Wetland mapping with multi-temporal sentinel-1 & -2 imagery (2017 – 2020) and LiDAR data in the grassland natural region of alberta

Figures & data

Figure 1. Map of the study area showing (A) the United States and Canada provinces in the Prairie Pothole Region, (B) the natural subregions of Alberta overlayed with the northwestern extent of the Prairie Pothole Region and (C) a summer Sentinel-2 natural composite of the study area for 2020

Figure 2. Sentinel-2 multispectral time-series images illustrating the dynamic nature of prairie pothole wetlands in spring/summer (May–August) and fall (September–November) months of 2017, 2018, 2019 and 2020

Table 1. Description of wetland and upland classes considered in the study

Landcover class	Description
Marsh*	Marshes are mineral wetlands that represent the transition zone between shallow open water and basin edge and, in some instances, could exist in isolation. Marshes usually exhibit variable water levels at or above the ground surface at different periods throughout the seasons. Marshes periodically turn dry, exposing the ground surface to oxygen, which results in nutrient enrichment. This condition promotes the growth of diverse emergent and grasslike vegetation (graminoids).
Shallow Open Water (later known as open water)*	Shallow open water bodies are mineral wetlands with a water depth up to a maximum of 2 meters. Open water wetlands receive water through a combination of precipitation, runoff and stream inflow. Deep portions of the open water zone support floating and/or submerged aquatic vegetation contingent on nutrient availability. Open water wetlands are usually permanently flooded and experience water level fluctuations over seasons.
Swamp*	Swamps are considered mineral wetlands (≤40 cm peat depth) and could also exist as peatlands (>40 cm peat depth) in some instances (such as a woody plant cover) comprises more than 25% of the total area. The ground surface is characterized by soils derived from minerals and/or rocks containing little organic material even though sub-surface peat exist in certain situations. The water sources for swamps include a combination of precipitation, runoff, and groundwater. Swamps are transition zones between other wetlands and non-wetlands and typically support high biodiversity.
Uplands**	Uplands comprise croplands, grasslands, urban and developed, forests (broadleaved, coniferous, and mixed), exposed and barren, and shrubland classes. Cultivated cropland or cropland are farmlands planted with annual crop species. The croplands includes small grains (wheat, barley, oats and mixed grains), oilseeds (canola, flax), specialty crops (peas, lentils), and row crops (such as potatoes, sugar beets, corn, vegetables). Fallow describes areas used for producing the crops that do not exhibit visible vegetation due to being cultivated. This cropland information was sourced from a combination of ABMI HFI and the Canada Annual Crop Inventory.

* Sourced from the Government of Alberta (GoA) Alberta Merged Wetland Inventory database

** Sourced from the Alberta Biodiversity Monitoring Institute Human Footprint Inventory database and the Canada Annual Crop Inventory. The shallow open water class cited in the GoA merged wetland inventory was renamed open water for the purpose of this study. This reference is reflected in subsequent text and maps cited or generated in the study.

Figure 3. Training and testing data (from Canadian data inventories) used for image classification and accuracy assessment in the study

Table 2. Training and testing data used in the study for image classification and accuracy assessment

LCC ID	Landcover class	Training data		Testing data
		Pixel counts	Area (sq km)	Pixel counts	Area (sq km)
1	Marsh*	544,590	54.5	363,398	36.3
2	Open Water*	468,916	46.9	267,761	26.8
3	Swamp*	2340	0.2	1,588	0.2
4	Cropland**	2,981,795	298.2	4,280,635	428.1
5	Grassland**	3,340,260	334.0	901,194	90.1
6	Urban and Developed**	101,323	10.1	119,503	12.0
7	Forest**	18,364	1.8	42,854	4.3
8	Exposed Land and Barren**	193,342	19.3	119,978	12.0
9	Shrubland**	41,948	4.2	41,041	4.1

* Sourced from the Government of Alberta (GoA) Alberta Merged Wetland Inventory database

** Sourced from the Alberta Biodiversity Monitoring Institute Human Footprint Inventory database and the Canada Annual Crop Inventory.

Table 3. Characteristics of Sentinel-1, Sentinel-2, and LiDAR elevation data explored in the study

Satellite	Input features
	Dates of Sentinel-1 and Sentinel 2 images (dd/mm/yyyy)	Spring/summer: 15/05/2017 – 15/08/2017 15/05/2018 – 15/08/2018 15/05/2019 – 15/08/2019 15/05/2020 – 15/08/2020	Fall: 01/09/2017 – 15/11/2020 01/09/2018 – 15/11/2020 01/09/2019 – 15/11/2020 01/09/2020 – 15/11/2020
Sentinel-1	Channels	VV, VH
	Ratio feature	$\frac{VV}{VH}$
	Number of overlapping images (Sentinel-1)	Spring/summer:	Fall:
		2017 = 51	2017 = 38
		2018 = 49	2018 = 32
		2019 = 54	2019 = 41
		2020 = 53	2020 = 46
Sentinel-2	Selected channels	B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), B11 (SWIR-1), B12 (SWIR-2)
	Spectral indices	$NDVI=\frac{\left(NIR-Red\right)}{\left(NIR+Red\right)}$, $NDWI=\frac{\left(Green-NIR\right)}{\left(Green+NIR\right)}$, $NDSI=\frac{\left(SWIR-NIR\right)}{\left(SWIR+NIR\right)}$
	Number of overlapping images (Sentinel-2)	Spring/summer:	Fall:
		2017 = 55	2017 = 19
		2018 = 207	2018 = 58
		2019 = 164	2019 = 50
		2020 = 178	2020 = 35
	Elevation Data	LiDAR bare earth Digital Elevation Model
	Topographic measure	$TWI=In\left(\frac{A}{tan\left(\beta \right)}\right)$

B: Band, NIR: Near infrared, SWIR: Shortwave infrared, NDVI: Normalized Difference Vegetation Index, NDWI: Normalized Difference Water Index, NDSI: Normalized Difference Soil Index, LiDAR: Light Detection and Ranging, VH: Vertical transmit and horizontal receive polarization, VV: Vertical transmit and vertical receive polarization, and TWI: Topographic Wetness Index (where A is the upslope contributing area, and β is the local slope angle).

Figure 4. Proposed workflow for wetland mapping using multi-temporal freely available Sentinel 1 SAR and Sentinel 2 optical data combined with LiDAR-derived topographic metric

Table 4. Fused datasets used for experiments utilizing multi-seasonal (spring/summer & fall) Sentinel-1 radar and Sentinel-2 multispectral data and LiDAR-derived topographic variables (where B8 = near-infrared band, B11 = shortwave infrared band, VH = vertical and horizontal dual-polarization channel, TWI = topographic wetness index)

Experiment #	Dataset #	Fused data	Classifier	Number of bands	Year
1	1	(B8 + B11 + VH) + TWI	RF	7	2017
2	2	(B8 + B11 + VH) + TWI	RF	7	2018
3	3	(B8 + B11 + VH) + TWI	RF	7	2019
4	4	(B8 + B11 + VH) + TWI	RF	7	2020
5	5	Datasets 1 + 2 + 3 + 4	RF	25	2017-2020
6	6	Datasets 1 + 2 + 3 + 4	CART	25	2017-2020
7	7	Datasets 1 + 2 + 3 + 4	SVM	25	2017-2020

Figure 5. Examples of the reclassified Alberta Biodiversity monitoring institute 3 × 7 photo plot data reference layer used for wetland comparison

Table 5. McNemar’s test results of classifications performed in the study

Dataset/Classifier	Tests	z-score
Dataset 1–5 (RF)	1 vs 2	4.58**
	1 vs 3	1.83
	1 vs 4	6.63**
	1 vs 5	9.06**
	2 vs 3	4.93**
	2 vs 4	8.06**
	2 vs 5	7.81**
	3 vs 4	6.38**
	3 vs 5	9.24**
	4 vs 5	11.22**
Dataset 5–7 (optimal RF classification, SVM and CART classification)	RF vs SVM	17.81**
	RF vs CART	14.53**
	SVM vs CART	10.31**
Optimal RF classification and AMWI wetland map	AMWI vs RF	1.63

Considered significant at a confidence levels of 95% (z ≥ 1.96 (positive or negative)) and 99% (if z ≥ 2.58). Where ** denotes 99% significant.

Table 6. Per-class accuracy results of multi-temporal Sentinel SAR and optical data fused with LiDAR topographic index evaluated in the study (best results in bold and italics)

Random Forest classification experiments (Experiments 1, 2, 3, 4, and 5)
Class	PA (%)	UA (%)	F-score
Dataset 1 – RF
Marsh	39.5	68.7	0.50
Open water	83.1	89.5	0.86
Swamp	73.5	79.1	0.76
Upland	94.1	60.6	0.74
Dataset 2 – RF
Marsh	44.3	68.7	0.54
Open water	84.9	83.6	0.84
Swamp	67.1	80.7	0.73
Upland	95.7	65.0	0.77
Dataset 3 – RF
Marsh	43.8	66.6	0.53
Open water	84.2	85.9	0.85
Swamp	66.3	78.7	0.72
Upland	93.9	63.1	0.75
Dataset 4 – RF
Marsh	39.2	62.4	0.48
Open water	83.4	85.5	0.84
Swamp	65.0	74.8	0.70
Upland	94.3	62.8	0.75
Dataset 5 – RF
Marsh	50.0	78.9	0.61
Open water	86.7	78.4	0.82
Swamp	67.5	85.5	0.75
Upland	97.0	68.0	0.80
CART and SVM classification experiments (Datasets 6 and 7)
Dataset 6 – CART
Marsh	38.4	59.5	0.47
Open water	83.5	83.2	0.83
Swamp	50.4	79.2	0.62
Upland	93.8	57.1	0.71
Dataset 7 – SVM
Marsh	38.4	79.5	0.58
Open water	83.5	83.2	0.75
Swamp	50.4	79.2	0.56
Upland	93.8	57.1	0.64

Figure 6. Graph showing the classification accuracy results in the study. The best performing dataset (Dataset 5) generated using the RF classifier (* denotes RF, ** CART, and *** SVM classified datasets)

Figure 7. Optimal wetland classification generated using multi-temporal Sentinel-2 multispectral and Sentinel-1A radar imagery fused with LiDAR-derived Topographic Wetness Index

Table 7. Summary statistics of the Alberta merged wetland inventory dataset and optimal wetland map generated in this study

Wetland type	Count	Min (ha)	Max (ha)	Mean (ha)	Median (ha)	Sum (ha)	Area percentage
AMWI wetland
Marsh	32,958	0.000025	127.89	0.44	0.21	14,591	66.9%
Open water	3033	0.000025	2,318.11	2.37	0.19	7,177	32.9%
Swamp	153	0.000025	4.57	0.36	0.21	55	0.3%
Total (polygons)	36,144	0.000075	2,450.57	3.17	0.21	21,823	100.0%
Dataset 5 (Experiment 5)
Marsh	31,014	0.04	247.74	0.44	0.14	13,700	54.8%
Open water	1,624	0.04	2,598.85	5.14	0.91	8,355	33.4%
Swamp	9,517	0.04	20.85	0.31	0.12	2,935	11.7%
Total (polygons)	42,155	0.12	2,867.44	5.90	0.14	24,990	100.0%

Table 8. Classification accuracy results of AMWI and Dataset 5 wetland maps

Class Name	Reference totals	Classified totals	Number correct	Producers Accuracy (%)	Users Accuracy (%)
AMWI wetland map
Unclassified	1	54	0	-	-
Marsh	150	172	112	74.7	65.1
Open water	150	160	140	93.3	87.5
Swamp	86	0	0	-	-
Total	386	386	252
Overall Accuracy – 65.3% & Overall kappa value – 0.48
Dataset 5-RF wetland map
Unclassified	0	35	0	-	-
Marsh	150	171	104	69.3	60.8
Open water	149	159	140	94.0	88.1
Swamp	87	21	16	18.4	76.2
Total	386	386	260
Overall Accuracy – 67.4% & Overall kappa value – 0.50

Figure 8. Comparison of AMWI and optimal multi-temporal wetland products of selected sites in the study area – Deadhorse Lake (A and B), Crawling Lake (C and D), and Mattoyekiu Lake (E and F)

Figure 9. Examples of selected areas in Figure 7 showing a visual comparison of the optimal wetland inventory (a, c, and e) and medium resolution Sentinel-2 natural composite image (b, d, and f)

Data and Codes Availability Statement

The authors confirm that the data supporting the findings of this study are available within the article and its supplementary materials.