Effect of spatial resolution of satellite images on estimating the greenness and evapotranspiration of urban green spaces

Urban green spaces (UGS), like most managed land covers, are getting progressively affected by water scarcity and drought. Preserving, restoring and expanding UGS require sustainable management of green and blue water resources to fulfil evapotranspiration (ET) demand for green plant cover. The heterogeneity of UGS with high variation in their microclimates and irrigation practices builds up the complexity of ET estimation. In oversized UGS, areas too large to be measured with in situ ET methods, remote sensing (RS) approaches of ET measurement have the potential to estimate the actual ET. Often in situ approaches are not feasible or too expensive. We studied the effects of spatial resolution using different satellite images, with high‐, medium‐ and coarse‐spatial resolutions, on the greenness and ET of UGS using Vegetation Indices (VIs) and VI‐based ET, over a 780‐ha urban park in Adelaide, Australia. We validated ET with the ground‐based ET method of Soil Water Balance. Three sets of imagery from WorldView2, Landsat and MODIS, and three VIs including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI) and Enhanced Vegetation Index 2 (EVI2), were used to assess long‐term changes of VIs and ET calculated from the different imagery acquired for this study (2011–2018). We found high correspondence between ET‐MODIS and ET‐Landsat (R2 > 0.99 for all VIs). Landsat‐VIs captured the seasonal changes of greenness better than MODIS‐VIs. We used artificial neural network (ANN) to relate the RS‐ET and ground data, and ET‐MODIS (EVI2) showed the highest correlation (R2 = 0.95 and MSE =0.01 for validation). We found a strong relationship between RS‐ET and in situ measurements, even though it was not explicable by simple regressions; black box models helped us to explore their correlation. The methodology used in this research makes a strong case for the value of remote sensing in estimating and managing ET of green spaces in water‐limited cities.

and MSE =0.01 for validation). We found a strong relationship between RS-ET and in situ measurements, even though it was not explicable by simple regressions; black box models helped us to explore their correlation. The methodology used in this research makes a strong case for the value of remote sensing in estimating and managing ET of green spaces in water-limited cities.

| INTRODUCTION
Newly emerged and still evolving concepts of "greening cities" and "moving towards carbon neutrality" are getting more attention in recent times.
The benefits of urban green spaces (UGS) and their significance to public health and wellbeing are getting more recognized (Elgström, Erman, Klindworth, & Koenig, 2014;Li, Sutton, Anderson, & Nouri, 2017;Li, Sutton, & Nouri, 2018). The sustainable green city concept based on SDG-Goal 11 of "sustainable cities and communities" should not be limited to a narrow scope of only expanding UGS; sustainability scope must be expanded to the urban water system to avoid a possible conflict between "water-saving" and "greening cities" goals.
UGS are crucial elements in urban ecohydrology (linkage of water cycles and urban landscape) and their ecosystem services, which include their potential to mitigate climate change impacts and make cities more inhabitable and sustainable (Costanza & Liu, 2014;Wagner & Zalewski, 2009). Evaluation and enhancing their services require a better understanding of their metrics and indicators such as their extent and greenness, evapotranspiration (ET) and water use. These metrics help ecohydrological models to quantify the ecosystem services provided by UGS (Revelli & Porporato, 2018).
Water consumption, known as water lost to the atmosphere, or ET, takes a vital role in urban water management focused on UGS, particularly in dry regions where green plant covers are more affected by water scarcity and drought. In arid and semi-arid regions, UGS are mostly irrigated, thus making UGS a significant competitor for other water-demand sectors (Evans & Sadler, 2008;Nouri, Chavoshi Borujeni, & Hoekstra, 2019).
Greenness, shading and ET play key roles in the cooling effect of the green spaces, and in mitigating the urban heat island (UHI) effects, and reduction of energy consumption. These roles mean that maintaining and expanding green plant covers of trees, shrubs and grasses require more water resources (Maheng, Ducton, Lauwaet, Zevenbergen, & Pathirana, 2019;Qiu et al., 2017;Skelhorn, Lindley, & Levermore, 2018).
To balance urban greening and blue water saving, the water footprint of urban green coverand in particular their ET, needs to be quantified.
There are numerous studies on the water use of UGS (Hilaire et al., 2008;Mini, Hogue, & Pincetl, 2014;Ouyang, Wentz, Ruddell, & Harlan, 2014). The majority of these studies refer to water withdrawal or water application/use of UGS rather than water consumption that is measured by ET. Total water consumption includes both blue water (irrigation water from surface and groundwater resources) and green water (rainwater) (Nouri et al., 2019;Velpuri & Senay, 2017). The available literature on UGS ET estimation is mainly limited to small scales, for example, backyard gardens, due to the complication of making ET measurements in heterogeneous urban vegetation. For instance, an in situ method of soil water balance was established in an urban park in Adelaide -Adelaide Parklandsand was compared with three observational factor-based methods of Water Use Classifications of Landscape Species (WUCOLS), Irrigated Public Open Space program (IPOS), and Plant Factor (Nouri et al., 2016). Two factorbased approaches of Landscape Irrigation Management Program (LIMP) and WUCOLS were compared in two green spaces, a botanic garden and an urban forest in Isfahan, Iran (Shojaei, Gheysari, Nouri, Myers, & Esmaeili, 2018). Two methods of WUCOLS and portable chambers were used to measure ET from eight irrigated turfgrass regions in the Los Angeles Metropolitan area (Litvak, Bijoor, & Pataki, 2014;Litvak & Pataki, 2016). The outcomes of this later study were compared with an approach that calculates total ET by adding ET of both grasses and trees using empirical models and remote sensing maps of land covers (Litvak, Manago, Hogue, & Pataki, 2017). By this approach, a considerable portion of green shrubbery was not included in the total ET of these irrigated sites. This work did illustrate the necessity for proper spatial scale when assessing the ET of oversized UGS (i.e., UGS that are too large and complex to be studied by in situ approaches of ET estimation).
In the last two decades, remote sensing (RS) has introduced innovative approaches to measuring ET in large-scale regions, but it was primarily applied to agricultural systems, and to a lesser extent in riparian and lowland/upland forests, yet remote sensing-based ET has rarely been applied to UGS (Bastiaanssen, Menenti, Feddes, & Holtslag, 1998;Huete et al., 2002;Nagler et al., 2005;Su, 2002). Nouri, Beecham, Anderson, Hassanli, and Kazemi (2015) reviewed the potential of optical remote sensing to facilitate ET measurements in heterogeneous urban vegetation and later they employed remote sensing-based ET estimation methods in a small urban park ($10 ha) using high-resolution images of WorldView 2 & WorldView 3 Nouri, Beecham, Anderson, & Nagler, 2014).
Land cover mapping from free satellite imagery, such as MODIS, has been practised for decades (García et al., 2013;Yebra, Van Dijk, Leuning, Huete, & Guerschman, 2013). While free of charge, this class of synoptic sensors is usually too coarse to accurately map UGS (Gómez, White, & Wulder, 2016;Grekousis, Mountrakis, & Kavouras, 2015;Xian, Shi, Dewitz, & Wu, 2019). The high-frequency and high-quality MODIS data and VI time series (1-2 days) make it a promising tool to study vegetation change (greenness changes over time) in UGS, including their ET. Similarly, free images of Landsat with a higher spatial resolution (compared to MODIS) has the potential of supporting a more accurate estimation of green covers. Both of these advantages need to be tested in the context of an oversized urban landscape.
A new class of high-resolution, space-borne sensors, such as WorldView, IKONOS, Quickbird, and GeoEye with sub-meter resolution, can resolve biophysical parameters of the heterogeneous urban environment and bring a new dimension to the mapping of UGS.
However, the cost of these images, their limited availability and footprint and associated processing make them not suitable for mapping of oversized UGS in cities.
In all remote sensing-based efforts, the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), and improved version of the Enhanced Vegetation Index 2 (EVI2), from high, medium and coarse resolution images, showed promising results in monitoring greenness and ET of large green covers in non-urban landscapes (Glenn, Neale, Hunsaker, & Nagler, 2011;Groeneveld, Baugh, Sanderson, & Cooper, 2007;Jarchow et al., 2020;Nagler et al., 2005). This research investigated the value of estimating greenness (VIs) and ET of oversized UGS using freely available remote sensing imagery. In this study, we evaluated the relationship between VIs This research study is timely as cities continue to explore carbonneutral management strategies. Accurate estimation of the ET of large UGS is particularly important for Adelaide and other water-limited cities that have increasingly suffered from aridity and drought. These cities are rapidly expanding and are in pressing need for more proactive and sustainable water management strategies.

| STUDY AREA
The city of Adelaide is surrounded by green infrastructure known as the Adelaide Parklands (called "Parklands" in this study) consisting of 29 parks covering approximately 780 ha ( Figure 1). The Adelaide Parklands is one of the largest urban parks in the world; it is almost three times larger than London's Hyde Park and is called "The heart and lungs of Adelaide". Over 510 plant taxa, about 40% native and 60% introduced species, are reported to exist in the Parklands. This diverse and distinctive landscape provides urban habitat for more than 150 species of birds, 33 species of mammal, 18 species of reptiles and six species of amphibian. This internationally unique green belt provides a precious social, environmental and recreational resource. Due to the expanse and heterogeneity of species, microclimate, plant density, and accessibility, we hypothesize that remote sensing data and tools would be a useful, practical for studying the water demand of this green belt.  Table 1.
Worldview2 (8 bands, spatial resolution of 0.46 m panchromatic and 1.85 m multispectral); WorldView2 images were pre-processed (orthorectified and atmospherically corrected), NDVI from band 5 and 7 were calculated, and the Modifiable Areal Unit Problem (MAUP) effects on vegetation pixels in NDVI calculation were tested . Bands 5 (red) and 7, near-infrared1 (NIR1), of WV2 were selected as the most accurate combination to calculate the ET of UGS (Nouri et al., 2014). Earth Explorer (https://earthexplorer.usgs.gov/). Each scene consisted of individual GeoTIFF bands files (surfaces reflectance, VIs and pixel quality information). All scenes were then subset to a common geographic region of interest, and cloudy and poor quality pixels were removed based on the per-pixel QA information. Due to gaps in the data resulting from the Scan Line Corrector issue in Landsat 7 ETM+ (Storey, Scaramuzza, & Schmidt, 2005), we simply ignored the missing data and averaged whatever was present within the scene.
Aqua MODIS (with 36 bands, spatial resolution of 250 m for the red and NIR bands, 500 m for the remaining land bands, and 1 km for all other bands); MODIS Aqua data from tile h29v12 (collection 6) for the period 2011-2018 were downloaded from the LP DAAC (https:// lpdaac.usgs.gov/) data pool.
All images were re-projected to a common geographic lat/lon grid using the WGS84 datum. 03.04.2019). The variables included precipitation, minimum temperature, maximum temperature, minimum relative humidity, maximum relative humidity, average wind speed and solar radiation.

NDVI, a normalized ratio of the NIR and red bands [Equation
The EVI was designed to complement NDVI over denser canopies with more sensitivity as a result of the de-coupling of the canopy background signal and a reduction in atmosphere influences (Huete et al., 2002. Similar to NDVI, EVI uses a ratio of NIR, red, and blue bands [Equation (2)].

| Vegetation index scaling
To address scene-to-scene variability resulting from atmosphere correction and/or sun and viewing geometry, we removed NDVI outliers at the high and low ends, by calculating the max/min NDVI values for each scene using barren soil or verdant (green vegetation) pixels, typically from agricultural fields (Groeneveld et al., 2007). We identified the barren soil area within the scene and averaged all NDVI pixels to generate the NDVI min . Similarly, using verdant agricultural fields in the scene, we determined the absolute highest NDVI value; we then averaged all pixels within 95% of that absolute value to generate NDVI max .

| ET computation
We estimated the greenness using NDVI, EVI, and EVI2 and then calculated ET by Equations (5) and (6) The ET 0 refers to the reference ET defined as "the rate of evapotranspiration from a hypothetical reference crop with an assumed crop height of 0.12 m, a fixed surface resistance of 70 sec m -1 and an albedo of 0.23, closely resembling the ET from an extensive surface of green grass of uniform height, actively growing, well-watered, and completely shading the ground" (Irmak & Haman, 2003). Daily ET 0 data were acquired from the Kent-Town weather station in Adelaide.
The long-term trend and correlation between RS-ET including ET-

| Ground data
The ground data were collected using the soil water balance (SWB) method to quantify the main components of the water budget equation . ET was calculated from the mass balance of water inflows and outflows over the study period using in situ measurements from the experimental site; a full description of the methodology and data are available from Nouri et al. (2016).
Water balance in the root zone, for a given time period, is calculated by Equation (7).
where P is precipitation, ET is evapotranspiration, IR is irrigation, DR is drainage, RO is runoff and δS is the change in soil moisture status.
To capture the heterogeneity of landscape plant species, density and microclimate within the Parklands, two layers of data for soil salinityusing proximal sensing by electromagnetic -EM38and landscape coverusing airborne images to differentiate turfgrasses from trees and shrubswere overlapped and four sampling zones were defined. These four zones included 1low salinity (<1.2 dS/m)primarily covered with turf grasses with few trees and shrubs, 2high salinity (>1.2 dS/m)primarily covered with turf grasses with few trees and shrubs, 3low salinity (<1.2 dS/m)mostly trees and shrubs with intermittent turf grasses, and 4high salinity (> 1.2 dS/m)mostly trees and shrubs with intermittent turf grasses. Samples for drainage and soil moisture status were collected from these four zones, and the average was assumed a representative for the full site.
Meteorological data were acquired from two stations, an in situ wireless weather station in the Parkland and the closest station of the

| Comparison of RS-ET and ground data
To relate RS-ET dataset to in situ measurement, both, linear and nonlinear, approaches of modelling were studied and compared. In model- The performances of scenarios were evaluated by conventional statistical criteria of goodness of fit, including mean squared error (MSE), and the coefficient of determination (R 2 ) using Equations (8) and (9). showed inter-and intra-annual variations (Figures 2 and 3

| Comparison of RS-ET
The trends of ET from time-series data produced using EVI, EVI2 and ET from Landsat and MODIS were lowest in June, July, and August, the winter in Adelaide, and highest in late November, December, January, and sometimes early February, the summer in Adelaide ( Figures 5 and 6). Comparing ET from Landsat, ET-Landsat (NDVI) was consistently highest, and ET-Landsat (EVI2) was lowest. These trends had substantial agreements with observational methods of ET estimation in the Parklands  and our ground data. Irrigation of the Parklands usually stops in winter; water resource for the Parklands in winter is limited to green water.

| Validation of RS-ET against SWB
RS-ET derived from Landsat (NDVI, EVI, and EVI2) and MODIS (EVI and EVI2) were compared against ground data of SWB from December 2011 to November 2012. Monthly RS-ET rates were calculated and evaluated in comparison with SWB measurements, as presented in Figure 7. RS-ET and SWB showed similar trends throughout the study period; the highest ET occurred in summer, and lowest ET in winter, as expected.
Since the regression coefficients of daily and monthly RS-ET against SWB data was not strong (R 2 $ 0.41), the black box models F I G U R E 2 Seasonal trends of NDVI, EVI, and EVI2 from Landsat during 2011-2018. We note the high similarity between EVI and EVI2.  Figure 8 shows the correlation as well as the equation between actual F I G U R E 3 Seasonal trends of NDVI, EVI, and EVI2 from MODIS during 2011-2018. All indices are highly similar as with Landsat, however, due to the mixing resulting from the larger MODIS pixel size the data is noisier, especially EVI and EVI2. All indices show higher values in the Winter/Spring months (J-O) (X-axis) and predicted (Y-axis) values in both training (left) and validation (right) steps for the selected models.
The ANN models show that SWB agrees with ET-Landsat (NDVI),

ET-Landsat (EVI), ET-Landsat (EVI2), ET-MODIS (EVI) and ET-MODIS
(EVI2    In the Parklands, the maximum urban heat island (UHI) reduction in temperature was recorded at 1.5 C in the daytime and 0.5 C in the nighttime (Guan et al., 2013;Guan et al., 2016), which is a much lower reduction in temperature than would be provided by optimum greenness of the Parklands which would benefit Adelaide. A possible solution is to have green plant cover that is comprised of droughttolerant native species which would have lower water demand, in addition to applying the principles of water sensitive urban design (WSUD) (Chavoshi, Pezzaniti, Myers, & Sharma, 2017;Sharma et al., 2013) to free up more blue water from the limited available amount required to maintain (and possibly expand) the green cover of the Parklands.

| CONCLUSION
Accurate estimation of ET and water demand/water consumption for large UGS that are too large for ET monitoring from the ground (as well as areas that are inaccessible and not cost-effective) remains critical, especially in arid and semi-arid urban environments. Measuring ET remotely will help decision-makers and water managers formulate strategies and plans for sustainable green cities across the globe.
In this study, three satellites, WV2, Landsat, and MODIS were used to assess the greenness and ET of a large park, the Adelaide Parklands, a 780-ha public green space within a dry state, South Australia. Different satellite-based VIs were analysed, including those from WV2-NDVI, Landsat-NDVI, Landsat-EVI, Landsat-EVI2, MODIS-NDVI, MODIS-EVI, and MODIS-EVI2. We then compared these RS-ET with the in situ method of SWB using ANN machine learning techniques which offered a better understanding of the interactions across the parameters that define ET than regular regression analysis. Our findings point to a series of key conclusions and suggestions: • Public and free-access satellite images can cost-effectively assist in characterizing the greenness and ET of large UGS.
• EVI2, a 2-band alternative to EVI, performed well across the three sensors, indicating its robustness and demonstrating across sensors EVI continuity • RS-based ET performed very well and proved to be an accurate long-term monitoring tool for greenness and ET trends over large UGS. Remotely sensed estimates of ET are timely, cost-F I G U R E 9 NDVI changes for two consecutive years from 2011 to 2018. The first 5 years use Landsat 7 ETM+ data and the remaining 2 years use Landsat 8 OLI effective, and minimize reliance on ground-based methods. Our use of the black box models helped tease out ground-truth validation.
• While 30 m is still coarse for an urban setting, Landsat outperformed MODIS (250 m) by capturing the finer-scale and seasonal variation of VIs across the complex Parklands. The 250 m MODIS footprint was too coarse to deal with adjacency of a mixed landscape.
• The composite and divergent phenology signals resulting from trees and grasses confounded the seasonal variation in the VI profiles, implying that scale and resolution are critical in these heterogeneous urban landscapes.
• Our ET derived data captured the differences in dry, hot summers and mild, wet winters in Adelaide. This finding was captured by all of the RS-ET methods as well as the ground data estimation techniques.
Our work, based on multi-sensor remote sensing data fusion and analytical methods in an urban landscape demonstrated that these techniques are transferrable to other water-limited cities, particularly urban areas in arid and semi-arid climates, such as those in North Africa, Southern Europe, the Middle East, Southwest USA, Northern China, and Southern India. Managing water use with these novel techniques can help identify responses to the sustainable city green space management.

ACKNOWLEDGEMENTS
This article is part of the special issue in honour of Prof. Edward P. Glenn. This paper is dedicated to Edward P. Glenn to express our sincere and deep gratitude for being such an inspiration. Edward P. Glenn was the referee of my PhD thesis, and later during my postdoc, I had the chance to work with him remotely and during my visit to the U. S. Geological Survey (USGS) and University of Arizona. It was truly a great privilege and honour to work with him and learn from him. He is sincerely missed. We thank the University of South Australia to support the in situ measurements in the Adelaide Parklands; without their help, this project would not have been possible. We are also extremely grateful to Mr. Hamed Noori, the graphic designer, for his great assistance with the graphic abstract.
We wish to thank Dr. Jon Michael Hathaway for his constructive review. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

DATA AVAILABILITY STATEMENT
Data available on request due to privacy/ethical restrictions.