Discussion
Understanding vegetation water requirements and losses are important to
inform environmental water management and underpin equitable water
sharing plans. Given the significant advances in digital technology and
high costs of in-situ monitoring, new innovative cost-effective
methods are vital to monitor large land tracts both in Australia and
other regions across the world (Manfreda et al. , 2018). Woody
vegetation ET can provide a line of evidence to improve monitoring and
inform water management, however downscaling of low spatial resolution
data is required to provide robust remotely sensed ET estimates. The
performance of the RFall predictor model presented
within, indicates that a model has been developed that can accurately
predict FTCC for both sparse and densely vegetated areas semi-arid and
likely arid, floodplain environments.
As mentioned previously, while other fractional vegetation products are
available (Guerschman et al. , 2015; Guerschman and Hill, 2018;
DEE, 2019) the classification and spatial resolution of these did not
suit the purpose of improving remotely sensed ET outputs. Guerschman and
Hill (2018), for example, provide landscape fractional cover including
percent photosynthetic vegetation, non-photosynthetic vegetation and
bare soil across 250 m MODIS pixels. In contrast, the model presented
here, provides FTCC in 10% increments of canopy cover related only to
trees.
Important outcomes of method
development
LiDAR imagery collected from the regions of interest (Yanga and Barmah
National Parks) proved invaluable to the development of the reported
method. LiDAR provides a proxy for field derived canopy cover, against
which Sentinel data was trained. As the LiDAR output is composed of
‘point clouds’ representing 3-D land surface features, it was possible
to separate trees over 2 m in height from other surface features, to
provide ‘field-based’ canopy cover. The results, from a remote sensing
perspective, are also important to understand critical bands that are
required to monitor vegetation and water to inform future satellite
development.
Additional method
application
While remote sensing methods can be used to derive FTCC such as aerial
imagery (Melville et al. , 2019), LiDAR (Wasser et al. ,
2013) and fine resolution satellite imagery like WorldView2 and 3
(Immitzer et al. , 2018), acquiring imagery is costly and requires
‘tasking’ (i.e. imagery it is not collected regularly and needs to be
ordered) for specific areas of interest. As a result, national scale
imagery is not available and temporal availability is poor. In
comparison, developing a method using open-access Sentinel-1 and -2
imagery, provides a mechanism to monitor vegetation cover change from
2015 and into the future at desired intervals such as monthly,
seasonally or annually, depending on the application.
The FTCC method, is however, likely to be very valuable to other areas
of catchment water management. The significant bushfires across southern
Australia over the summer of 2019/2020 are likely to have significant
future impacts on water resources and especially changes to water yield
in both quality and quantity over the next decade (Brown, 1972; Lee,
2020; Moreno et al. , 2020). The FTCC method would enable accurate
estimates of tree area, pre and post bushfires, to underpin future
hydrological catchment yield forecasting. Current methods are unlikely
to be suitable to disentangle woody tree vegetation, which is a dominant
water user, from other vegetation sources. This may lead to errors in
water yield estimation pre and post fires. Tree reduction also increases
streamflow locally, although this is quickly reversed as regeneration
occurs, particularly in Australia with bush tolerant native species
(Kuczera, 1987; Brookhouse et al. , 2013). There is an opportunity
to link broadscale FTCC predictions with modelling of water fluxes
through Land Surface Models, enabling modelling to understand the
effects of fires (or any land cover changes) on hydrologic fluxes
(Barlage and Zeng, 2004; Fang et al. , 2018). As severe bushfires
have also featured in other areas around the world such as the United
States and Europe, the method is relevant internationally.
Sources of error
Sentinel data was trained against LiDAR which was collected between
2009-2015. The very high correlation between LiDAR FTCC and predicted
FTCC, provides some confidence that although time has passed,
substantial changes to vegetation crowns were not apparent in the
trained areas. As vegetation might have changed slightly between the
training data (LiDAR) and the covariates (multi-spectral and SAR bands),
part of the error is actually not attributable to FTCC modelling, i.e.
the discrepancy between the LiDAR FTCC and the predicted FTCC represent
a maximum error margin. Yanga LiDAR, collected in 2009, occurred before
the break in the Millennium Drought from 1997 to 2009 (Leblanc et
al. , 2012), while Barmah LiDAR was collected after (2015). This might
explain the poorer prediction at Yanga using the RFYangamodel as substantial improvement in tree canopy crowns occurred over the
2010-2012 flood period (Doody et al. , 2015) leading to a
discrepancy between amount of crown cover pre and post flood. The match
at Barmah was likely higher due to closer match between imagery dates
(2015 LiDAR and 2016 Sentinel). While the date gap between Yanga imagery
is not ideal, it was suitable for this project, however more recent
LiDAR imagery is preferred. Additional sources of error could have been
introduced to the model from use of two different LiDAR collecting
platforms and differences in their acquisition altitudes as well as the
spatial mismatch between 20 m training data and 10 m LiDAR data.
Further research
While the initial method shows considerable promise for widespread
application and identification of FTCC, to scale the method across the
MDB, additional areas will need to be trained to incorporate vegetation
in different climate and especially rainfall zones. It is unclear if
regions with higher rainfall will fit the RFall model,
so further investigation is required. The objective moving forward, is
to provide a universal model to predict FTCC across the MDB and examine
reducing FTCC resolution further to <10 m. Building further
upon that, will be investigation of the feasibility of producing FTCC
timeseries over the period of Sentinel availability (~5
years), focusing on seasonal and annual predictions which will be
valuable for monitoring of temporal vegetation canopy cover change at a
fine resolution. As mentioned in relation to bushfire and water yield
research applications, provision of <10 m woody tree canopy
cover would substantially improve vegetation water use estimates based
on tree area and aid forecasts of how water yield and hydrologic fluxes
(ET, recharge and runoff) will change into the future.