Seeing the floods through the trees: Using adaptive shortwave infrared thresholds to map inundation under wooded wetlands

Accurate information about the extent, frequency and duration of forest inundation is required to inform ecological, biophysical and hydrological models and enables floodplain managers to quantify the efficacy of flood mitigation/modification activities. Open water classifiers derived from optical remote sensing typically underestimate or fail to detect floodplain forest inundation. This paper presents a new method for detecting forest inundation dynamics using freely available Landsat and Sentinel 2 data, referred to as short‐wave infrared mapping under vegetation. The method uses a dynamic threshold that accounts for the additional shortwave infrared reflectance caused by the presence of tree canopies over floodwater. The method is demonstrated at five Ramsar listed River Red Gum floodplain forest wetlands in southeastern Australia. Accuracy assessment based on independent drone imagery from a wide range of vegetated wetlands showed an absolute accuracy of 67%–70% and a fuzzy accuracy of 81%–83%. We found the method is conservative, and underestimates inundation (16%–18%) but very rarely misclassifies dry pixels as inundated (0.3%–0.6%). When compared to river gauge data, the method shows similar trends to an open water classifier (i.e., the area of inundated vegetation increases with increasing river height). The method is conservative compared to lidar‐based floodplain inundation models but can be applied wherever cloud‐free scenes of Landsat or Sentinel 2 have been acquired, thereby enabling floodplain managers with the ability to quantify changes in inundation dynamics in places/time‐periods where lidar is unavailable.


| INTRODUCTION
Located at the interface between terrestrial and aquatic ecosystems, floodplain forests, mangroves, and floodplain wetlands provide a wide array of crucial ecosystem functions and services (Brown et al., 1997;Hughes, 1988;Hughes & Rood, 2003).Floodplain forests are a source of carbon and nutrients to in-stream ecosystems (Bunn et al., 2003;Robertson et al., 1999); provide a site for carbon sequestration (Sutfin et al., 2016); provide shade and large woody debris to in-stream ecosystems (Brown et al., 1997;Mac Nally et al., 2001, 2011); stabilize river banks (Abernethy & Rutherfurd, 1998;Brown et al., 1997); and serve as essential habitat for both terrestrial and aquatic species, such as waterbirds and fish (Brown et al., 1997;Hughes, 1988;Mac Nally et al., 2001), and provide refugia during periods of hydrological and climatic extremes (Fink & Scheidegger, 2021;Sedell et al., 1990).
In southeastern Australia, these ecosystem functions are provided by the predominant floodplain forest species the River Red Gum (Eucalyptus camaldulensis Dehnh.), and the importance of these River Red Gum forests is reflected in the selection of multiple River Red Gum forests for listing under the Ramsar international convention on wetlands (Finlayson, 2013).As a floodplain forest species River Red Gums are dependent on inundation from both a vegetation condition and reproductive point of view (Catelotti et al., 2015;Dexter et al., 1986;Doody et al., 2014;Steinfeld & Kingsford, 2013).
However, river flow regulation has altered the inundation dynamics in the majority of River Red Gum forests (Bren, 1988;Doody et al., 2014).
Accurate spatial characterization of the frequency, extent and timing of floodplain inundation is required to quantify carbon budgets, infer riparian condition, and quantify the lateral and longitudinal connectivity of riparian and floodplain forests (Bren et al., 1988;Horner et al., 2009;Hughes & Rood, 2003;Roberts & Marston, 2011;Robertson et al., 1999).However, the spatial and temporal variability of canopy density in floodplain forests makes the mapping of inundation extent beneath the canopy extremely challenging (Huang et al., 2018;Smith, 1997;Zhiqiang et al., 2016).
Existing methods for mapping inundation extent underneath vegetation fall into three categories: (1) synthetic aperture radar (Hess et al., 1995); (2) synthetic aperture radar/optical hybrid methods (Ward et al., 2013(Ward et al., , 2014)); and (3) hydrodynamic inundation models (typically based on lidar digital elevation models) (Zhiqiang et al., 2016).Whilst these approaches are effective for the areas and epochs for which the relevant datasets are available, the high cost of lidar and the limited temporal coverage of systematic SAR acquisitions limit the spatio-temporal applicability of such methods.
Medium resolution optical satellites such as Landsat and Sentinel 2 have been extensively used to map open water (Mueller et al., 2016;Pekel et al., 2016;Ticehurst et al., 2022).However open water classification algorithms are not designed to detect pixels that contain mixtures of water and other targets for example, floodplain forest canopies.Wetlands have been characterized using optical satellite imagery previously, using both coarse (Bian et al., 2017;Deng et al., 2014;Kordelas et al., 2019;Li et al., 2013) and moderate resolution sensors (Ji et al., 2009;Wolski et al., 2017).However, such methods are typically based around an empirical approach or require subjective threshold adjustment for individual scenes.To address these limitations, this paper proposes a new approach that leverages the properties of open water and River Red Gum forests in the shortwave infrared wavelengths (SWIR), specifically the SWIR centred around 2100 nm which is strongly absorbed by water (Wolski et al., 2017).To remove the subjectivity of threshold selection, flooded imagery is used to identify the range of SWIR values occupied by flooded vegetation in contrast to adjacent dry areas, and the seasonal influences of canopy shading are accounted for resulting in an algorithm that can be applied across all available images.Whilst vegetation canopy reflectance in the SWIR is a function of canopy moisture, SWIR reflectance of vegetation canopies is higher than that of open water (Huete, 2012).The conceptual model behind the approach is presented in Figure 1.
The aim of this paper is to present a new method of mapping the extent of inundation underneath tree canopies.We achieve this by:

| Algorithm conceptual basis
The fundamental concept that underpins the method presented in this paper is that persistent overstory/emergent vegetation reflects more strongly in the SWIR than open water, and that if the density of the overstory/emergent vegetation is known then it is possible to include that additional SWIR reflectance into a threshold-based approach for detecting inundation.This method uses images captured during the rising stages of a river gauge to identify areas that are completely inundated (all SWIR reflectance is assumed to be from the overstory/emergent vegetation) and not-inundated (SWIR reflectance is assumed to be a combination of overstory/emergent vegetation and the ground surface).The SWIR thresholds that best distinguish these two targets are used as thresholds within the method.
The SWIR reflectance of canopies varies with solar angle, with higher solar zenith angles resulting in increasing SWIR reflectance because a greater portion of the canopy is sunlit (Schaaf & Strahler, 1993).SWIR thresholds are determined for rising river stages for both summer and winter, and the threshold adjusted between these two extremes as a function of solar zenith angle.Summer and winter are used as they represent the extremes for solar angle, which influences the SWIR reflectance of tree canopies due to higher amounts of self-shading at low sun angles.
The influence of open water on a pixel's SWIR reflectance becomes increasingly diminished as canopy closure increases and becomes impossible to distinguish at high canopy cover levels.This extinction point is taken as 80% canopy cover in this method.
Low SWIR reflectance in isolation is not diagnostic of open water, with targets such as terrain shadow, burn scars and dark bare soils also having low SWIR reflectance.This method uses a floodplain mask to exclude this source of potential false positives.

| Platform
The analysis was undertaken on the Digital Earth Australia platform (Dhu et al., 2017) using analysis ready data (Dwyer et al., 2018) that adheres to the CEOS Analysis Ready Data for land specification (Siqueira et al., 2019).

| Satellite imagery
The Landsat (Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI and Landsat 9 OLI) and Sentinel 2 (Sentinel 2a and 2b) data used in this analysis were corrected for nadir view angle, bidirectional reflectance distribution function, atmospheric and terrain effects (NBART) using the methods described in Li et al. (2012).The solar zenith angle values used in the NBART algorithm were also used to correct for canopy self-shading effects.Cloud and cloud shadow affected pixels in the Landsat data were masked using the fmask algorithm (Zhu et al., 2015), and Sen2Cloudless algorithm for the Sentinel 2 data (https://docs.sentinel-hub.com/api/latest/user-guides/cloud-masks/).
The 20 m SWIR 2100 nm bands from Sentinel 2 were resampled to 30 m to enable multi-sensor results.
A fractional cover algorithm (Scarth et al., 2010) was used to calculate the green, bare and non-photosynthetic cover fractions for all Landsat scenes.The fractions were summarized on a percentile basis for each calendar year, with the 10th percentile of green cover (pv10) assumed to represent vegetation cover that had persisted throughout the year.

| Ancillary data
The maximum floodplain extent was taken as the extent of floodplains contained within the Australian National Aquatic Ecosystems for the Murray Darling Basin version 3.0 (Brooks, 2021).This step was taken to mask out features within the agricultural landscapes adjacent to the Ramsar wetlands that may be conflated with inundation patterns within the wetlands.
River gauge data from the Bureau of Meteorology's Water Data online portal http://www.bom.gov.au/waterdata/ were used to assess F I G U R E 2 Ramsar listed floodplain forest wetlands in southeastern Australia used in this study.The numbers refer to sites described in the text (background imagery from Bing Maps).
the relationship between river level and the extent of inundation within the five Ramsar sites.

| Open water classifier (water observations from space)
This paper uses the water observations from space (WOfS) (Mueller et al., 2016) as an open water classification algorithm to identify areas of open water.(Figure 3).

| Water underneath vegetation-short wave infrared mapping under vegetation
Our new short-wave infrared inundation mapping under vegetation (SWIM-UV) method is based on a dynamic SWIR threshold, where the threshold increases as a function of the amount of persistent vegetation cover observed at that pixel.To establish the values for that threshold, imagery was selected with reference to a river gauge for that floodplain forest.Satellite images from the rising limb of the hydrograph and high river stage (Figure 4) were used to identify areas of floodplain forest that were either dry or inundated.This process was repeated for all the Ramsar listed floodplain forests used in this study (Figure 2) to ensure that the range of canopy cover encountered within floodplain forests was captured.
To account for the influence of solar angle on SWIR reflectance the threshold selection process was repeated for both summer and winter rising limb/high water stage satellite images (Figure 5).The threshold was varied between the summer maximum values and the winter minimum values as a function of solar angle (Equations 1 and 2).Let β represent the portion of sunlit crown and α represent the seasonal SWIM-UV threshold adjustment factor.
This term was used to adjust both the amount of SWIR reflectance from open water and the maximum SWIR contribution from the tree canopies.This was done to represent the reduction in the amount of sunlight reaching the water surface under forests at low sun angles, and the reduction in SWIR reflectance due to increased canopy selfshading at low solar angles.(Table 1).
The SWIM UV threshold is calculated as follows (with values from 3. Thresholds were scaled to each observation using the α term calculated using Equation ( 2).
The SWIM-UV algorithm workflow is described in Figure 6.It consists of the following steps for each time step: 1. Identify 'open water' pixels using the WOfS algorithm 2. Calculate the 'water under vegetation' threshold for that date as a function of the 10th percentile of green fractional cover (pv10) for that year and solar angle for that day (SWIM-UV threshold) 3. Identify pixels where the vegetation cover was too dense for the presence/absence of water to have an influence on the SWIR reflectance (set as a constant of 80% green fraction) 4. Generate the map of inundation extent for that time step.
To evaluate the accuracy of the SWIM-UV algorithm, we used labelled drone footage from 20 sites across Australia that covered the range of water-vegetation mixture types, including floating, emergent and over-story mixture gradients.

| Validation data
The drone imagery used in this paper was captured as part of a broader campaign to evaluate remote sensing inundation algorithms across a wide range of vegetation types.Consequently, the drone The dates and locations used in this study are summarized in Table 2.In some cases, there was no cloud-free imagery available • Overstorey (trees above surface water), • Emergent (vegetation over the surface water), • Floating (vegetation floating on the surface water), • OpenWater (no vegetation, just surface water), • GreenVeg (green vegetation on dry land), • DryVeg (dry vegetation on dry land), • Bare (bare ground), and • NA (no data).
The 3 | RESULTS

| Accuracy assessment
The A fuzzy accuracy assessment (Foody, 1996) method is used to compare the categorical (wet/dry) classification results with the percentage inundated results from the labelled drone footage.The 'absolutely correct' percentage represents the sum of both dark green cells divided by the number of samples for that sensor.The same method is used to calculate the fuzzy correct (pale green), fuzzy incorrect (pink) and absolutely incorrect (red) values.The results also show that the SWIM-UV classifier is a conservative classifier with SWIM-UV = wet correctly identifying pixels dominated by water (<50 water present) at a rate of 92.3% (Sentinel 2) and 89.1% (Landsat) but will misclassify completely wet pixels as dry in 9% of cases for Sentinel 2 and 8% of cases for Landsat.
Looking at 'total incorrect' as a combination of 'fuzzy' and 'absolutely' incorrect in Table 4 shows error rates between 18% (Sentinel 2) and 16% (Landsat).It is important to note that these errors are on the conservative side, that is SWIM-UV fails to detect pixels with water present between 18% (Sentinel 2) and 16% (Landsat) of the time, but only classified predominantly dry pixels as 'wet' 0.6% (Sentinel 2) and 0.8% (Landsat) of the time (red/pink cell values in the top rows for each sensor (Table 4) divided by total number of samples for that sensor).Put simply, the SWIM-UV classifier will underestimate overall inundation extent (16%-18%) but will very rarely misclassify dry areas as water (0.6%-0.8%).
The similarity in accuracy between Landsat and Sentinel 2 data allows the results from both sensors to be combined into a single multi-sensor SWIM-UV classification result, thereby providing an inundation mapping method from 1987 onwards, with more temporally rich inundation information from 2016 onwards.

| Daily inundation extent
The daily inundation extent results are shown in the bottom row of In the low river level image, some limited inundation under vegetation south of the Murray River can be observed.In the rising river level image, more extensive inundation under vegetation can be observed, and in the high river level image inundation can be seen extending out all the way to the levee system in the northeast.F I G U R E 7 Landsat and Sentinel 2 pixels labelled using drone imagery, with the short wave infrared mapping under vegetation summer threshold.The dot colours represent the different levels of percentages of water and aquatic vegetation present within that pixel.WO = water observations from space.SWIR short-wave infrared wavelengths.
T A B L E 3 Accuracy assessment of labelled drone pixels.

| Annual inundation summaries
The daily inundation maps can be combined across a calendar year T A B L E 4 Fuzzy accuracy assessment summary (cell colours refer to Table 3).

| Comparing inundation extent and river stage heights
Whilst the comparison with drone imagery (Figure 7) provides a direct validation of algorithm performance, a comparison of algorithm results with river gauge data provides a qualitative assessment of algorithm performance within a hydrological context.From first principles, we know that inundation extent should increase as some function of increasing river height, with the shape of that function determined by the morphology of the channel and the floodplain.Comparing the hydrographs (Table 5) with the extent of inundated vegetation  The SWIM-UV method provides a complement to proven open water classifiers such as WOfS, and in combination they enable the determination of overall inundated extent, and when combined with hydrograph data can be used to calculate timing and duration of inundation events.This information is critical to estimating floodplain productivity (McInerney et al., 2023;Wohl & Knox, 2022) and for river management, such as estimating blackwater or hypoxia risks arising from inundated organic matter on floodplains (Gibbs et al., 2022).
Inundation plays an essential role in maintaining and provisioning key habitat for a range of biota, such as macro-invertebrates (Benke, 2001;Jenkins & Boulton, 2003), for fish habitat and recruitment (Gibbs et al., 2023;Gorski et al., 2011).
Whilst SAR-based methods can also map water underneath vegetation (Hess et al., 1995), the SWIM-UV method can be applied to earlier epochs when global SAR missions such as Sentinel 1 are not available, thereby providing the ability to quantify changes to floodplain hydrology further back in time.
The SWIM-UV method advances the approach proposed by Wolski et al. (2017) by applying a SWIR based thresholding approach at a 30 m, rather than 500 m pixel scale, thereby providing end users with a far more detailed assessment of inundation extent, and resolving narrower inundation events on smaller river reaches.The SWIM-UV approach advances the method proposed by Ji et al. (2009) by providing an a-priori estimate of the vegetation fraction within the pixel and accounting for that vegetation fraction's additional SWIR reflectance.
Systematic, large scale wetland inventories rely upon (semi)automated methods for delineating polygons that define the extent of wetlands.Whilst open water classifiers such as WOfS can be used for this purpose for delineating lacustrine wetlands, systematic delineation of palustrine wetlands has historically been more challenging, relying on higher levels of manual digitisation and expert knowledge.
The SWIM-UV method may provide a solution to the long-standing challenge of delineating the extent of vegetated wetlands.

| Constraints
The single biggest constraint of the SWIM-UV method is its reliance on cloud free observations of the floodplain.For small, steep or rapidly draining catchments where the high cloud cover that accompanies flood events obscures the floodplain, this method will fail to detect floodwaters.The usefulness of this method is constrained to larger, low relief catchments where the floodwaters persist for days to weeks after the flood generating rainfall event(s) have passed.
The SWIM-UV method is demonstrated on River Red Gum forests.Leaf optical property datasets such as those described in Jacque- The SWIM-UV method relies on the availability of fractional cover percentiles to estimate the amount of persistent green vegetation such as overstory vegetation and persistent macrophyte beds.
Fractional cover percentiles are currently available for Australia and Africa, but would need to be calculated for other regions, or an alternate persistent vegetation cover surrogate would need to be developed and tested.The use of persistent green fraction is unlikely to be suitable for mapping inundation extents under the wetland vegetation types dominated by seasonally green dense emergent macrophyte beds (Phragmites australis, Typha sp.) or those dominated by nonphotosynthetic vegetation cover, such as lignum (Duma florulenta).
SWIM-UV uses a floodplain mask to mask out results from the surrounding landscape.This is particularly important in areas where shaded hillside forests, fire scars and/or permanently irrigated pastures occupy the same SWIR reflectance range as inundated vegetated wetlands.In applying this method to other locations, a terrainbased floodplain mask could be developed/applied to suppress spurious results from the surrounding landscape.
The SWIR thresholds used by SWIM-UV are intentionally conservative, meaning that the classifier will tend to err on the side of underestimation.This may limit its utility for applications that require delineations of maximum extent of inundation, particularly if the areas at the edges of inundation consist predominantly of mixed pixels.

1.
Identifying the SWIR reflectance of water in Ramsar listed River Red Gum forests in southeastern Australia 2. Estimating the additional SWIR reflectance due to persistent vegetation 3. Identifying inundated pixels by including the SWIR contribution of persistent vegetation when applying a SWIR threshold 4. Evaluating the accuracy of the method using UAV data 5. Comparing inundation extents from this method with: a. River gauge data b.Modelled flood plain inundation extent 2 | METHOD 2.1 | Australia's Ramsar listed floodplain forests The major floodplain forest wetlands used in this study (Figure 2) and protected by the Ramsar convention in Australia are listed below.Site specific descriptions from the Australian Government Department of Climate Change, Energy, the Environment and Water are contained within the links to each site: 1. Barmah forest (Barmah-Millewa) 2. Gunbower forest (Gunbower, Perricoota and Koondrook) 3. Macquarie marshes (Macquarie Marshes North) F I G U R E 1 Conceptual model re: short-wave infrared wavelengths (SWIR) reflectance, open water, and tree canopies.4. NSW Central Murray (Edwards Wakool) 5. Riverland (Chowilla floodplain)

F
I G U R E 4 False colour Landsat satellite imagery during low flows (left) and high flows (right).In the low flow image water (black) is only distinguishable in the main channel, whereas water can be seen across the floodplain in the high flow image.The red dots in the hydrograph (bottom) show water depth during the low-flow and high-flow images.footage includes sites with River Red Gum forests and a wide range of other wetlands that include overstory, floating and emergent vegetation.However, there was insufficient drone coverage of River Red Gum forests to use as the only source of validation.Therefore, we used all available drone imagery for sites that were not dominated by macrophyte vegetation to quantify the behaviour of SWIM-UV over a variety of water-vegetation mixture gradients.The majority of drone data was captured during six field campaigns to the following regions and with the following partners: 1. Stradbroke Island, Moreton Bay, Southeast QLD.In collaboration with Elders in Council, Uncle Norm Enoch (February 2020) 2. Kakadu National Park, NT.In collaboration with Dr Renee Bartolo and Traditional Owners Jackie Cahill and Natasha Nadji (October 2020) 3. Macquarie Marshes, Gwydir Wetlands, NSW.In collaboration with Dr Rachael Thomas and Traditional Owner Phil Duncan (December 2020)

F
I G U R E 5 Short-wave infrared wavelengths-fractional cover scatterplots (summer and winter) showing threshold positions (red line).The blue and orange dots represent inundated and dry pixels, respectively.The green line shows location of open water threshold.close to the date (<12 days) of the drone imagery, so these sites were not used in the comparison.Data were acquired at 15 different wetlands with a UAV, Phantom 4 DJI.The height of the flight was set between 80 and 100 m and the resolution of the photographs was of 12-megapixels.The pictures were joined and processed with DroneDeploy and converted to a GeoTiff image.Five additional drone mosaics captured by Renee Bartolo from the Office of the Supervising Scientist in 2013, 2017 and 2019 were included in the analysis.The high-resolution (1 m 2 pixel) drone imagery was classified into eight classes: Abbreviations: SWIR, short-wave infrared wavelengths.
accuracy assessment methodology follows the best practice guidelines set out in Olofsson et al. (2013) of using independent highresolution imagery to validate Earth observation algorithms.The accuracy of the SWIM-UV algorithm was quantified by comparing the per pixel 'percent water present' data with the per pixel SWIR reflectance from the corresponding satellite image with reference to the SWIM-UV SWIR threshold as shown in Figure 7 and Table 3.Any pixels that consisted solely of open water in the drone data (pure water pixels) were excluded from the accuracy assessment.Note the dark green (moderate resolution satellite pixels that consist of only drone labelled aquatic/over-story vegetation and open water) pixels fall predominantly under the SWIM-UV threshold.

Figure 8
Figure 8 for the following dates in 2022 (low river level: 11th March, rising river level: 2nd September, high river level: 11th December), the false colour images in the top row are shown by way of comparison.
and presented as a normalized count to describe the percentage of observations classified as 'inundated' by both the open water classifier (WOfS) and the inundated vegetation (SWIM-UV) classifier (Figure 9).The open water classifier clearly identifies the Moira Lake lacustrine wetland (bottom left) within the broader River Red Gum forests of the Barmah-Millewa Ramsar site, and shows that water was observed in Moira Lake consistently throughout 2016.The inundated vegetation classifier shows the far more extensive inundation that occurred throughout the Ramsar site.The percentage of time inundated for this class varies depending on floodplain topography.The two inundation classifiers can be combined to provide a combined view of inundation frequency as per the panel on the right.The annual inundation summaries can be produced for each calendar year and this provides an opportunity for a secondary level of product validation.Annual summaries of inundation can be compared with the river flows for the corresponding year as shown in Figure 10.The hydrograph for the River Murray immediately upstream of the Gunbower, Perricoota and Koondrook Ramsar site has 2 years highlighted.The high flow year of 2016 is highlighted in purple, and the lower flow year of 2019 is highlighted in yellow.The inundation frequency maps for those 2 years reflect the hydrograph, with extensive overbank inundation observed in 2016, and predominantly inchannel inundation observed in 2019.It is worth noting the discrepancy between stage height (high) and observed inundation (low) that occurs in 2011 through 2012.This represents a period when there was only one Landsat satellite operating, which reduces observation frequency, and therefore reduces the chances of observing short lived high stage events.
I G U R E 8 Per time-slice short wave infrared mapping under vegetation (SWIM-UV) results for low-mid-high stage flows in the Gunbower, Perricoota and Koondrook forest.The top row shows the Landsat false colour image, and bottom row shows SWIM-UV extent.One of the advantages of annual summaries is that it allows you to compare and identify changes to patterns of inundation between two wet years as shown in Figure 11.This figure shows the wet years of 1991 and 2016 for the Chowilla Ramsar site, areas wetter in 1991 are shown in blue, whereas areas wetter in 2016 are shown in red.Changes in flow regulation to the sequence of lakes along the north west have resulted in starkly different inundation patterns between the 2 years.

(
SWIM-UV) and the extent of open water (WOfS) at a daily timestep as shown in Figure 12 provides the opportunity to compare the behaviour of both the open water classifier and the inundated vegetation classifier over multiple decades and all six sensors (Landsat 5, Landsat 7, Landsat 8, Landsat 9, Sentinel 2a and Sentinel 2b).There are no discernible differences in the results between sensors (minimum extent of open water and inundated vegetation track consistently across the entire time series) and the overall relationship between river height and inundation extent is similar for both the open water and inundated vegetation classifier.Note that the nonlinear nature of these cross plots is to be expected given that inundation extent is a non-linear function of stage height depending on the topography of the floodplain (Bates et al., 2003).The majority of results are 'limited inundation extent' and 'low flow' as represented by the yellow-green blob centred around the 2-metre stage height.The extent of open water and inundated vegetation both start to rise as the stage height exceeds 5 m indicating that this is the 'commence to overbank flow' height for this floodplain.It is worth noting the maximum values on the y axis for both of the area plots: 400 km 2 for open water and 700 km 2 for inundated vegetation.This highlights the value of the SWIM-UV method for detecting inundation under vegetation, as the open water classifier in isolation would dramatically under-estimate the true extent of floodplain inundation.

Figure 13
Figure 13 shows the relationship between stage height and both the open water classifier and inundated vegetation classifier (SWIM-UV) for the other four Ramsar listed wetlands in southeastern Australia.The Macquarie Marshes Northern Ramsar site shows limited open water extent (less than 25 km 2 except for one outlier) and far more extensive inundated vegetation (more than 100 km 2 ).In the Barmah-Millewa Ramsar site both open water inundation and inundation under trees increase steadily within increasing river height, but the extent of inundated vegetation ($700 km 2 ) is far greater than that of open water ($300 km 2 +), the same is true for the Edward-Wakool with inundated vegetation ($200 km 2 ) compared to open water ($100 km 2 ).For the Chowilla floodplain (the most downstream floodplain), this pattern is reversed, with more open water ($250 km 2 ) compared with inundated vegetation (50 km 2 ).
moud and Ustin(2019)  show that leaves reflect more strongly than open water in the SWIR.Applying SWIM-UV to other wooded wetlands such as mangroves would require the user to capture the range of SWIR values observed when the wooded wetlands were completely inundated, and the SWIR reflectance from adjacent dry areas, to ensure that the SWIR threshold included the influence of that canopy type on SWIR reflectance.

Table 1
representing the summer and winter extremes for the threshold): 1.For pixels with pv10 values <15% use the open water SWIR threshold 2. For pixels with pv10 values >15% calculate dynamic threshold as 'open water SWIR threshold + 15 Â pv10 value' until the minimum dry pixel value is reached Drone image sites used for validation of Landsat and Sentinel 2 data.
Note that while the open water (WOfS) classifier identifies open water within the channel and near the limits of the floodplain in the high river level image, the T A B L E 2