Summary
The detection and monitoring of surface water and its extent are critical for understanding floodwater hazards. Flooding and undermining caused by surface water flow can result in damage to critical infrastructure and changes in ecosystems. Along major transportation corridors, such as railways, even small bodies of water can pose significant hazards resulting in eroded or washed out tracks. In this study, heterogeneous data from synthetic aperture radar (SAR) satellite missions, optical satellite-based imagery and airborne light detection and ranging (LiDAR) were fused for surface water detection. Each dataset was independently classified for surface water and then fused classification models of the three datasets were created. A multi-level decision tree was developed to create an optimal water mask by minimizing the differences between models originating from single datasets. Results show a water classification uncertainty of 4–9% using the final fused models compared to 17–23% uncertainty using single polarization SAR. Of note is the use of a high resolution LiDAR digital elevation model (DEM) to remove shadow and layover effects in the SAR observations, which reduces overestimation of surface water with growing vegetation. Overall, the results highlight the advantages of fusing multiple heterogeneous remote sensing techniques to detect surface water in a predominantly natural landscape.
Methodology
The study area is located at the Queen’s University Biological Station (QUBS) in Kingston, Ontario, Canada, approximately 50 km north of the eastern end of Lake Ontario. QUBS was chosen due to the abundance of water bodies that vary in size from small inundated forests to large open bodies of water. The 2 km × 2 km square area where the SAR, LiDAR and optical data sets overlap is shown in Figure 1.
Five TSX staring spotlight mode scenes were acquired over QUBS from April to September 2016. These scenes represent single look slant range (SSC) products of descending path, single polarization (HH), right looking with an incidence angle of 44°. They have a slant range and azimuth resolution of up to 0.6 m and 0.24 m, respectively. The LiDAR data were acquired using an Optech Gemini ALTM, which is a small-footprint, single wavelength, discrete return system. The survey was flown in June 2015 at an altitude of 1200 m, and resulted in a spatial resolution of 1 pt/m2. WorldView-2 imagery was acquired in August 2016 consisting of 8-band multispectral optical images with a resolution of 2 m. Table 1 outlines the data sets used in this study.
In situ field investigations were performed on three different occasions, which overlap with three of the five SAR acquisition dates. During these field investigations to five chosen study locations, GPS coordinates were recorded as well as water and site conditions, including the presence of open water, flooded vegetation, and the dominant vegetation type such as reeds, shrubs or forest. The five sites were chosen for having different environmental and surface water conditions as follows:
A. Poole Lake: A large, open body of water bordered by wetlands and mixed forest.
B. Marsh A: A dense cattail marsh connected to a small lake, with the potential to be flooded during parts of the year. A small stream runs along the periphery of the marsh.
C. Inundated Forest: A small area of observed inundation beneath a mixed forest canopy with some emergent shrubbery. Depth of water in April was approximately 1 m.
D. Marsh B: A small pond bordered by a sparse cattail marsh, which immediately backs onto a flat field to one side and a forest to the other side. The forested side is beyond the chosen study area. Water level was observed to be highest in the spring and receded throughout summer.
E. Vegetated Lake: A wetland composed of a central pond/marsh, transitioning into a dense cattail marsh along the periphery. In the central pond/marsh, there are sparse cattails and the remnant trucks of dead trees. A central stream cuts through the marsh, connected to peripheral streams throughout the marsh.