Riparian vegetation NDVI dynamics and its relationship with climate, surface water and groundwater

Riparian vegetation NDVI dynamics and its relationship with climate, surface water and groundwater

Journal of Arid Environments 113 (2015) 59e68 Contents lists available at ScienceDirect Journal of Arid Environments journal homepage: www.elsevier...

2MB Sizes 2 Downloads 39 Views

Journal of Arid Environments 113 (2015) 59e68

Contents lists available at ScienceDirect

Journal of Arid Environments journal homepage: www.elsevier.com/locate/jaridenv

Riparian vegetation NDVI dynamics and its relationship with climate, surface water and groundwater Baihua Fu*, Isabela Burgher Fenner School of Environment and Society, Australian National University, Canberra, ACT 0200, Australia

a r t i c l e i n f o

a b s t r a c t

Article history: Received 9 July 2013 Received in revised form 2 June 2014 Accepted 24 September 2014 Available online

Maintaining the integrity of riparian ecosystems whilst continuing to reserve and extract water for other purposes necessitates a greater understanding of relationships between riparian vegetation and water availability. The Normalised Difference Vegetation Index (NDVI) is a good indicator for identifying longterm changes in vegetated areas and their condition. In this study, we use regression tree analysis to investigate long term NDVI data (23 years) at semi-arid riparian areas in the Namoi catchment, Australia. Climatic factors (temperature and rainfall), surface water (flow and flooding) and groundwater levels are analysed collectively. We find that in general maximum temperature is the variable that primarily splits NDVI values, followed by antecedent 28-day rainfall and then inter-flood dry period and groundwater levels. More rain is required in the warmer months compared to cooler months to achieve similar mean NDVI values in tree patches or areas of high NDVI in riparian zones, presumably because of higher evaporation. Inter-flood dry period is shown to be important for maintenance of NDVI levels, particularly when rainfall is limited. Shallower groundwater levels sustain the NDVI and hence vegetation greenness when conditions are cooler and wetter. © 2014 Elsevier Ltd. All rights reserved.

Keywords: Remote sensing Environmental flow Regression tree analysis Groundwater NDVI

1. Introduction Dams, surface water extraction and groundwater pumping for human uses have contributed to serious changes in the functioning of riparian ecosystems e areas which directly adjoin and influence inland water bodies (Allan and Castillo, 2007; Nilsson and Berggren, 2000; Poff et al., 1997). This is especially so for riparian ecosystems in arid and semi-arid regions where water is scarcer yet in high demand for human use, resulting in greater extraction of surface water and groundwater resources (Stromberg et al., 1996). Maintaining the integrity of riparian ecosystems whilst continuing to reserve and extract water for other purposes necessitates a greater understanding of relationships between riparian vegetation and water availability (Cunningham et al., 2011). As water resources that we have some level of control over, this is particularly true of river flow and groundwater, an understanding of which can help inform what surface water and groundwater regimes are needed to maintain the integrity of riparian ecosystems. Vegetation dynamics, especially over large scales, can be monitored using remote sensing (Barbosa et al., 2006; Gaughan et al., 2012; McGrath et al., 2012; Wang et al., 2003). Of the

* Corresponding author. E-mail address: [email protected] (B. Fu). http://dx.doi.org/10.1016/j.jaridenv.2014.09.010 0140-1963/© 2014 Elsevier Ltd. All rights reserved.

spectral indices derived from remote sensing which identify vegetated areas and their condition, the Normalised Difference Vegetation Index (NDVI) is still the most well-known and frequently used (Bulcock and Jewitt, 2010; Sims and Colloff, 2012). NDVI is based on the differential reflectance that plants exhibit for different parts of the solar radiation spectrum. Healthy green leaves strongly absorb photosynthetically active radiation for energy in photosynthesis, whereas internal mesophyll structures in the leaf scatter radiation in the near-infrared region to prevent overheating of the plant (Bulcock and Jewitt, 2010; Wang et al., 2003). Calculated by obtaining the difference between the remotely sensed visible (red) and near-infrared bands and normalising it over the sum of the two, NDVI is a good indicator of the ability of vegetation to absorb photosynthetically active radiation and therefore of landcover which comprises unstressed vegetation (Otto et al., 2011; Wang et al., 2003). NDVI values have been correlated with a number of vegetation structures and functions such as biomass (Cho et al., 2007; Hansen and Schjoerring, 2003), primary productivity (Goward and Dye, 1987), and Leaf Area Index (Bulcock and Jewitt, 2010). The importance of water availability to vegetation vigour has seen NDVI commonly used to investigate relationships between terrestrial vegetation and climate (e.g. Barbosa et al., 2006; Gaughan et al., 2012; Ji and Peters, 2003; McGrath et al., 2012).

60

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

The linear relationship between vegetation NDVI and antecedent rainfall in arid to semi-arid regions has been well documented (Groenvald and Baugh, 2007; Ji and Peters, 2003; Peng et al., 2012; Wang et al., 2003). Less documented are NDVI responses to hydrological factors. Marked NDVI responses to flooding events in large areas of naturally vegetated floodplain have been observed in eastern inland Australia (Parsons and Thoms, 2013; Sims and Colloff, 2012). Relationships have also been found for NDVI and changes in groundwater levels (Aguilar et al., 2012) and groundwater flow discharge (Petus et al., 2012). However, there is little literature that investigates the impact of climatic, surface water and groundwater factors collectively on NDVI for riparian vegetation. In this paper, we used 23 years of NDVI to examine semi-arid riparian vegetation responses to rainfall, flow, flood and groundwater across nine sections of the Namoi River Catchment in eastern Australia. We specifically looked at NDVI values of tree patches (dominated by river red gums) and riparian vegetation zones with denser and sparser trees. 2. Study area The Namoi River Catchment forms part of the Murray-Darling Basin and drains an area of approximately 42,000 km2 in northern New South Wales, Australia (Fig. 1). From east to west across the catchment, mean maximum temperatures in January range from 27 to 35  C, while mean maximum temperatures in July range from 12 to 17  C (Australian Bureau of Meteorology, 2013b). Rainfall generally decreases across the catchment from east to west, with annual averages of 945 mm rainfall at Niangala near the headwaters, 620 mm at Gunnedah in the midsection of the catchment and

480 mm at Walgett in the low lying plains of the west (Australian Bureau of Meteorology, 2013b). Potential evaporation increases across the catchment from east to west, with very high potential evaporation in the summer months compared to winter months. Mean daily evaporation rate during 1981e2010 at Gunnedah is 8.2 mm in January, compared to 2.0 mm in June (Australian Bureau of Meteorology, 2013b). This study focuses on the mid to lower sections of the Namoi River Catchment, downstream of Gunnedah. Annual flows generally increase with catchment area but catchment flows in the Namoi decrease downstream of Gunnedah due to increased evaporation, transmission losses and water use. The mid to lower Namoi has a long history of river regulation and flow regime has been significantly altered (Sheldon et al., 2000). Namoi also has the highest groundwater use in the Murray-Darling Basin. In 2004e2005, groundwater extraction in the Namoi was estimated to be 255 GL, accounting for 15.2% of the total groundwater use in the Murray-Darling Basin (CSIRO, 2007). About 35% of the groundwater extraction in the Namoi River Catchment was from the Lower Namoi Alluvium Groundwater Management Unit (CSIRO, 2007). The major streams and rivers of the catchment are dominated by river oak (Casuarina cunninghamiana) and river red gum (Eucalyptus camaldulensis). Native floodplain vegetation communities include open grassy woodlands dominated by poplar box (Eucalyptus populnea), black box (Eucalyptus largiflorens) and coolibah (Eucalyptus coolabah), and native grasslands dominated by plains grass (Austrostipa aristiglumis). Large areas of riverine land in the Namoi River Catchment have been converted to cropping and pastoral uses, thus except for habitat corridors and patches of

Fig. 1. Namoi River Catchment, showing the nine river sections we examined (i.e. assets), and the broader catchment area for which Landsat derived NDVI values were obtained. Here the NDVI values across the catchment area are shown for two years e 1991 a wet year and 2002 a dry year in the Namoi.

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

riverine vegetation most of the native vegetation has been cleared (Eco Logical, 2009). Within the study area, we specifically looked at dynamics between NDVI and water availability along nine sections of river (Fig. 1). These were ecological assets (specific locations of ecological interest) previously selected by another study researching watering needs for the development of broader environmental flow guidelines (Barma Water Resources et al., 2012). All assets except Maules Creek - which is considered to be in relatively natural condition have a history of river regulation. Additionally, large areas of these assets are classified as groundwater-dependent ecosystems by the Australian Groundwater Dependent Ecosystems Atlas (Australian Bureau of Meteorology, 2013a).

3. Methods 3.1. Calculating NDVI Existing riverine vegetation in the Namoi River Catchment often occurs in a fairly narrow fashion along water courses (see Fig. 2). We thus used remote sensing data acquired by the Landsat 5 TM and Landsat 7 ETM satellites as their 30 m resolution can target such a fine spatial configuration. This avoids the inclusion of potentially irrigated crops which would have been almost impossible with commonly used larger resolution imagery such as MODIS or AVHRR (with resolutions of 250 m or greater). Another important advantage of using Landsat imagery is the early start time of Landsat 5 which has been operating since 1984, allowing for analysis of a relatively long time-series of data. In contrast, 250 m resolution data from MODIS are only available from the year 2000. We acquired Landsat images (path 91/row 81) from Geoscience Australia (1987e1998) and the United States Geological Survey (1999e2010) comprising a data time-series of 23 years. Ideally, Landsat offers images on a periodic 16-day basis. However, image availability from the data distribution agencies and obscuring of assets by periodic cloud cover means that the number of usable data points for the years examined is inconsistent. We extracted a total of 228 usable images, with most years having between 8 and 15 images, although the years 1987, 1995, 1998, 1999, 2005 and 2008 have 5 images or fewer. Using ENVI 5.0 (Exelis Visual Information Solutions, 2012), images were corrected for sensor defects and sensor differences by converting to top of atmosphere reflectance using published postlaunch gains and offsets (Chander et al., 2009). Dark object subtraction was then performed using the band minimum from each image. Dark object subtraction was chosen because this method is one of the simplest yet most widely used methods of atmospheric correction for land-use classification and change-detection

61

purposes (Song et al., 2001). Following this correction, NDVI was calculated for all images as per Eq. (1) using bands 3 and 4 in Landsat which have been calibrated to sense radiation in the visible (Red) and near-infrared (NIR) regions of the spectrum respectively.

NDVI ¼ ðNIR  RedÞ=ðNIR þ RedÞ

(1)

NDVI values range between 1.0 and 1.0 with values nearing zero and below indicating features which are not vegetated such as water, snow, ice, clouds and barren surfaces. We estimated NDVI values for two different types of riparian areas: general riparian vegetation zones which consist of riparian trees and grasses, and selected tree patches within the riparian vegetation zones which are predominantly comprised of river red gums. For the tree patches, polygons were manually drawn around selected areas of forest adjacent to river channels within assets using Google Earth (Fig. 2). Mean tree patch NDVI across each asset was calculated for each date. For riparian vegetation zones, the river channels within assets were buffered by 200 m on each side and areas corresponding with crops clipped out (Fig. 2). Then, in each asset the area of each NDVI class: [e1, 0), [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, 1] was calculated. This area was then standardised for each asset by calculating the proportional area of each class in each asset. We used NDVI classes rather than mean NDVI values for riparian vegetation zones because the variations of NDVI are too high due to large areas and different riparian vegetation types. Mean NDVI values cannot fully reflect the NDVI values in the riparian vegetation zones in each asset. In contrast, variations of NDVI in tree patches are very small and mean NDVI values are good indicators of NDVI values in each asset. 3.2. Rainfall and hydrological data Daily rainfall records were obtained from the Australian Bureau of Meteorology website, using the most proximate station to each asset which had complete or near complete data records (Australian Bureau of Meteorology, 2013b). In the vicinity of Duncan's Junction, Wee Waa to Bugilbone and Bugilbone to Walgett, only one rainfall station had sufficient data, and this station is used for all three assets. Daily surface flow data before 2008 were extracted from PINNEENA 9.2 (NSW Department of Water and Energy, 2008), with more recent data taken from the NSW government water information website (waterinfo.nsw.gov.au). River gauges were selected based on their proximity to the assets. Historical groundwater bore data were obtained from Groundwater PINNEENA 3.2 (NSW Office of Water, 2011). The groundwater bore data were interpolated into daily time series using a linear regression (Blakers, 2011). Bores were selected based on their proximity to the asset, and the mean value of the daily groundwater

Fig. 2. Section of Wee Waa to Bugilbone as viewed by Google Earth showing selected tree patch polygons (left) and LandSat derived NDVI values, showing tree patch and 200 m river buffer with crops clipped out (right). NDVI values tend to increase near the river channel due to presence of riparian vegetation.

62

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

levels at Pipe 1 (the pipe with the shallowest opening) of these bores was used to represent the groundwater level for that asset. 3.3. Exploratory analyses Three types of analyses were undertaken to investigate NDVI dynamics of tree patches and riparian vegetation zones: 1) long term trend of NDVI in the past 23 years; 2) seasonal behaviour of NDVI; and 3) relationships between NDVI and climate and hydrological variables using classification. Annual average NDVI tree patch and class area percentage values for each asset were used to identify long-term trends and variability over the 23 year period. Average monthly NDVI values over the 23 year period were used to assess seasonality over the year. Regression tree analysis was used for classification. Regression tree analysis is a commonly used statistical method for non-parametric regression and classification that has been widely applied for ecological data (De'Ath and Fabricius, 2000). It generates a regression tree that classifies a response variable based on explanatory variables; in doing so thresholds can be identified to best separate values of the response variable. We used the ctree() function in the party package in R (Hothorn et al., 2006) for regression tree analysis. This method tests the global null hypothesis of independence between any of the explanatory variables and the response variable. It selects the explanatory variable with the strongest association to the response variable. This association is measured by a p-value corresponding to the hypothesis test (we used KruskaleWallis test). Then the algorithm implements an optimal binary split in the selected explanatory variable using a permutation test. The classification stops when the p-value is greater than a threshold (we used 0.01). The algorithms can be found in Hothorn et al. (2006). The explanatory variables used for regression tree analysis include daily maximum temperature (Temp_max), antecedent rainfall (Rainfall_total), antecedent flow (Flow_total), groundwater level (Groundwater), inter-flood dry period (indicated by days since last flood, Days_lastflood), the proportion of flow relative to the overbank flood threshold (Perc_low), and the mean Perc_low over the past days (Perclow_mean). We used Perc_low to account for variation in channel capacity across assets, thus a Perc_low value of 1 indicates that the flow is at the overbank flood threshold regardless of the channel size at different assets. We used antecedent rainfall, antecedent flow and Perclow_mean to account for potential lag response of NDVI to rainfall and flow. The lag was identified by calculating 1e60 days (at a 1 day interval) of total rainfall, total flow and mean percentage low, then estimating correlations between NDVI and these antecedent values. The lags that

correspond to the highest correlation were used to calculate Rainfall_total, Flow_total and Perclow_mean. These explanatory variables were related to three sets of response variables for regression tree analysis: 1) mean NDVI values of the tree patches at all assets; 2) proportion of area that has high NDVI (i.e. NDVI > 0.6) in each of the three riparian vegetation zones that have dense trees (i.e. assets 5e7: Duncan's Junction, Wee Waa to Bugilbone, Bugilbone to Walgett); and 3) proportion of area that has high NDVI (i.e. NDVI > 0.6) in each of the two riparian zones that have sparse trees (i.e. assets 3 and 4: Upstream Mollee and Mollee to Gunidgera). Calculation of these NDVI was described in Section 3.1. We separated NDVI data for riparian zones with dense and sparse tree vegetation in order to reduce the impact of land clearing on areas of NDVI values. 4. Results 4.1. Long-term trends Annual rainfall averaged across assets shows high variability between years. The years 1994 and 2002 were comparatively very dry, whereas relatively large amounts of rain fell in the years 1991, 1998, 2004 and 2010 (FigS1, electronic version only). Annual flows are mostly less than 300 GL/year in most years across the assets. High annual flows are found for periods during 1989e91, and the years 1998, 2000 and 2010. Mean groundwater levels are fairly constant across the 23 years, but have slightly deepened since 2003. Annually averaged NDVI values for tree patches are fairly stable for all assets over the 23 year period examined (Fig. 3). There is a clear difference in the NDVI ranges between assets, with Pian Ck showing a lower range (0.41e0.51) quite apart from the rest of the group. Gunnedah on the other hand shows the highest range (0.52e0.74). Eastern assets including Gunnedah, Maules Ck, and Barbers Lagoon have higher ranges of annual NDVI values and higher variability over the time period than do western assets (Pian Ck, Duncan's Junction and Wee Waa to Bugilbone). Trends of NDVI change for tree patches through the 23 years are similar across assets, with significant falls in NDVI in the years 1994, 2002 and 2007 for all assets (except for Gunnedah in 2007). These low NDVI periods broadly correspond to low rainfall periods in the region. Notable rises in NDVI occur in the 2003e04 period and 2008 across assets. Other periods exhibit mixed change in tree patch NDVI. For example, in 2010 western assets decreased in NDVI in while eastern assets increased for this year. Annual mean NDVI class areas within assets showed patterns largely reflecting those of annual mean tree patch NDVI, with 1994

Annual Mean Treepatch NDVI

0.8

Gunnedah Barbers Lagoon Mollee to Gunidgera Duncans Junction Wee Waa to Bugilbone Bugilbone to Walgett Pian Creek

0.7

0.6

0.5

Maules Creek

0.4

Upstream Mollee

0.3 1985

1990

1995

2000

2005

2010

Fig. 3. Annual mean NDVI of selected tree patches across assets from 1987 to 2010.

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

and 2002 showing markedly large areas of low NDVI (<0.4) and therefore reduced areas of higher NDVI across all assets (Fig. 4). Years where there were markedly large areas of high NDVI were inconsistent across assets, although the years 1990, 1996, 2003 and 2008 showed relatively large areas of high NDVI (>0.6) for many assets. 4.2. Seasonality We found strong seasonality over the years with very low NDVI (i.e. 1994, 2002 and 2007 e presumably dry periods) (FigS2, electronic version only). In these cases NDVI values tend to be highest in the middle months of the year (corresponding with the southern hemisphere winter) and lowest in summer. Seasonality over the years which have average or high levels of NDVI is less obvious. The NDVI values are slightly lower in summer months particularly November for most assets, but are slightly higher in March, April and May in the western assets, and August, September and October for the eastern assets. Similar seasonality patterns were found for NDVI class areas in riparian vegetation zones. 4.3. Regression tree analysis 4.3.1. Mean NDVI for tree patches Correlation analysis between NDVI and 1e60 days of antecedent rainfall and flow values suggested that 28-days (e.g. total rainfall of the current day and the previous 27 days) is the best lag time. For mean NDVI for tree patches, the correlations are 0.30, 0.25 and 0.25 between NDVI and 28 days of total rainfall, total flow and mean Perc_low (i.e. proportion of flow relative to the overbank flood threshold) respectively. The highest correlations were also found between proportion of high NDVI areas (i.e. NDVI > 0.6) and 28 days of total rainfall, total flow and mean Perc_low (correlations are 0.30,

Fig. 4. Annual mean area of NDVI class value as a percentage of entire asset: a) Barbers Lagoon and b) Wee Waa to Bugilbone.

63

0.32 and 0.32 respectively). Therefore, 28 days of total rainfall (Rainfall_total), total flow (Flow_total) and mean Perc_low (Perclow_mean) were used as explanatory variables in classification analysis. Datasets with Perc_low values greater than 1 (i.e. dates during flooding) were excluded in the classification analysis because the NDVI values for these datasets are underestimated due to of the reflectance of flood water. Regression tree analysis of mean NDVI for tree patches at all assets showed that the major variable to separate NDVI values was groundwater levels. In areas where the groundwater is deeper than 18.94 m below ground, the mean NDVI values for tree patches are 0.48 (sd ¼ 0.08); while in areas where the groundwater is equal to or shallower than 18.94 m below ground, the mean NDVI values are 0.57 (sd ¼ 0.08). The former category (Groundwater > 18.94 m) is mostly comprised of the asset Pian Ck where groundwater levels range from 19 to 27 m in the 23 years period. When groundwater levels exceed 18.94 m, dates with shorter inter-flood dry period (Days_lastflood  96 days) have higher NDVI values (Node 2 in Fig. 5, mean NDVI ¼ 0.54, sd ¼ 0.07) than dates with longer inter-flood dry period (mean NDVI ¼ 0.47, sd ¼ 0.07). After this, maximum temperatures are seen to relate to NDVI values with temperatures below 27.2 more related to higher NDVI (mean NDVI ¼ 0.49, sd ¼ 0.07) (Node 4 in Fig. 5) than temperatures above this threshold (mean NDVI ¼ 0.44, sd ¼ 0.06) (Node 5 in Fig. 5). This implies that stressed riparian trees in deeper groundwater zones are associated with longer dry periods (more than 3 months) and hotter weather (more than 27 maximum temperature). Pian Ck was then removed from regression tree analysis so as to examine the relationships between NDVI and climatic and hydrological variables at areas where groundwater levels did not always exceed 19 m below ground. In this case, maximum temperature is the primary variable classifying the NDVI data, with cooler maximum temperatures (<33.9  C) tending to relate to slightly higher NDVI for tree patches than warmer maximum temperatures (p < 0.001) (Fig. 6). In either case, after temperature, antecedent rainfall (i.e. 28 days total rainfall) becomes the next variable which splits the data. When maximum temperatures are cooler (33.9  C), higher antecedent rainfall (>23.8 mm) is related to

Fig. 5. Regression tree showing classification of mean NDVI for the tree patches for dates when groundwater levels exceed 18.94 m.

64

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

Fig. 6. Regression tree showing classification of mean NDVI for the tree patches in all assets except Pian Ck.

observations with the highest NDVI ranges (mean ¼ 0.60, sd ¼ 0.07) in the tree patches (Node 6 in Fig. 6). When antecedent rainfall is lower (<23.8 mm) inter-flood dry period becomes an important split for the data (p < 0.001), with shorter dry periods associated with higher NDVI. When the weather is hotter (Temp_max > 33.9  C) and when the antecedent rainfall in the past 28 days exceeds 70 mm, high NDVI values (mean ¼ 0.59, sd ¼ 0.07) were recorded for the tree patches (Node 11 in Fig. 6). This is similar to cooler weather though with a lower rainfall threshold (23.8 mm) to achieve similarly high NDVI values (Node 6 in Fig. 6). However, if the weather was hotter (Temp_max > 33.9  C) and dryer (Rainfall_total  70 mm), and with longer dry periods (Days_lastflood > 271 days), the lowest range of NDVI values was recorded for the tree patches in the assets with a mean of 0.46 (sd ¼ 0.07) (Node 10 in Fig. 6). 4.3.2. NDVI class areas for dense riparian vegetation zones Compared to tree patches (Fig. 6), a similar regression tree was generated when relating proportion of NDVI class areas with climatic and hydrological variables (Fig. 7). It depicts maximum

temperature above or below 25.5  C as the primary determinant of the proportion of areas of high NDVI (i.e. NDVI > 0.6). Lower maximum temperatures are generally related to larger proportions of high NDVI area. When maximum temperatures are lower than 25.5  C, larger proportions of high NDVI area were found in situations when the antecedent rainfall is low (Rainfall_total  42 mm) but cooler (Temp_max  14.4  C) (Node 4 in Fig. 7), or when antecedent rainfall is high (Rainfall_total > 42 mm) and groundwater is shallower (Groundwater  16.04 m) (Node 7 in Fig. 7). In contrast, smaller proportions of areas of high NDVI were found related to situations when it is warm and dry (Node 5 in Fig. 7), or wet but with deeper groundwater (Node 8 in Fig. 7). When maximum temperatures are cooler (Temp_max  25.5  C) and antecedent rainfall is higher (Rainfall_total > 42 mm), area of high NDVI tends to comprise more than 40% of the assets if groundwater is shallower than 16.04 m (Node 7 in Fig. 7), but less than 40% of asset area if groundwater is deeper (Node 8 in Fig. 7). When the weather is warmer (Temp_max > 25.5  C), the next determinant of high NDVI area is the inter-flood dry period.

Fig. 7. Regression tree showing classification of proportion of area with high NDVI (>0.6) for the assets with dense riparian vegetation (assets 5e7).

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

Situations when the last flood occurred more than 52 days ago are related to very low areas of high NDVI (in average 7%, Nodes 13 and 14 in Fig. 7), unless antecedent rainfall has been very high (Rainfall_total > 95.5 mm) (Node 15 in Fig. 7). In comparison, the NDVI value is higher when the weather is warmer and the interflood dry period is less than 52 days. Under such conditions on average 20% of the asset areas has high NDVI (Node 10 in Fig. 7). 4.3.3. NDVI class areas for sparse riparian vegetation zones Similar to tree patches and dense riparian vegetation zones, the primary split in high NDVI area for riparian vegetation community with sparse trees is determined by maximum temperature, with 21.2  C as the pivotal temperature (Fig. 8). Below or equal to 21.2  C, area of high NDVI has a large spread but has an average of about 40% of the assets. Above a maximum temperature of 21.2  C, rainfall becomes the next most important variable. When antecedent rainfall is less than 54.8 mm and days since last flood exceed 36 days, areas of high NDVI are very low (Node 6 in Fig. 8). Warmer temperature, lower rainfall but recently flooded dates (Node 5 in Fig. 8) have higher NDVI areas which are within a similar range to cooler temperature (Node 2 in Fig. 8). If antecedent rainfall is greater than 54.8 mm and maximum temperature is higher than 21.2  C (Node 7 in Fig. 8), this can produce a slightly lower range of high NDVI area as compared to when the weather is cooler (Node 2 in Fig. 8). 5. Discussion This paper presents an exploratory analysis of how water availability affects vegetation vigour in the riparian zone by relating a range of climatic and hydrological variables to NDVI. Of these variables, the regression trees presented here consistently indicate that maximum temperature is the variable that primarily splits NDVI values, followed by antecedent 28-day rainfall and then the

65

hydrological variables of inter-flood dry period and groundwater levels. We believe this is because while rainfall, flood and groundwater are water supplies to riparian vegetation, temperature moderates the availability of this water and hence NDVI.

5.1. Climatic effects on NDVI The relative impacts of climatic factors (including rainfall and temperature) on NDVI differ between climates and bioregions. Research on arid and semi-arid regions where water is the limiting factor on productivity tends to focus on rainfall alone (Barbosa et al., 2006; Gaughan et al., 2012). On the other hand, temperature is a strong driver of NDVI in higher latitudes (Ichii et al., 2002), at high altitudes (Hu et al., 2011; Liang et al., 2012) and for bioregions where deciduous vegetation exists (see Wang et al., 2008). Our exploratory regression tree analyses consistently show maximum temperature to be the dominating variable influencing riparian NDVI values in our study area (excluding Pian Ck). This is different with some other studies at the mid-latitudes, in which temperature is often reported to be secondary to rainfall in terms of influencing NDVI (Peng et al., 2012; Wang et al., 2003). However, these studies are often focussed on biomes that are not semi-arid and look at NDVI only within the distinct growing season. Arguably there is no distinct growing season for the eucalypt dominant vegetation of our study region. And if we were to characterise seasons by temperature, looking at individual seasons in our study region would show antecedent rainfall to be the dominant influence on NDVI. Wen et al. (2012) found that at Macquarie Marshes (located in a study region biogeographically relatively similar to Namoi Catchment), monthly total rainfall is more influential to monthly mean NDVI than the mean minimum daily temperature in a month. However, maximum temperature was not included in the modelling. In contrast, our study investigated daily NDVI and its relationships with daily maximum temperature.

Fig. 8. Regression tree showing classification of proportion of area with high NDVI (>0.6) for the assets with sparse riparian vegetation (assets 3 and 4).

66

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

In our study, maximum temperature is negatively related to NDVI. Higher maximum temperature is found associated with lower mean NDVI for the tree patches and smaller areas of high NDVI in riparian zones. This is consistent with other studies which suggested temperature to be negatively correlated with NDVI during spring (Yang et al., 1997) and summer (Wang et al., 2003). This can be due to lower soil moisture caused by higher temperature, especially in semi-arid regions when rainfall is depleted. Soil moisture is widely seen to be the link between rainfall, temperature and NDVI (Wang et al., 2003). Though temperature does not directly influence how much water enters the soil, it moderates soil moisture levels through its ability to drive potential evaporation. The importance of evaporation and its negative impact on NDVI have been reported in other studies (Yang et al., 1998, 1997). In our study, more rain falls in the warmer months of the year in the Namoi catchment, but these months are also strongly coupled with high potential evaporation, driven in great part by the higher temperatures. Our results suggest that the moderating effect of temperature on soil moisture can be significant so that much of the water entering the soil at this time soon becomes unavailable to plants, resulting in lower NDVI. Conversely, the smaller amount of rain that enters the soil in the cooler months is likely to be more available to vegetation due to low potential evaporation, resulting in generally higher NDVI. Once temperature is controlled for, antecedent rainfall becomes the dominant variable which splits NDVI values, with higher rainfall related to higher NDVI values and larger areas of high NDVI. The regression trees consistently show that more rain is required in the warmer months compared to cooler months to achieve similar mean NDVI values in tree patches or areas of high NDVI in riparian zones, presumably because of higher evaporation. For example, to achieve similar mean NDVI values for tree patches, more than 70 mm of antecedent rainfall is required when the maximum temperature is greater than 33.9  C, in comparison to 23.8 mm of antecedent rainfall required when the maximum temperature is lower than 33.9  C (see Fig. 6).

5.2. Surface water influences on NDVI Healthy riparian ecosystems depend on suitable flow regimes. The natural regime of stream flow e encompassing frequency, magnitude, duration and timing of water e is what many riparian organisms have specifically adapted to (Poff et al., 1997). Changes to the predictability and variability of stream flow due to water impoundment and extraction are likely to test an organism's adaptability and thus ability to survive (Allan and Castillo, 2007). This situation has been observed in riparian ecosystems the world over (Nilsson and Berggren, 2000; Poff et al., 1997) including along Australia's highly altered Murray-Darling River system where changes in the natural flow regime have led to reduced growth rates, accelerated mortality and reduced recruitment for eucalypts in its river red gum (E. camaldulensis) forests (Bacon et al., 1993; Cunningham et al., 2009). Positive NDVI responses to flooding events in large areas of naturally vegetated floodplain have been documented in eastern inland Australia (Parsons and Thoms, 2013; Sims and Colloff, 2012). For example, Sims and Colloff (2012) reported a nearly 20% increase in NDVI for 13 months following flood recession as a result of a flood which inundated more than 50% of the Paroo River Wetlands. However, when combining hydrological and climatic factors, Wen et al. (2012) reported that rainfall rather than flow levels into a wetland has a more significant relationship with mean NDVI. Sun et al. (2008) found a negative correlation between stream flow and NDVI during growing season in the Minjiang River region of

western China, and attributed this to water uptake by vegetation in the growing season. In our study, surface water variables were not indicated to be as influential as rainfall. Various surface water variables were investigated, including antecedent flow (Flow_total), inter-flood dry period (Days_lastflood), the proportion of flow relative to the overbank flood threshold (Perc_low), and the mean Perc_low over the past days (Perclow_mean). Inter-flood dry period was identified to be the only significant surface water variable in regression trees. Generally this variable becomes important when limited water is available from rainfall. For tree patches and dense riparian zones, when antecedent 28-day rainfall is low, areas with less than 7e9 months of inter-flood dry periods have relatively higher NDVI compared to those haven't been flooded for more than 7e9 months. For sparse riparian zones, much shorter dry period and hence more frequent flooding is associated with high NDVI. In this case, when antecedent 28-day rainfall is low (<54.8 mm), areas that have not been flooded in the past 36 days have significantly low areas of high NDVI in the riparian zones (see Fig. 8). This is likely due to the quicker browning off effect of non-tree vegetation, especially grasses, as opposed to trees which are more buffered from changes in water availability due to deeper rooting systems and greater carbohydrate reserves (Peng et al., 2012; Wang et al., 2003). 5.3. Groundwater influences on NDVI Groundwater is an important water source for maintenance, abundance and composition of many riparian communities, especially in arid and semi-arid regions. Depletion of groundwater has been associated with stress, mortality and lack of recruitment in groundwater-dependent vegetation in many parts of the world (Cunningham et al., 2009; Horton et al., 2001; Scott et al., 1999; Stromberg et al., 1996). Aguilar et al. (2012) reported strong linear relationships between mean and maximum NDVI and groundwater levels in dry years; however the relationships are much weaker in wet years. In our study, the importance of groundwater level was identified in the classification of tree patches and of dense riparian zones. For the tree patches, groundwater level was the primary split, mostly due to the effect of the asset Pian Ck which had significantly deeper groundwater levels than the rest of the assets. This is associated with markedly lower NDVI compared to rest of the assets (see Fig. 3). For dense riparian zones, groundwater is important in splitting NDVI at situations when the weather is cooler and wetter. In this case, situation with groundwater levels shallower than 16 m below ground is associated with larger areas of high NDVI. Groundwater level is not an important variable in splitting NDVI for riparian zones with sparse trees (assets 3 and 4). This could be the effect of fewer trees in these riparian zones, as non-tree vegetation e especially grasses e are unlikely to be effected by changes of groundwater levels at the depths observed in our study area. 5.4. Limitations and implications for management Regression tree analysis allowed us to classify NDVI and identify critical explanatory variables without the assumption about their linear relationships (as is the case for linear regression analysis used in other literature). It provides insight into the roles of one explanatory variable in relation to other explanatory variables. For example, inter-flood dry period often becomes important when rainfall is below certain thresholds. The results of regression tree analysis are sensitive to the dataset used. However, in our study a consistent story emerged in terms of the relationships between NDVI and climatic, surface water and groundwater variables. This

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

demonstrates the robustness of the findings. The thresholds identified in the regression trees apply to the dataset used in this study. Therefore, these thresholds should be seen as indicative and the exact values should not be used for water planning purposes. Lack of field studies to cross-verify remotely-sensed observations is arguably a limitation of this study. Field studies would have allowed us to make physiological observations of vegetation stress brought on by low water availability as well as observations of land clearing and grazing which would affect NDVI values in ways not accounted for in this study. This said, remotely sensed NDVI has allowed this study to rapidly collect data not only across a sizeable landscape, but more importantly, across a time span of 23 years, for which field data on vegetation vigour in our study region does not exist. In this way the use of remote sensing can be viewed as a particular strength especially for this exploratory analysis: the large number of data points generated by remote sensing data gives strength to the data mining technique of decision tree analysis. Another limitation is that NDVI does not account for soil reflectance which can make NDVI difficult to interpret when vegetation cover is low and the surface substrate is unknown. It is possible that NDVI can indicate a range of values when soils in the background are different, despite vegetation vigour being effectively the same (Rondeaux et al., 1996; Sims and Colloff, 2012). Care was taken to demarcate dense patches of trees for tree patch NDVI, so as to minimise the problem of soil reflectance. Among the variables we investigated, flow and groundwater levels are manageable through human intervention. Though not as influential as temperature and rainfall, our study indicates these variables are important for assisting the maintenance of healthy riparian vegetation communities. Inter-flood dry period is shown to be important for maintenance of NDVI levels, particularly in summer when high temperatures can reduce the amount of available water in the soil fed from rain. Our study suggests that environmental flows which allow water to overflow the bank in summer or in particularly dry winters will assist to maintain healthy riparian vegetation. 6. Conclusion Regression tree analysis was used to investigate how water availability affects vegetation vigour in riparian zone in nine ecological assets in the Namoi catchment, Australia. This was achieved by relating a range of climatic and hydrological variables to 23 years of NDVI data. Of these variables, the regression trees presented here consistently indicate that maximum temperature is the variable that primarily splits NDVI values, followed by antecedent 28-day rainfall and then the surface water variable (inter-flood dry period) and groundwater levels. Maximum temperature is the dominant variable and is negatively related to NDVI. This can be due to lower soil moisture caused by higher temperature, especially in semi-arid regions when rainfall is depleted. More rain is required in the warmer months compared to cooler months to achieve similar mean NDVI values in tree patches or areas of high NDVI in riparian zones, presumably because of higher evaporation. Interflood dry period was identified to be the only significant surface water variable in regression trees. Generally this variable becomes important when rainfall is limited. Our study also suggested that shallower groundwater levels sustain the NDVI and hence vegetation greenness when cooler and wetter. Acknowledgement We would like to thank Geoscience Australia for providing Landsat 5 data between 1987 and 1998. This study was funded by the Cotton Research and Development Corporation and the

67

National Centre for Groundwater Research and Training (Grant ID 22654), an Australian Government initiative supported by the Australian Research Council and the National Water Commission. Appendix A. Supplementary data Supplementary data related to this article can be found at http:// dx.doi.org/10.1016/j.jaridenv.2014.09.010. References Aguilar, C., Zinnert, J.C., Polo, M.J., Young, D.R., 2012. NDVI as an indicator for changes in water availability to woody vegetation. Ecol. Indic. 23, 290e300. Allan, J.D., Castillo, M.M., 2007. Stream Ecology: Structure and Function of Running Waters, second ed. Springer, Dordrecht. Australian Bureau of Meteorology, 2013a. Atlas of Groundwater Dependent Ecosystems. http://www.bom.gov.au/water/groundwater/gde/map.shtml (accessed 19.12.2012). Australian Bureau of Meteorology, 2013b. Climate Data Online. http://www.bom. gov.au/climate/data/ (accessed 17.6.2013). Bacon, P.E., Stone, C., Binns, D.L., Leslie, D.J., 1993. Relationships between water availability and Eucalyptus camaldulensis growth in a riparian forest. J. Hydrol. 150, 541e561. Barbosa, H.A., Huete, A.R., Baethgen, W.E., 2006. A 20-year study of NDVI variability over the northeast region of Brazil. J. Arid Environ. 67, 288e307. Barma Water Resources, Thurtell, L., Wettin, P., 2012. Environmental Water Delivery: Namoi River., Prepared for Commonwealth Environmental Water, Department of Sustainability, Environment, Water, Population and Communities, Canberra. Blakers, R., 2011. Documentation: CleanGroundwaterData_v2-x.m. ANU, Canberra. Bulcock, H.H., Jewitt, G.P., 2010. Spatial mapping of leaf area index using hyperspectral remote sensing for hydrological applications with a particular focus on canopy interception. Hydrol. Earth Syst. Sci. 14, 383e392. Chander, G., Markham, B., Helder, D., 2009. Summary of current radiometric calibration coefficients for landsat MSS, TM, ETMþ, and EO-1 ALI sensors. Remote Sens. Environ. 113, 893e903. Cho, M.A., Skidmore, A., Corsi, F., van Wieren, S.E., Istiak, S., 2007. Estimation of green grass/herb biomass from airborne hyperspectral imagery using spectral indices and partial least squares regression. Int. J. Appl. Earth Obs. Geoinf. 9, 414e424. CSIRO, 2007. Water Availability in the Namoi, A Report to the Australian Government from the CSIRO Murray-Darling Basin Sustainable Yields Project. CSIRO, Australia, p. 154. Cunningham, S.C., Mac Nally, R., Read, J., Baker, P.J., White, M., Thomson, J.R., Griffioen, P., 2009. A robust technique for mapping vegetation condition across a major river system. Ecosystems 12, 207e219. Cunningham, S.C., Thomson, J.R., Mac Nally, R., Read, J., Baker, P.J., 2011. Groundwater change forecasts widespread forest dieback across an extensive floodplain system. Freshwater Biol. 56, 1494e1508. De'Ath, G., Fabricius, K.E., 2000. Classification and regression trees: a powerful yet simple technique for ecological data analysis. Ecology 81, 3178e3192. Eco Logical, 2009. Riverine Vegetation in the Namoi Catchment: an Assessment of Type and Condition: Final Report Prepared for Cotton Catchment Communities CRC and Namoi Catchment Management Authority. Eco Logical Australia Pty Ltd. Exelis Visual Information Solutions, 2012. ENVI 5.0. Gaughan, A.E., Stevens, F.R., Gibbes, C., Southworth, J., Binford, M.W., 2012. Linking vegetation response to seasonal precipitation in the Okavango-KwandoZambezi catchment of southern Africa. Int. J. Remote Sens. 33, 6783e6804. Goward, S.N., Dye, D.G., 1987. Evaluating North American net primary productivity with satellite observations. Adv. Space Res. 7, 165e174. Groenvald, D.P., Baugh, W.M., 2007. Correcting satellite data to detect vegetation signal for eco-hydrologic analyses. J. Hydrol. 344, 135e145. Hansen, P.M., Schjoerring, J.K., 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalised difference vegetation indices and partial least squares regression. Remote Sens. Environ. 86, 542e553. Horton, J.L., Kolb, T.E., Hart, S.C., 2001. Responses of riparian trees to interannual variation in ground water depth in a semi-arid river basin Plant. Cell Eviron. 24, 293e304. Hothorn, T., Hornik, K., Zeileis, A., 2006. Unbiased recursive partitioning: a conditional inference framework. J. Comput. Graph. Stat. 15, 651e674. Hu, M.Q., Mao, M., Sun, H., Hou, Y.Y., 2011. Study of normalized difference vegetation index and its correlation with climate factors in the three-river-source region. Int. J. Appl. Earth Obs. Geoinf. 13, 24e33. Ichii, K., Kawabata, A., Yamaguchi, Y., 2002. Global correlation analysis for NDVI and climatic variables and NDVI trends: 1982e1990. Int. J. Remote Sens. 23, 3873e3878. Ji, L., Peters, A.J., 2003. Assessing vegetation response to drought in the northern Great Plains using vegetation and drought indices. Remote Sens. Environ. 87, 85e98.

68

B. Fu, I. Burgher / Journal of Arid Environments 113 (2015) 59e68

Liang, S., Ge, S., Wan, L., Xu, D., 2012. Characteristics and causes of vegetation variation in the source regions of the Yellow River. China Int. J. Remote Sens. 33, 1529e1542. McGrath, G.S., Sadler, R., Fleming, K., Tregoning, P., Hinz, C., Veneklass, E.J., 2012. Tropical cyclones and the ecohydrology of Australia's recent continental-scale drought. Geophys. Res. Lett. 39, L03404. Nilsson, C., Berggren, K., 2000. Alterations of riparian ecosystems caused by river regulation. BioScience 50, 783e792. NSW Department of Water and Energy, 2008. Continuous Monitoring PINNEENA 9.2 DVD. NSW Department of Water and Energy, Sydney. NSW Office of Water, 2011. Interconnection of Surface and Groundwater Systems River Losses from Losing-disconnected Streams. NSW Office of Water, Sydney. Otto, M., Scherer, D., Richters, J., 2011. Hydrological differentiation and spatial distribution of high altitude wetlands in a semi-arid Andean region derived from satellite data. Hydrol. Earth Syst. Sci. 15, 1713e1727. Parsons, M., Thoms, M.C., 2013. Patterns of vegetation greenness during flood, rain and dry resource states in a large, unconfined floodplain landscape. J. Arid Environ. 88, 24e38. Peng, J., Dong, W., Yuan, W., Z, Y., 2012. Responses of grassland and forest to temperature and precipitation changes in Northeast China. Adv. Atmos. Sci. 29, 1063e1077. Petus, C., Lewis, M., White, D., 2012. Using MODIS normalized difference vegetation index to monitor seasonal and inter-annual dynamics of wetland vegetation in the Great Artesian Basin: a baseline for assessment of future changes in a unique ecosystem. In: International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XXII ISPRS Congress, Melbourne, Australia, pp. 187e192. Poff, N.L., Allan, J.D., Bain, M.B., Karr, J.R., Prestegaard, K.L., 1997. The natural flow regime. BioScience 47, 769e784. Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ. 55, 95e107.

Scott, M.L., Shaforth, P.B., Auble, G.T., 1999. Responses of riparian cottonwoods to alluvial water table declines. Environ. Manag. 23, 347e358. Sheldon, F., Thoms, M.C., Berry, O., Puckridge, J., 2000. Using disaster to prevent catastrophe: referencing the impacts of flow changes in large dryland rivers. Regul. Rivers Res. Manag. 16, 403e420. Sims, N.C., Colloff, M.J., 2012. Remote sensing of vegetation responses to flooding of a semi-arid floodplain: Implications for monitoring ecological effects of environmental flows. Ecol. Indic. 18, 387e391. Song, C., Woodcock, C.E., Seto, K.C., Lenney, Mary Pax, Macomber, S.A., 2001. Classification and change detection using landsat TM data: when and how to correct atmospheric effects? Remote Sens. Environ. 75, 230e244. Stromberg, J.C., Tiller, R., Richter, B., 1996. Effects of groundwater decline on riparian vegetation of semiarid regions: the San Pedro, Arizona. Ecol. Appl. 6, 113e131. Sun, P., Liu, S., Jiang, H., Lu, Y., Liu, J., Lin, Y., Liu, X., 2008. Hydrologic effects of NDVI time series in a context of climatic variability in an upstream catchment of the Minjiang River. J. Am. Water Resour. Assoc. 44, 1132e1143. Wang, J., Meng, J.J., Cai, Y.L., 2008. Assessing vegetation dynamics impacted by climate change in the southwestern karst region of China with AVHRR NDVI and AVHRR NPP time-series. Environ. Geol. 54, 1185e1195. Wang, J., Rich, P.M., Price, K.P., 2003. Temporal responses of NDVI to precipitation and temperature in the central Great Plains, USA. Int. J. Remote Sens. 24, 2345e2364. Wen, L., Yang, X., Saintilan, N., 2012. Local climate determines the NDVI-based primary productivity and flooding creates heterogeneity in semi-arid floodplain. Ecol. Model. 242, 116e126. Yang, L., Wylie, B., Tiezen, L.L., Reed, B.C., 1998. An analysis of relationships among climate forcing and time-integrated NDVI of grasslands over the U.S. Northern and Central great Plains. Remote Sens. Environ. 65, 25e37. Yang, W., Yang, L., Merchant, J.W., 1997. An assessment of AVHRR/NDVIecoclimatological relations in Nebraska, U.S.A. Int. J. Remote Sens. 18, 2161e2180.