Figure 1. Study area indicating investigated reaches A and B.
Base map reprinted from ArcGIS Online maps under a CC BY license, with
permission from Esri, original Copyright © 2018 Esri (Basemaps supported
by Esri, DigitalGlobe, GeoEye, Earthstar Geographics, CNES/Airbus Ds,
USDA, AEX, Getmapping, and the GIS User Community).
Data and Methods
Satellite Video
Satellite video was acquired over our study area on 5 February 2022 at
23:12 UTC by a Jilin-1 GF-03 sensor, part of the Jilin-1 constellation
operated and developed by Chang Guang Satellite Technology Company. The
Jilin-1 satellite video has a spatial resolution of 1.22 m and was
acquired at native frame rate of 5 Hz for a duration of 28 seconds. To
counter sensor platform movement as well as scene ‘morphing’ due to the
changing view angle of the satellite overpass, we stabilized the video
using FIJI’s TrakEM2 plugin (Cardona, 2006; Cardona et al., 2012).
TrakEM2 relies on a Scale Invariant Feature Transform (SIFT) algorithm
to align image stacks based on common features. To avoid geometric
distortions and since frames from the video were acquired at a similar
resolution, we utilized an affine transform to register our image
stacks. Quantitative metrics detailing the minimum, maximum and mean
displacement errors related with image stabilization are reported.
Large-scale Particle Image Velocimetry
LSPIV, based on Eulerian principles of motion (Euler, 2008), was
originally introduced by Fujita et al ., (1998), enabling the
estimation of instantaneous flow velocities from a series of consecutive
images. LSPIV velocities were computed using PIVlab (Thielicke and
Sonntag, 2021; Thielicke and Stamhuis, 2014) developed in MATLAB
(R2022b, Mathworks, Natick, MA, USA). Computation of surface flow
velocities in PIVlab is attained by cross-correlation algorithms applied
to orthorectified images recorded at a known time interval, δt .
Here, we evaluate the accuracy of both Fast Fourier Transform window
deformation (direct FFT correlation with multiple passes and deforming
windows) and Ensemble correlation (Figure 2). The multi-pass FFT window
deformation approach allows for the spatial resolution of velocity
measurements to be improved through multiple refinements of
interrogation areas. Interrogation areas (IA), which are small windows
of defined size (in pixels), are used to track the displacement of image
patterns within a chosen larger search area (SA) in subsequent images.
Ensemble correlation is better suited for sparsely seeded images as it
relies on averaging correlation matrices followed by detecting a
correlation peak with the resultant benefit of lower bias and
displacement errors. Lewis et al . (2018) and Muste et al .
(2008) provide comprehensive detail on the theory and application of
LSPIV in riverine environments.
We cropped the video to reduce computational cost while focusing on two
cloud-free and straight river reaches A and B (Figure 1). Individual
frames were extracted from the cropped video at frame rates of 1, 0.5
and 0.25 Hz. Image pre-processing was performed to amplify the
visibility of surface tracers with respect to the background
(riverbanks/static ground), applying a Contrast-limited adaptive
histogram equalization (CLAHE) filter to enhanced image contrast (Li and
Yan, 2022; Masafu et al ., 2022). Distinct features on the water
surface were difficult to discern in the raw images, which would be
expected in natural rivers and given the height of the optical sensor.
However, CLAHE contrast enhancement enabled the tracking of seeding
surrogates in the image sequences, which occur when specular reflection
formed by incident light interacts with free-surface deformations on the
river. Image intensity variations associated with these surface
deformations were visible in our post-processed images.
Sensitivity to Image Frame Rate and PIV algorithm
The primary free parameters in LSPIV are the sampling frequency (frame
extraction rate), interrogation (IA) and search (SA) areas, and optimal
configurations vary significantly (Kim et al., 2008; Legleiter and
Kinzel, 2020; Sharif, 2022). Earlier studies (Tauro et al ., 2018;
Zhu and Lipeme Kouyi, 2019) have demonstrated that the IA should be
small enough to eliminate spurious velocities whilst being large enough
to accommodate an adequate window for surface pattern tracking. Sampling
frequency (frame extraction rate) and the IA are closely coupled and
must be considered in tandem, with frame-to-frame displacement rates
influencing the accuracy of pattern/particle detection on images.
FFT window deformation and Ensemble correlation algorithms were utilized
with the maximum allowable number of PIV algorithm passes allowed within
PIVlab (four) for our sensitivity analysis (see Zhu and Lipeme Kouyi,
2019). We processed images using an IA of 128×128 pixels with successive
passes based on IA sizes of 64×64, 32×32 and 8×8 pixels, all with 50%
overlaps, corresponding to a minimum spatial distance of 9.8 m. For the
~70 m wide River Darling at Tilpa, this was sufficient
to allow the detection of displaced surface features. Whilst smaller IAs
would allow for higher-resolution vector maps, this would also
significantly increase noise and the amount of erroneous correlations.
We process two configurations based on FFT window deformation and
Ensemble correlation algorithms at three sampling rates (1, 0.5 and 0.25
Hz) resulting in 6 different LSPIV runs for each scenario. This resulted
in image sequences consisting of 28, 14 and 7 frames which enabled us to
experiment with varied frame extraction rates for image-based velocity
analysis. Following LSPIV cross-correlation, we post-processed the
resultant velocity fields to filter out spurious velocities.
Specifically, we utilized filters that removed velocity vectors that
differed by 8 x (PIVLab’s default threshold) the standard deviation from
the mean velocity and further applied a local median filter threshold of
3 x 3 pixels to remove outliers. Velocity vectors were georeferenced
within PIVlab from an image coordinate system back into a projected
coordinate reference system (GDA 1994 MGA Zone 55). We used the
coordinates of the same distinct features as those used in PIVLab to
assess the accuracy of our georeferencing against actual locations,
based on 1 m Maxar satellite imagery in ArcGIS Pro.
Validation of PIV velocity vectors
We use a calibrated 2D hydraulic model to evaluate the accuracy of LSPIV
velocities. 2D models offer particular value as they can map velocities
in diverse hydraulic conditions rather than at a few idealized sections,
including locations where the range and resolution of traditional
equipment (such as aDcps and current meters) is limited or where the
deployment of velocity sensors can be complex, time-consuming, and
hazardous, particularly during flood events when flow depths and
velocities prevent field deployment. We detail our model calibration
process in the supplementary material (section 1).
Discharge estimation using LSPIV velocities
The velocity-area method was used to calculate discharge (Q )
(Turnipseed and Sauer, 2010). Channel depth and velocity are integrated
from discrete locations along a channel’s width. Discharges estimated at
each vertical sections spanning the channel width are summed to total
discharge (Q ) (Cohn et al ., 2013).
\begin{equation}
Q=\sum_{i=1}^{m}{A_{i}v_{i}}=\sum_{i=1}^{m}{b_{i}d_{i}v_{i}}\nonumber \\
\end{equation}where m = number of verticals across channel;Ai = cross-sectional area of vertical i? ;bi = width of vertical i =
(x i+1 – x i-1)/2 withx = horizontal distance of vertical from the edge of water;di = average depth of vertical i ; andvi = average downstream velocity in verticali . We define a minimum of 25 vertical subsections at each
cross-section, with sub-sectional area extending half the distance to
the preceding and following measurements.