The North American Prairie Pothole Region (PPR) represents a large system of wetlands with great importance for biodiversity, water storage and flood management. Knowledge of seasonal and inter-annual surface water dynamics in the PPR is important for understanding the functionality of these wetland ecosystems and the changing degree of hydrologic connectivity between them. Optical sensors that are widely used for retrieving such information are often limited by their temporal resolution and cloud cover, especially in the case of flood events. Synthetic aperture radar (SAR) sensors can potentially overcome such limitations. However, water extent retrieval from SAR data is often impacted by environmental factors, such as wind on water surfaces. Hence, robust retrieval methods are required to reliably monitor water extent over longer time periods .
The aim of this study was to develop a robust approach for classifying open water extent in the PPR and to analyse the obtained time series covering the entire available Sentinel-1 observation period from 2015 to 2020 in the hydrometeorological context. Open water in prairie potholes was classified by fusing dual-polarised Sentinel-1 data and high-resolution topographical information using a Bayesian framework. The approach was tested for a study area in North Dakota. The resulting surface water maps were validated using high-resolution airborne optical imagery. For the observation period, the total water area, the number of waterbodies and the median area per waterbody were computed. The validation of the retrieved water maps yielded producer’s accuracies between 84 % and 95 % for calm days and between 74 % and 88 % for windy days. User’s accuracies were above 98 % in all cases, indicating a very low occurrence of false positives due to the constraints introduced by topographical information.
The observed dynamics of total water area displayed both intra-annual and inter-annual patterns. In addition to differences in seasonality between small (
Surface water dynamics in wetland ecosystems play an important role in water storage variability of a region
The Prairie Pothole Region (PPR) of North America covers an area of over 780 000 km
Direct evidence to support such conceptual models of water surface dynamics in the PPR is mainly based on Earth observation
However, HR satellite or airborne imagery is typically not available at the very short time intervals necessary to resolve intra-annual variations in waterbody sizes. Moreover, the flood mitigation potential of the wetlands in the PPR is a function of the water volume existing at the beginning of a flood event
Such limitations, along with the short observation periods currently available for most SAR missions (in comparison to e.g. Landsat), pose the greatest hindrances for a wider uptake of SAR data for a long-term monitoring of wetland dynamics. In contrast to most SAR missions to date, the two-satellite Sentinel-1 constellation focuses on providing consistent data over longer time periods
Here, a retrieval algorithm for open waterbodies in the PPR based on dual-polarisation Sentinel-1 data is proposed. We use a probabilistic approach based on Bayes' theorem combining SAR backscatter and information derived from a lidar-based digital elevation model in order to minimise the occurrence of false positives, caused by bare areas, tarmac or wet snow, which has been identified as a limiting factor in the aforementioned study
Digital elevation model of the Pipestem Creek catchment. The inset shows the location of the study area within the PPR. Dashed lines show the footprints of the aerial imagery used for validation. OpenStreetMap is used as the background. The elevation data are provided by the North Dakota State Water Commission (2018) under the Creative Commons license.
The study area comprises the Pipestem Creek catchment in North Dakota (ND), USA (Fig.
Discharge in the Pipestem Creek shows large variability (Fig.
Surface water dynamics were derived from a stack of Sentinel-1 interferometric wide swath (IWS) images. Sentinel-1A was launched in April 2014, and its twin satellite Sentinel-1B followed in April 2016. Both satellites carry a C-band SAR instrument operating at a wavelength of ca. 5.6 cm. The ground range detected (GRD) product has a spatial resolution of ca. 20 m
A digital terrain model (DTM) with a resolution of 1 m was available from the North Dakota
As a source of information on land use/land cover, the U.S. Department of Agriculture (USDA) NASS CDL for North Dakota was downloaded (henceforth referred to as CDL). The CDL for 2015 is based on data from Landsat 8, DEIMOS-1 and UK2 that were acquired during that year. The CDL has a spatial resolution of 30 m and is provided under a Creative Commons licence
The surface water extents were validated for three different dates in 2016, 2017 and 2019 using aerial imagery from the National Agriculture Imagery Program (NAIP, U.S. Department of Agriculture, Farm Service Agency). These years were selected to have a representation of dry and wet years. The three NAIP scenes cover different parts of the study area (Fig.
Calm, open water surfaces typically cause specular reflection of incident microwave radiation. In SAR scenes with an approximately balanced mix of open water and land classes, this phenomenon leads to bimodal grey-value histograms. However, these classes rarely occur at similar proportions within a scene, and, as a result, bimodality of the histogram is not commonly observed. Efforts have been made to balance water and non-water classes by subsampling areas from SAR images where water and land pixels are equally represented
The locations of prairie potholes, however, are governed by the landscape features of the PPR and have been reported to be relatively stable over longer time periods
An overview of the water classification workflow is shown in Fig.
Sentinel-1 processing workflow for dynamic open water classification. Bold numbers in parentheses indicate key datasets and processing steps referred to in the text.
In order to estimate the backscatter distribution expected for open water, an independent reference layer is required
The estimated
The left panel shows the DTM hillshade and derived potholes, the middle panel displays the Sentinel-1
In addition, the final water area should take the information from both VV and VH polarisations into account. The respective water probabilities, there is a the derived water area is connected with the original pothole area (taking all eight neighbours of each pixel into account).
For validation, producer’s and user’s accuracies were computed for the water class using the combined product against the reference data described in Sect.
Prairie wetlands can merge over time with neighbouring wetlands or split into separate waterbodies. Hence, monitoring the area of individual waterbodies is challenging. To track surface water dynamics across the study area, we computed the total water area, areas covered by individual waterbodies and the number of waterbodies for each of the observation dates. Furthermore, we were interested in the inter- and intra-annual dynamics of wetlands of different sizes. For this purpose, waterbodies were divided in four different size bins, and the contribution of each bin to the total water extent was tracked along the observation period. The metrics were computed using the “landscapemetrics” R package
The obtained
Map of prior water probabilities
Open water maps were produced for the months from May to October for the period from 2015 to 2020 (Fig.
Panels
Comparison with independent NAIP data, which was carried out for three dates, resulted in high producer’s accuracies (
Accuracy values of open water classification for three dates.
The accuracies reported here are similar to the ones reported by
Wind is an important environmental factor in the prairies and has been shown to affect SAR water extent retrieval by causing scattering from roughened water surfaces
Backscatter in
The effect of wind on SAR backscatter from water surfaces was a frequently encountered problem in this study, especially for VV-polarised imagery. VH data were shown to be less affected by this problem, which is in line with our expectations and findings available in the literature
Overall, the proposed approach for water extent classification represents an efficient way of fusing SAR backscatter with topographical information via the derivation of HAND. Such probabilistic approaches have drawn some attention in recent years to combine SAR imagery with multi-temporal and ancillary sources of information
Throughout the study period, the number of potholes with open water surfaces strongly varied between ca. 2300 and more than 5000. The total water surface area varied between ca. 14 000 ha and more than 18 000 ha (Fig.
We first report and discuss seasonality in the derived observed water area, number of waterbodies and average area per waterbody. Subsequently, inter-annual differences are reported and discussed in the context of ancillary hydrometeorological information.
In most years, the number of waterbodies and total water area displayed a seasonal behaviour: the highest numbers are observed in spring, there is an annual minimum in late summer, and the numbers increase again during September and October (Fig.
In general, the temporal patterns found in these variables followed the expected seasonality. In a modelling study,
Summed areas of waterbodies in size classes
In general, the number of waterbodies, the total water area and the median area per waterbody were relatively stable between 2015 and 2019, whereas 2020 differed considerably from the rest of the observation period in terms of these three metrics (Fig.
The strong increase in water area and the emergence of small wetlands in October 2019 with respect to an earlier year can also be observed in the water extent maps in Fig.
Water extent dynamics in
To our knowledge, this study is the first time that surface water dynamics in the PPR have been monitored using Sentinel-1 over an extended period covering both dry and wet years. This enables us to also support findings from previous studies
Inter-annual changes in prairie wetland extent have been tracked using Landsat data
In this study, a novel approach for retrieving dynamic open water extent in prairie pothole wetlands from dual-polarised Sentinel-1 SAR data was presented. Using a Bayesian framework, topographic information was integrated in the retrieval processed via HAND. The results demonstrate that the approach was successful in mapping changes in water extent in prairie potholes when their location was known a priori. The inclusion of topographic information, at least in some cases, helped to mitigate the adverse effects of non-water areas resembling water surfaces due to low backscatter and of wind roughening the water surface. The impact of the latter factor was further decreased by the combination of co-polarised and cross-polarised SAR data, as the latter are typically less affected by surface roughness. The obtained time series of total water area, number of open waterbodies and median wetland area covering a time period of 6 years showed clear intra-annual as well as inter-annual patterns. The different responses of small (
NAIP false-colour composites (R – near-infrared, G – red and B – green) acquired in
Maps of posterior probabilities,
The code used in this paper can be made available upon request from the first author.
The data used in this study were retrieved from the following sources: Sentinel-1 data are available through the Copernicus Open Access Hub (
StS designed the study, developed the Sentinel-1 classification algorithm, processed the data and analysed the results. MC extensively contributed to the development of the classification algorithm, the structure of the paper and the discussion of the results. WD extensively contributed to the structure of the paper and the discussion. SP contributed to the data processing and to the discussion of the results. StS prepared the paper with contributions from all co-authors.
The contact author has declared that neither they nor their co-authors have any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
The authors would like to thank the following organisations for providing the data used in this study: Copernicus (Sentinel-1 data), the USGS (runoff data), NOAA NCDC (precipitation data), the National Agriculture Imagery Program administered by the USDA Farm Service Agency (aerial imagery), the North Dakota State Water Commission (topographic data) and the USDA NASS (Cropland Data Layer). The authors acknowledge the helpful comments from Geoff Pegram and three anonymous reviewers. The authors further acknowledge the TU Wien Bibliothek for financial support through its Open Access Funding Programme.
The publication cost was covered by the TU Wien Bibliothek through its Open Access Funding Programme.
This paper was edited by Zhongbo Yu and reviewed by Geoff Pegram and three anonymous referees.