wetlands (NRCan, 2023); National Terrestrial Ecosystem Monitoring System for Canada, Land Cover version 2 (Hermosilla et al., 2016); and the Land Cover of Canada (NRCan, Canada Centre for Remote Sensing, 2024).^Note 

For each 250-kilometre grid tile, the most accurate source of wetland data was determined based on visual interpretation of a number of randomly chosen sampling sites using true colour high-resolution imagery and Sentinel-2 false colour composites, as well as supporting vector data sources that included hydrographic and topographic data. In some grid squares, the best option was to combine multiple wetland datasets. Figure 3 shows the source of wetland data used for each 250-kilometre tile in the north.

Description

The title of this figure is “Source of wetland data for each 250 kilometre by 250 kilometre tile.” It provides a visual representation of the data source used for wetlands for each tile in the land cover register.

The figure is a map of Canada, with provinces and territories shown in grey, and their boundaries represented by black lines. North of 60°N latitude, a grid of 250 kilometres by 250 kilometres squares has been overlaid. The source of wetland data for each square is indicated by a different colour. There is a legend in the top right corner and a scale bar in the bottom left corner.

For most of Nunavut and northern Quebec and Labrador, the Canadian Wetland Inventory Map version 3 (CWIM3) was used as the source of wetland data. For the Northwest Territories and Yukon, a variety of different sources were used in addition to CWIM3, including the Land Cover of Canada, Land Cover version 2, CanVec Wetlands and the Global Land Cover and Land Use datasets.

Note: See Table 2 for descriptions of acronyms. For some grid tiles, multiple wetland sources were combined.

Source: Statistics Canada, Environment Accounts and Statistics Division.

Legend Table summary
The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description	Acronym used for data source(s)
Dark blue	Land Cover of Canada and Canadian Wetland Inventory Map version 3 used as sources of wetland data.	CCMEO LCC/CWIM3
Medium green	Land Cover version 2 used as source of wetland data.	CFS LCV2
Dark green	Land Cover version 2, Land Cover of Canada and CanVec Wetlands used as sources of wetland data.	CFS LCV2/CCMEO LCC/CanVec
Pink	Land Cover version 2 and Canadian Wetland Inventory Map version 3 used as sources of wetland data.	CFS LCV2/CWIM3
Burgundy	Land Cover version 2 and CanVec Wetlands used as sources of wetland data.	CFS LCV2/CanVec
Light orange	Land Cover version 2, CanVec Wetlands and Canadian Wetland Inventory Map version 3 used as sources of wetland data.	CFS LCV2/CanVec/CWIM3
Light blue	Canadian Wetland Inventory Map version 3 used as source of wetland data.	CWIM3
Medium blue	CanVec Wetlands used as source of wetland data.	CanVec
Dark purple	CanVec Wetlands and Canadian Wetland Inventory Map version 3 used as sources of wetland data.	CanVec/CWIM3
Light yellow	Global Land Cover and Land Use used as source of wetland data.	GLCLU
Lime green	Global Land Cover and Land Use and Land Cover of Canada used as sources of wetland data.	GLCLU/ CCMEO LCC
Orange	Global Land Cover and Land Use and Canadian Wetland Inventory Map version 3 used as sources of wetland data.	GLCLU/ CWIM3
Dark grey	Global Land Cover and Land Use and CanVec Wetlands used as sources of wetland data.	GLCLU/ CanVec

4.4.1.2 Built-up and artificial surfaces

4.4.1.2.1 Northern Canada

For the built-up and artificial surface class in northern Canada (including the territories, Labrador and northeastern Quebec), additional data sources were used to supplement the base data layers. These included NORDIS, the 2021 Road Network File (Statistics Canada, 2022), the National Road Network file (NRCan, 2024) and the National Railway Network file (NRCan, 2021).

Where these features were not already present in the base layers, railways from the National Railway Network file and roads from the 2021 Road Network File and the National Road Network file were added to the built-up and artificial surface class.

In addition, features from the following classes were integrated from NORDIS: settlement, airstrip, cleared land, mine, oil/gas, roads, powerlines and dams. In some cases, features from this dataset were corrected for positional accuracy using high-resolution imagery prior to integration.

Finally, in regions where the CFS LCV2 was used as the source of the base data, the built-up class from the CCMEO LCC was also burned into the LCR.

4.4.1.2.2 All of Canada

Peat extraction sites were identified using the Annual Crop Inventory (ACI) 2019 and 2023^Note  (AAFC, 2024) and integrated into the built-up class. Resource extraction sites, which were inconsistently mapped as built-up and artificial surfaces in the LCR, were identified using the CanVec Resources Management Features (NRCan, 2023). These were corrected for positional accuracy using high-resolution imagery and then integrated into the built-up class.

4.4.1.3 Cropland

Orchards and other fruit-growing parcels were not always classified as cropland in the preliminary LCR dataset. To rectify this, the 2021 Census of Agriculture (Statistics Canada, 2023) was used to identify areas with the highest density of fruit-growing parcels. The ACI was used to identify fruit parcels (such as orchards, blueberries and cranberries) within these regions, and these were then integrated into the cropland class of the LCR.

4.4.1.4 Grassland and shrubland in the Prairie ecozone

Confrontation of multiple data sources revealed frequent misclassification of grassland as cropland in the Eastern Boreal Plains and Eastern Prairie ecoprovinces. To improve this, pixels classified as grassland in ECCC’s Prairie Ecozone Grasslands dataset (Pouliot et al., 2021) and also in the ACI (AAFC, 2024) in either 2020 or 2021 were reclassified as grassland and shrubland in the LCR.

4.4.1.5 Treed area disturbance

4.4.1.5.1 Northern Canada

The treed area disturbance class was created in the territories using the CFS’s Canada Landsat Burned Severity product 1985-2020 (CanLaBS) data (Guindon et al., 2020) for fire disturbance and ECCC’s NORDIS data (ECCC, 2020) for harvest disturbance, as described in the following paragraphs.

The CanLaBS data product contains variables related to fire burn severity for all fire events in Canada between 1985 and 2020.^Note  Pixels with a fire severity (based on delta normalized burn ratio [dNBR]) of at least 100 and that occurred during or after the year 1990 were included in the treed area disturbance class if those locations were identified as forest according to the CFS’s FAO Forest layer (Wulder et al., 2020). Since the CFS FAO Forest layer classifies both existing forest and trees that have been removed by disturbance as forest, this helped to ensure only areas treed before the disturbance were included in the treed area disturbance class.

Polygon features classified as either harvest or cutblocks in NORDIS were integrated into the LCR to build the harvest portion of the treed area disturbance class in the territories.

4.4.1.5.2 Southern Canada

South of 60°N latitude, the treed area disturbance class originated from the AAFC SDLU. However, disturbances that occurred in the years just before 2020 were not always captured in this layer. Using CanLaBS and the CFS’s Forest Harvest layer (Hermosilla et al., 2016) for fires and harvest, respectively, areas south of 60°N latitude were updated with more recent disturbance data. Areas were classified as disturbed only if they were forest according to the FAO Forest layer and, in the case of fire, if they had a dNBR value of at least 100.

4.4.1.6 Treed wetlands

Inconsistencies in the treed wetland class were evident in several locations in the LCR in the Taiga Plains and Boreal Shield ecozones, where sharp geometric boundaries between the treed (non-wetland) and treed wetland classes were evident. Confronting the LCR against other data sources, including a 250-metre resolution peatland map produced by NRCan (2011), indicated that the treed wetland class was likely underestimated in these areas. Since a national 30-metre resolution data source for treed wetlands could not be identified, treed wetlands were estimated using the procedure below.

The presence of black spruce (Picea mariana) and tamarack (Larix laricina), as well as low forest stand height and stand age greater than 75 years, have been found to be strong predictors of the presence of peatlands (Thompson et al., 2016). Using 30-metre resolution data from the Spatialized CAnadian National Forest Inventory data product (SCANFI; Guindon et al., 2023) for tree species cover and vegetation height, combined with forest age data from the CFS (Maltman et al., 2023), treed wetland presence was estimated using thresholds based on the findings of Thompson et al. (2016). Any pixel that met the following criteria was considered as treed wetland:

1. within the areas indicated as treed or forested peatland in the NRCan 250-metre peatland map
2. more than 50% black spruce cover or more than 20% tamarack cover
3. not more than 40% pine cover (jack pine and lodgepole pine combined)^Note 
4. age greater than 50 years^Note 
5. height less than 15 metres.

The resulting pixels were integrated into the LCR as treed wetland in the territories and the following ecoprovinces: Hay-Slave Lowlands, Western Boreal Shield and Mid-Boreal Shield. Only pixels that were classified as treed (non-wetland) in the LCR at the time of integration were reclassified as treed wetland.

4.4.2 Correcting systematic misclassifications and inconsistencies

Several steps were implemented to improve large-scale patterns of inconsistencies or misclassifications on a class-by-class and region-by-region basis. These misclassifications were identified by confronting multiple data sources, including high-resolution imagery, against each other. After identifying areas for improvement, alternate data sources were evaluated to determine a more suitable data source, a coverage area for the improvement was delineated, and decision rules were created for integrating the data into the LCR (for example, only replacing pixels of certain classes). The steps were executed in a sequence to preserve precedence. Not all data integration steps are included in this document; only those with the largest impact on the final product are described in the following paragraphs.

Scattered pixels of certain land cover classes in improbable locations were identified, such as cropland in the far north, scattered built-up areas in inaccessible rocky regions in northern Quebec and Labrador, and treed pixels located above the known tree line (based on the Circumpolar Arctic Region Tree Line [Walker & Raynolds, 2018]). These were corrected after verifying from photo interpretation of high-resolution imagery that they had been misclassified. These pixels were reclassified using either a nearest neighbour operation or the most probable land cover type for the area, determined based on visual analysis of high-resolution imagery. In the north, segments of winter roads occurring over water had been classified as built-up in the source data; these were reclassified as water in the LCR.

On steep slopes facing north and east in the mountainous terrain of the territories, areas frequently or permanently in shadow were often misclassified as water. To correct for this misclassification, the affected regions were delineated and, within these areas, only pixels intersecting water from the CanVec Waterbodies dataset were retained as water. Other water pixels were reclassified using a nearest neighbour operation.

The CanVec Waterbodies dataset was also used to improve the representation of watercourses in the LCR. Medium-sized watercourses were often inconsistently classified in the source datasets. They may have been classified as water, but also as wetland, grassland and shrubland, or barren land. To improve consistency, all watercourses (class 91) from CanVec Waterbodies were burned into the LCR for the entirety of the country.

Visual assessment of the boundaries between the different base datasets revealed sharp transitions from treed to non-treed areas. Confrontation of several datasets and high-resolution imagery indicated frequent misclassification of non-treed areas as treed areas in these cases. To mitigate this, pixels classified as treed in the LCR that did not have at least 10% canopy cover according to SCANFI and that were not classified as treed in either the CCMEO LCC or the CFS LCV2 were reclassified using the CCMEO LCC.

To improve the permanent snow and ice class, CanVec’s Snow and Ice dataset was integrated in certain locations where the quality was determined to be better than the base data. Where permanent snow and ice were overestimated in the source data (as determined by comparisons with CanVec Snow and Ice and high-resolution imagery), these pixels were reclassified either as barren land or by using the value of a different land cover dataset, depending on the location.

In some areas where the CFS LCV2 was used as the base data, a comparison of NDVI values from July and August 2020 (when the maximum value is reached for the year) indicated that pixels classified as barren land had values that were more closely aligned with the sparsely vegetated land class elsewhere in the LCR. These pixels were reclassified as sparsely vegetated land.

4.4.3 Ensuring complete coverage of Canada

Several steps were implemented to ensure that the LCR had complete coverage of Canada’s land and freshwater areas. The administrative boundaries of Canada (Statistics Canada, 2021) and the International Boundary Commission’s border (International Boundary Commission, 2018) were used to identify the existence of small gaps with no land cover data. This occurred where the AAFC SDLU was used as the base layer and was corrected by filling in these gaps with data from the CCMEO LCC.

Thousands of small islands, mostly less than one hectare in area, were not included in the base datasets. This may have been a result of techniques such as smoothing of the land cover data in the input datasets to remove isolated pixels. These islands were added to the LCR using the administrative boundaries of Canada (Statistics Canada, 2021). Since no land cover data were available for these islands, they were classified as barren land if they occurred in coastal waters and as treed (non-wetland) if they occurred in inland waters, according to the most probable land cover for each region (coastal or inland) based on image interpretation.

The LCR was clipped to a customized version of the ecoprovinces of Canada to ensure full and accurate coverage of Canada’s terrestrial and freshwater extent. The customized ecoprovince boundaries had been adjusted to include the Canadian portion of lakes that are shared between Canada and the United States (such as the Great Lakes). Portions of lakes that were added to the revised boundaries were assigned to the adjacent ecoprovince.

5. Technical specifications

Spatial resolution: 30 metres

Extent

Top: 5,427,990 metres
Bottom: 615,780 metres
Left: 3,613,650 metres
Right: 9,019,800 metres

The LCR uses a customized Albers equal-area projection that is used for all CoE data products. It preserves area and is a conical projection with two standard parallels and a central meridian well-suited for Canada.

PROJCRS["NAD83 / Statistics Canada Ecosystem Register Albers",
    BASEGEOGCRS["NAD83",
        DATUM["North American Datum 1983",
            ELLIPSOID["GRS 1980",6378137,298.257222101,
                LENGTHUNIT["metre",1]]],
        PRIMEM["Greenwich",0,
            ANGLEUNIT["degree",0.0174532925199433]],
        ID["EPSG",4269]],
    CONVERSION["Statistics Canada Ecosystem Register Albers",
        METHOD["Albers Equal Area",
            ID["EPSG",9822]],
        PARAMETER["Latitude of false origin",63.390675,
            ANGLEUNIT["degree",0.0174532925199433],
            ID["EPSG",8821]],
        PARAMETER["Longitude of false origin",-91.8666666666667,
            ANGLEUNIT["degree",0.0174532925199433],
            ID["EPSG",8822]],
        PARAMETER["Latitude of 1st standard parallel",49,
            ANGLEUNIT["degree",0.0174532925199433],
            ID["EPSG",8823]],
        PARAMETER["Latitude of 2nd standard parallel",90,
            ANGLEUNIT["degree",0.0174532925199433],
            ID["EPSG",8824]],
        PARAMETER["Easting at false origin",6200000,
            LENGTHUNIT["metre",1],
            ID["EPSG",8826]],
        PARAMETER["Northing at false origin",3000000,
            LENGTHUNIT["metre",1],
            ID["EPSG",8827]]],
    CS[Cartesian,2],
        AXIS["(E)",east,
            ORDER[1],
            LENGTHUNIT["metre",1,
                ID["EPSG",9001]]],
        AXIS["(N)",north,
            ORDER[2],
            LENGTHUNIT["metre",1,
                ID["EPSG",9001]]]

    ]

6. Results

Map 1 shows the percentage of pixels edited by the improvements described in the methodology. A change in pixel values indicates that a pixel was modified after the primary datasets were reclassified and mosaicked together based on their coverage area. Around 943,000 square kilometres of land were modified during the process, representing 9.5% of all land pixels. The total amount of land affected varies by ecoprovince from 0.1% to 30.4%, with most edits occurring between latitudes 55°N and 75°N. The classes most affected by the edits were treed (non-wetland), wetland (non-treed), and grassland and shrubland.

A map of the register is shown in Map 2.

Description

The title of this figure is “Proportion of pixels affected by improvements to base layers, by ecoprovince.” It provides a visual representation of the percentage of pixels edited by the improvements described in the Land Cover Register: Documentation, by ecoprovince.

The figure is a map of Canada, with ecoprovince boundaries represented by black lines. There is a legend in the top right corner and a scale bar in the bottom left corner. Ecoprovinces are labelled with codes. Below the map, ecozone and ecoprovince names are listed with their codes.

Ecoprovinces are symbolized in one of five classes, according to the percentage of pixels that were changed in the land cover register data integration process. A change in pixel values indicates that a pixel was modified after the primary datasets were reclassified and mosaicked together based on their coverage area. The proportion of improvements varies by ecoprovince from 0.1% to 30.4%, with most edits occurring between latitudes 55°N and 75°N.

Source: Statistics Canada, Environment Accounts and Statistics Division.

Legend Table summary
This table displays the results of . The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description
Light yellow	Less than 2% of pixels affected by improvements
Light green	Between 2% and 5% of pixels affected by improvements
Light blue	Between 5% and 10% of pixels affected by improvements
Medium blue	Between 10% and 20% of pixels affected by improvements
Dark blue	Between 20% and 30% of pixels affected by improvements

Description

The title of this figure is “Land cover register, circa 2020.” It provides a visual representation of the land cover register.

The figure is a map showing the spatial distribution of land cover types in Canada. There is a legend in the top right corner and a scale bar in the bottom left corner.

Each pixel in the map is represented by one of eleven land cover classes: built-up and artificial surface, cropland, inland water body, treed (non-wetland), treed wetland, treed area disturbance, grassland and shrubland, wetland (non-treed), sparsely vegetated land, barren land or permanent snow and ice.

In major urban areas, such as Toronto, Montréal, Ottawa, Edmonton, Calgary and Vancouver, built-up and artificial surfaces are visible in the map as the main land cover type. In the Prairie ecozone and much of the Mixedwood Plains, cropland is the dominant land cover type. A wide swath of the country from the Atlantic provinces in the southeast to British Columbia and Yukon in the west is distinguished by having treed (non-wetland), treed area disturbance and treed wetland as the dominant land cover types. South of Hudson Bay and west of James Bay, wetlands are present in a large concentration. In much of northern Canada above the tree line, grassland and shrubland is the dominant land cover class, while in the far northern islands of the Arctic Archipelago, sparsely vegetated land, barren land and permanent snow and ice are dominant.

Source: Statistics Canada, Environment Accounts and Statistics Division.

Legend Table summary
The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description
Burgundy	Built-up and artificial surface
Pink	Cropland
Light blue	Inland water body
Dark green	Treed (non-wetland)
Dark purple	Treed wetland
Medium green	Treed area disturbance
Lime green	Grassland and shrubland
Purplish pink	Wetland (non-treed)
Pale greenish yellow	Sparsely vegetated land
Grey	Barren land
White	Permanent snow and ice

6.1 Accuracy of the land cover register

As part of the design-based area estimation described in Textbox 1, CoE staff manually delineated land cover types for 5,011 100 m by 100 m sampling units.^Note  These units were selected across all ecoprovinces and included all land cover types occurring in each ecoprovince (with the treed [non-wetland] and treed wetland classes combined, as noted below). The spatial distribution of sampling units is shown in Appendix 1. Within these units, land cover features were manually delineated by trained CoE staff using high-resolution satellite imagery interpretation and the datasets listed in Table 2. For consistency, staff adhered to a document describing delineation guidelines to follow. Units were also validated by a second reviewer. The LCR was confronted with the sampling units to produce a confusion matrix, which was used to quantify the level of agreement between the two sources. The treed wetland and treed (non-wetland) classes were combined for this assessment because of the difficulty in distinguishing between these two classes using image interpretation.

Table 6 displays the level of agreement by land cover type at the national level. It is expressed in the form of the F-score, which quantifies how well the LCR maps the actual land cover type.  The F-score combines results of the user’s and producer’s accuracies.  The overall weighted F-score is 74.9%. The degree of agreement is not uniform across land cover classes. Four land cover classes have a high level of agreement (greater than 80%), covering 48.3% of the country’s inland area. They are cropland, inland water body, treed (including treed wetlands), and permanent snow and ice. The remaining classes are between 60.7% and 76.8% agreement, aside from sparsely vegetated land which has an agreement of 43.4%.  Appendix 2 shows the agreement (F-score) between the reference data and the LCR by ecozone and land cover class.

It should be noted that the difference in the spatial resolution of the LCR and the reference data (which are more accurate because they are derived from high-resolution imagery) partly contributes to the disagreement. Because of the spatial resolution of most of the input datasets, the LCR cannot map small features (less than 900 square metres in area) and coarsely represents boundaries between features. Open wetlands, in particular, are subject to omission in the mid-resolution LCR because they are numerous, they are often small, and they can have a narrow linear shape. Nonetheless, it is not surprising that the grassland and shrubland, wetland (non-treed), sparsely vegetated land, and barren land classes are less accurately mapped, given the similarities in spectral characteristics between certain pairings of these classes and the frequent presence of transitional zones or ecotones between these land cover types.

It should also be noted that, to a small extent, interpretation errors may have been a factor affecting the accuracy assessment. Some classes are difficult to interpret from high-resolution imagery, and despite the use of multiple lines of evidence (e.g., NDVI values, colour infrared imagery, various spatial data sources at different points in time), errors in manual classification may have occurred.

7. Limitations of the dataset

This dataset was created primarily for ecosystem accounting and analysis. It is meant to represent the best spatially explicit estimates of land cover assets, for spatial analysis in measuring ecosystem condition and services, and for use as an input for generating land cover area estimates using a design-based approach. The dataset is not primarily intended for mapping and visualization. For this reason, trade-offs were made in integrating the best available data on a large regional level, resulting in some spatial inconsistencies in the data between classes, particularly in classes that are more difficult to map or identify, such as wetlands in northern areas.

This dataset is meant for use at the national or large regional scale and is not recommended as a primary source in local or finer-resolution analysis or reporting. Additionally, the dataset is not designed for or intended to support change detection. While new versions of the LCR may be published in the future, it is not recommended to infer land cover change by simply comparing products, since perceived changes could be the result of differences in the datasets used as inputs. These can include differences in class definitions between datasets, temporal variability and misclassification, as opposed to real land cover change.

While misclassifications can occur in all classes, as explained in Textbox 1, certain class-specific limitations should be noted. The treed area disturbance class may include disturbances resulting from phenomena other than fire and harvest, such as insect infestation and wind damage. Further, partial harvesting (e.g., through selective logging) may not be captured in the disturbance data.

As noted in the previous section, wetlands are underrepresented in some regions. For example, in the Prairie region, prairie potholes are wetlands that are often small and/or narrow in shape. They collectively make up a large proportion of the land base, but are likely to be misclassified in this data product because of their size in relation to the spatial resolution of the available land cover data and the influences of temporal variability.

Since treed wetlands are difficult to differentiate from treed uplands using image interpretation, the accuracy of the treed wetland class was not assessed. Users are advised to exercise caution when using this class.

Linear features, such as roads, railways and powerlines were integrated into the LCR. Although these features are generally narrower than the 30 m resolution of the product and are often not the dominant land cover in a pixel, they were included since they can have significant impacts on ecosystems.

Users should exercise caution when using the LCR in coastal areas. Since boundaries between freshwater, intertidal and coastal waters are difficult to define, dynamic and transitional in nature, some areas along the coastline identified in the LCR as inland water may be coastal or marine waters. Where water courses drain into coastal waters, portions that could be considered coastal or marine may be classified as inland water in the LCR; alternately portions that could be considered freshwater may have been omitted. Additionally, misclassifications are more common in these dynamic and transitional areas along the coast because of factors such as shifting coastlines, fluctuating water levels, areas where freshwater and saltwater mix and other hydrological factors. Misalignment of different spatial data products may also lead to classification errors.

It should also be noted that merging and manipulating (e.g., reprojecting or snapping to a common grid) multiple spatial datasets can cause some spatial misalignments in the end product. For this reason, users may wish to refer to the source data, where suitable.

8. Future work

It is anticipated that the LCR will be updated as new technology, data sources and better-quality data become available. These improvements may include further integration of biophysical properties and characteristics to enable increased detail in the land cover classes.

9. References

AAFC [Agriculture and Agri-Food Canada] (2024). Annual Crop Inventory 2019 to 2023. [Dataset].

AAFC [Agriculture and Agri-Food Canada] (2023). Land Use 2020. [Dataset].

Di Gregorio, A., & Jansen, L. J. M. (2000). Land Cover Classification System: Classification Concepts and User Manual.

ECCC [Environment and Climate Change Canada] (2020). Northern Anthropogenic Disturbance Mapping (NORDIS). [Unpublished dataset].

FAO [Food and Agriculture Organization of the United Nations] (2018). Terms and Definitions, Global Forest Resources Assessment 2020.

Guindon, L., Villemaire, P., Correia, D. L. P., Manka, F., Lacarte, S., & Smiley, B. (2023). SCANFI: Spatialized CAnadian National Forest Inventory data product. [Dataset]. Natural Resources Canada, Canadian Forest Service, Laurentian Forestry Centre, Quebec, Canada.

Guindon, L., Villemaire, P., Manka, F., Dorion, H., Skakun, R., St-Amant, R., & Gauthier, S. (2020). Canada Landsat Burned Severity (CanLaBS): A Canada-wide Landsat-based 30-m resolution product of burned severity since 1985. [Dataset].

Hancock, A. (2013). Best Practice Guidelines for Developing International Statistical Classifications.

Hansen, M. C., Potapov, P. V., Pickens, A. H., Tyukavina, A., Hernandez-Serna, A., Zalles, V., Turubanova, S., Kommareddy, I., Stehman, S. V., & Song, X.-P. (2022). Global land use extent and dispersion within natural land cover using Landsat data. Environmental Research Letters, 17 (3).

Hermosilla, T., Wulder, M. A., White, J. C., Coops, N. C., Hobart, G. W., & Campbell, L. B. (2016). Mass data processing of time series Landsat imagery: Pixels to data products for forest monitoring. International Journal of Digital Earth, 9 (11), 1035-1054.

Hermosilla, T., Wulder, M. A., White, J. C., & Coops, N. C. (2022). Land cover classification in an era of big and open data: Optimizing localized implementation and training data selection to improve mapping outcomes. Remote Sensing of Environment, 268.

International Boundary Commission (2018). Canada / United States of America Boundary. [Dataset].

Latifovic, R., Pouliot, D., & Olthof, I. (2017). Circa 2010 Land Cover of Canada: Local Optimization Methodology and Product Development. Remote Sensing, 9 (11).

Mahdianpari, M., Brisco, B., Granger, J., Mohammadimanesh, F., Salehi, B., Homayouni, S., & Bourgeau-Chavez, L. (2021). The Third Generation of Pan-Canadian Wetland Map at 10 m Resolution Using Multisource Earth Observation Data on Cloud Computing Platform. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 8789-8803.

Maltman, J. C., Hermosilla, T., Wulder, M. A., Coops, N. C., & White, J. C. (2023). Estimating and mapping forest age across Canada’s forested ecosystems. Remote Sensing of Environment, 290.

NRCan [Natural Resources Canada] (2024). National Road Network. [Dataset].

NRCan [Natural Resources Canada] (2023). Topographic Data of Canada - CanVec Series. [Dataset].

NRCan [Natural Resources Canada] (2021). National Railway Network. [Dataset].

NRCan [Natural Resources Canada] (2011). Distribution of peatlands in Canada using National Forest Inventory forest structure and ancillary land cover data. [Dataset].

NRCan [Natural Resources Canada], Canada Centre for Remote Sensing (2024). 2020 Land Cover of Canada. [Dataset].

Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., & Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148 (25).

Pouliot, D., Alavi, N., Wilson, S., Duffe, J., Pasher, J., Davidson, A., Daneshfar, B., & Lindsay, E. (2021). Assessment of Landsat Based Deep-Learning Membership Analysis for Development of from–to Change Time Series in the Prairie Region of Canada from 1984 to 2018. Remote Sensing, 13 (4).

Statistics Canada (2023). 2021 Census of Agriculture. [Dataset].

Statistics Canada (2022). 2021 Census – Road network file. [Dataset].

Statistics Canada (2021). 2021 Census – Boundary files. [Dataset].

Thompson, D. K., Simpson, B. N., & Beaudoin, A. (2016). Using forest structure to predict the distribution of treed boreal peatlands in Canada. Forest Ecology and Management, 372, 19-27.

United Nations et al. (2021). System of Environmental-Economic Accounting—Ecosystem Accounting (SEEA EA). White cover publication, pre-edited text subject to official editing.

United Nations et al. (2014). System of Environmental-Economic Accounting 2012—Central Framework.

Venter, Z. S., Czúcz, B., Stange, E., Nowell, M. S., Simensen, T., Immerzeel, B., & Barton, D. N. (2024). “Uncertainty audit” for ecosystem accounting: Satellite-based ecosystem extent is biased without design-based area estimation and accuracy assessment. Ecosystem Services, 66.

Walker, D. A., & Raynolds, M. K. (2018). Circumpolar Arctic Vegetation, Geobotanical, Physiographic Maps, 1982-2003. [Dataset].

Wulder, M. A., Hermosilla, T., Stinson, G., Gougeon, F. A., White, J. C., Hill, D. A., & Smiley, B. P. (2020). Satellite-based time series land cover and change information to map forest area consistent with national and international reporting requirements. Forestry: An International Journal of Forest Research, 93 (3), 331-334.

Zhang, M., & Wei, X. (2013). Alteration of flow regimes caused by large-scale forest disturbance: A case study from a large watershed in the interior of British Columbia, Canada. Ecohydrology, 7 (2).

10. Appendices

10.1 Appendix 1

Description

The title of this figure is “Distribution of sampling units.” It provides a visual representation of the location and distribution of sampling units used in the design-based area estimation procedure.

The figure is a map of Canada with locations of sampling units shown as small red points. Ecoprovinces are shown in light grey with black boundaries. There is a scale bar in the bottom left corner and a legend in the top right corner. The sampling units were selected across all ecoprovinces and included all land cover types occurring in each.

Source: Statistics Canada, Environment Accounts and Statistics Division.

Legend Table summary
The information is grouped by Symbol (appearing as row headers), , calculated using (appearing as column headers).

Symbol	Description
Red point	Location of sampling unit
Grey	Ecoprovince

10.2 Appendix 2