Long-term effects of climate and land cover change on freshwater provision in the tropical Andes

Introduction Conclusions References


Introduction
Andean headwater catchments play a pivotal role to supply fresh water for downstream water users (Urrutia and Vuille, 2009;Roa-García et al., 2011). Although the ecosystems in the tropical Andes have been modified by anthropogenic disturbances for at least 7000 years (Bruhns, 1994), it is only since the early 20th century that natural ical conditions (Poveda and Mesa, 1997;Restrepo and Kjerfve, 2000). In the tropical Andes, the temporal variability of precipitation is strongly related to oceanic and atmospheric conditions over the Pacific Ocean and Amazon basin (Vuille et al., 2000;Marengo et al., 2004). The impact of El Niño-Southern Oscillation (ENSO) is clearly noticeable on the Western escarpment of the Andes in Ecuador and northern Peru 20 (Tapley and Waylen, 1990;Rossel, 1997), and decreases with altitude as the steep, high-altitude topography of the Andean range creates distinct microclimates (Mora and Willems, 2012).
Direct anthropogenic impact resulting from land cover change is rapidly transforming the hydrological functioning of tropical Andean ecosystems (Vanacker et al., 2003;25 Farley et al., 2004;Molina et al., 2012). The hydrological response is diverse, as changes in vegetation affect various components of the hydrological cycle including evapotranspiration (Nosetto et al., 2005), infiltration (Molina et al., 2007) and surface runoff (Bathurst et al., 2011). The clearance of native forest for arable and grazing land induces rapid changes in soil physical properties reducing soil infiltration capacity (Bosch and Hewlett, 1982;Molina et al., 2007), and increasing surface runoff as a result of soil compaction and reduced evapotranspiration (Ruprecht and Schofield, 1989). As a consequence, the conversion of native forests to agricultural land often results in an increase of the annual water yield, but a reduction of the low flows (Bruijnzeel, 5 1990; Andréassian, 2004). In contrast, afforestation and/or reforestation of grasslands and arable lands lead to a reduction in soil moisture and total water yield as a result of greater canopy interception and evapotranspiration (Bruijnzeel, 2004;Scott et al., 2005;Farley et al., 2005;Buytaert et al., 2007).
In tropical Andean ecosystems characterised by large inter-and intra-annual variability in hydrometeorological conditions, little is known about the relative importance of climate change and direct anthropogenic perturbations on streamflow. At large spatial scale (> 100 km 2 ), the patterns of land cover change are notoriously dynamic, both in space and time, and are commonly associated to climatic and altitudinal gradients. In this paper, we assess multi-decadal change in freshwater provision based on long time 15 series of hydrometeorological data and land cover reconstructions. Given the strong temporal variability of precipitation and streamflow data related to El Niño-Southern Oscillation (ENSO), we use Hilbert-Huang transformation to detrend the time series of streamflow and precipitation data. The adaptive data analysis is based on empirical mode decomposition techniques that are appropriate for nonlinear and nonstationary 20 time series data (Huang et al., 1998). After empirical mode decomposition, the remaining long-term trends in streamflow and precipitation are contrasted to the observed patterns of land cover change.
The study is realised in an exceptional setting, the Pangor catchment (c. 282 km 2 ) in the Ecuadorian Andes. Situated on the Western escarpment of Ecuadorian Andes, 25 the area is particularly affected by El Niño-Southern Oscillation cycles (Rossel, 1997 , 2015). By analysing long time-series of hydrometeorological data, we specifically tested the relative sensitivity of streamflow to climate and land cover change.  (Gonzalez Artieda et al., 1986), and are characterised by a remarkably high water-holding capacity and soil organic matter content when undisturbed (Podwojewski et al., 2002). The landscape pattern now reflects several decades of rapid land cover change. At mid and low altitudes, a complex patchwork of small agri-20 cultural plots, remnants of sub-alpine cloud forest, and patches of abandoned land with regeneration of natural shrub vegetation can be observed. Smallholder farming is the dominant agricultural activity, and crop rotation is a common practice where annual crops are alternated with pasture. Crop species vary with altitude, with maize (Zea mays) grown in association with common bean (Phaseolus vulgaris) at altitudes below Introduction are only remnant on steep slopes in areas with very low accessibility. Above the natural treeline, the páramo grasslands are dominant, but plantation forests with exotic tree species (Pinus radiata and Pinus patula) now cover extensive areas (Balthazar et al., 2015).

Land cover change detection
Land cover change for the period 1963-2009 was reconstructed based on panchromatic aerial photographs (IGM, Quito, Ecuador) and high resolution Landsat TM (15 October 1991) and ETM+ (3 November 2001, and 6 September 2009) images. A full coverage of aerial photographs at the scale of 1/60 000 was obtained for November 1963 and 1977, and land cover mapping was realized following the procedure described by Molina et al. (2012). Three Landsat scenes (1991,2001,2009, from the same season) with 1T level of pre-processing were acquired from the USGS archive, and images were atmospherically and topographically corrected with ATCOR3 . To support the definition of land cover classes, a WorldView II im- 15 age of 2010 with a horizontal resolution of 0.5 m (PAN) and 2 m (MS) was used (Digital Globe). A multi-source data integration method developed by Petit and Lambin (2001) was applied to reduce imprecision and inconsistency that may result from the comparison of heterogeneous datasets (Balthazar et al., 2015). Four land cover types were defined: 20 AL: agricultural land dominated by pastures and annual crops; F: montane cloud and subalpine forests (including primary and secondary forests); P: páramo grasslands dominated by tussock grasses and dwarf shrubs, and PP: exotic forest plantations dominated by Pinus radiata and patula. Introduction for the Pangor AJ Chimbo gauging station (Fig. 1). The time series are appropriate for studies of long-term trends in precipitation and streamflow, as they are of good quality (data gaps < 10 %) and cover a prolonged period of time (1974 to 2008). Only during 1997 and 2003, there is a gap in the series of observed streamflow and precipitation data of more than 3 months. The multiple linear regression method described by Mora 10 and Willems (2012) was used to fill gaps. This method estimates correlation coefficients between all pairs of hydrometeorological stations, either for the current and preceding month, and applies a multiple linear regression equation to predict missing flow or precipitation data. Given the low density of rain gauges in the Pangor catchment, we applied the re-15 gionalization method proposed by Mora and Willems (2012) to obtain catchment-wide or areal average precipitation depths. Based on data of altitude, vegetation pattern and precipitation regime, four meteorological regions were delineated, and the closest rain gauge station was assigned to each region. Additionally, an altitude correction factor was applied based on the observed relationship between mean annual precipitation 20 and altitude. The areal average precipitation for the entire Pangor catchment was then calculated by summing the weighted precipitation (by surface area) of the four regions, and dividing this value by the sum of the weights. The areal average daily precipitation depths, P d (mm), were aggregated into monthly data for the period 1974-2008. The time series of streamflow data  is based on daily water stage read- 25 ings at Pangor AJ Chimbo gauging station (Fig. 1). Stage records (m) were converted into discharge records (m 3 s −1 ) using the stage-discharge rating curve developed by IN-AMHI. The discharge was then converted to daily equivalent water depth, WD d (mm), to allow direct comparison with precipitation records. To condense the time series data, the daily water depths were aggregated at monthly time step. The monthly water depths were split into quick and slow flows using the hydro-statistical toolkit WETSPRO (Water Engineering Time Series PROcessing tool). This procedure is based on subflow separation techniques, and applies a generalization of the original Chapman filter. The

Conclusions
Chapman filter assumes exponential recession for the hydrological subflows and is derived from the general equation of a low pass filter (Willems, 2009).
To assess the consistency of the hydrometeorological datasets, two simple tests were carried out following Costa et al. (2003). First, the mean annual evapotranspiration, ET yr , was estimated as the difference between the mean areal average annual ). Second, the annual evapotranspiration data, ET yr , are highly associated with the areal average annual precipitation depths, P yr (Fig. 2), which is characteristic for tropical Andean basins (Mora and Willems, 2012). These two assessments show that the quality of the hydrometeorological data is acceptable, so that they can reliably be used for analyses of long-term trends.

Empirical mode decomposition EMD
The empirical mode decomposition (EMD), a technique developed by Huang et al. (1998), was here used to identify trends in the time series  of areal average monthly precipitation depths and equivalent water depths. The EMD method decomposes a given signal into a finite set of intrinsic mode function (IMFs) compo- 25 nents and a residual or trend. The method separates non-linear oscillatory patterns of higher frequencies from those of lower frequencies, till a constant or monotonic trend is ultimately obtained (Wu and Huang, 2004). In contrast to more traditional time series analysis techniques, such as Fourier transformation and wavelet analysis, EMD is not based on linear and stationary assumptions (Huang and Wu, 2008). As such, it can be applied to nonlinear and nonstationary data series (Peel and McMahon, 2006;Brisson et al., 2015). Huang et al. (1999) Brisson et al. (2015).
To assess the significance of the trend obtained through EMD, its variability was compared to the variability of non-significant trends. To ensure that non-significant trends 15 have similar characteristics than the one under assessment, a 3-step method is here proposed. First, the monthly values of the observed time series, P and WD, were randomly distributed. The resulting time series features a variability similar to the observed one, but without any meaningful trend. In total, 1000 random time series were generated. Second, following the process of the EEMD, a perturbation was added to the 20 1000 random time series. Finally, the trend in each random time series was derived using EMD. A trend was defined to be significant if its variability is higher than the 99th percentile of the variability of the trends derived from the random signal.

Estimation of the long-term water balance
A budget approach was used to approximate the different components of the water 25 cycle, including evaporation and transpiration (Bruijnzeel et al., 2006). First, the annual water balance for the entire catchment was approximated as: where P yr is the areal average precipitation (mm yr −1 ), HP yr is the horizontal rainfall and cloud interception (mm yr −1 ), WD yr is the equivalent water depth as derived from streamflow measurements (mm yr −1 ), ET yr is the evapotranspiration (mm yr −1 ), and ∆S is the change in soil water storage in the catchment (mm yr −1 . Long-term changes in 5 soil water storage, ∆S, can be neglected, as soils are typically shallow on the Western escarpment of the Andes so that deep infiltration is limited. Horizontal rainfall, HP yr , is here also considered to be negligible for the catchment-wide water balance, as additional water input from the interception of cloud water and wind-driven rain is typically constrained to the narrow(ing) belt of cloud forests (i.e. 11.6 % of the catchment area 10 in 2009). We can then estimate the annual evapotranspiration as Second, a partial water balance was established for the two ecosystems where major changes in land cover occurred (Table 1): the tropical montane cloud forest (defined as the landscape unit between 2200 and 3200 m a.s.l. originally covered by cloud forest), 15 and páramo ecosystems (here defined as the entire landscape unit of high altitude above the continuous forest line, 3200 m a.s.l.). Land cover data were used to estimate temporal changes in partial water balance over the period 1974-2009, as the main hydrological components were parametrized based on land cover type. As the date of the land cover maps does not correspond exactly to the time series of hydrometeorological 20 data, the land cover of 1974 and 2008 was reconstructed based on linear interpolation of existing land cover distributions. Chi-square analysis was used to analyse the significance level of the observed changes.
where K s is a water stress factor, K c is the crop coefficient and ET o is the reference crop evapotranspiration estimated at 1000 mm yr −1 for the páramo ecosystem based on INAMHI (2009 (Vanacker et al., 2003;Guns and Vanacker, 2013). The pattern of afforestation stands in sharp contrast to the deforestation pattern (Fig. 3). Afforestation is mainly concentrated in the subalpine and alpine zones, and started in the early 1990s. About 2/3 of the total decrease in páramo grasslands 15 (−23 km 2 ) results from exotic forest plantations, and only 1/3 from conversion to agricultural land.

Long-term trends in precipitation and streamflow (1974-2008)
The flow regime  largely mimics the yearly variation in precipitation, with maximum mean monthly streamflow in April (equivalent water depth of 86 mm) and 20 low flow in September (25 mm). More than 60 % of the annual flow is concentrated in the period between February and June. Annual values of precipitation and streamflow reveal strong inter-annual variation (Fig. 4). Given the nature of hydrometeorological data in tropical Andean basins, which often display an abrupt pattern of amplitude and frequency modulation at different time scales, EEMD is an ideal method to extract 25 physically meaningful signals. Using EEMD, the times series of monthly precipitation values and equivalent water depths were decomposed into six intrinsic mode function (IMFs 1-6) components plus the residual or trend (Fig. 5) The EEMD detrending analysis shows that the precipitation and streamflow regime changed significantly over time (Fig. 6). The EEMD analysis shows that the observed changes in streamflow  are not the result of long-term climate change. Despite increased precipitation, 5 there is a remarkable decrease in streamflow (Table 1). Over the period 1974-2008, the rate of change varied through time, and two periods of change can be distinguished based on the EEMD time series analysis. Between 1974 and 1991, the monthly precipitation amounts increased sharply, while streamflow and baseflow decreased. The rate of change decreased noticeably for the period 1992-2008 (Fig. 6).

Changes in water balance for montane cloud forest and páramo ecosystems
Land cover dynamics observed in the Pangor catchment are characteristic for the tropical Andes, with rapid deforestation of native forests and afforestation with exotic tree 5 species in more recent decades. Our land cover change analysis indicates that major changes occurred in the montane cloud forest and páramo ecosystems. Table 3 highlights the estimated evaporative gains and losses (hm 3 ) for these two ecosystems over the period 1974-2008. In montane cloud forests, there is a net reduction of annual ET by 1.1 hm 3 (corresponding to an overall ET loss of 4 mm at the catchment scale; Ta-10 ble 3) as a result of the conversion of 40 % of the surface area of montane cloud forest to agricultural land. This is likely to be a conservative estimate as the contribution of additional moisture from the interception of cloud water and wind-driven rain by the cloud forests is not taken into account, and might equal 5 up to 20 % of ordinary rainfall (Bruijnzeel, 2004). 15 On the other hand, the development of 15 km 2 of pine plantation in high alpine grasslands is estimated to have increased transpiration losses by about 8.6 hm 3 or 31 mm (Table 3). Pine forests' water use is very high compared to native páramo vegetation as result of the large total leaf surface area and deep root systems (Buytaert et al., 2007), and it largely affects the soil water storage and retention in organic-rich páramo soils 20 (Farley et al., 2004). In addition, the conversion of ∼ 6 km 2 páramo grassland to agricultural lands is expected to have further increased the transpiration losses by 3.0 hm 3 or 11 mm ( tems, where progressive colonization and afforestation of high alpine grasslands leads to a strong increase in transpiration losses. As such, the reduction in evaporative losses from the conversion of montane cloud forests (−4 mm) is cancelled out by the strong increase of transpiration in the páramo ecosystems following afforestation (+31 mm) and colonization (+11 mm). The observed land cover change in montane cloud forest 5 and páramo ecosystems is estimated to have resulted in a net loss of annual water yield by 38 mm (or 7 % of WD yr ) over the period 1974-2008, mainly as consequence of increased net evapotranspiration in páramo ecosystems. This observation further points to the importance of land use planning, to minimize the potential impact of land cover change on freshwater flow regimes in the tropical Andes.

Soil hydrology following land cover conversions
Land cover conversions are often followed by a phase of intense soil degradation that further exacerbates the anthropogenic impact on surface hydrology (Hofstede et al., 2002). Soil erosion measurements based on fallout-radionuclides for the Chimbo catchment (Central Ecuadorian Andes) clearly illustrate that soil erosion rates highly depend 15 on land cover and management (Henry et al., 2013). Erosion rates in páramo grasslands are estimated at 9 t ha −1 yr −1 , and are significantly higher in forest plantations, pastures and croplands with erosion rates of resp. 21, 24 and 150 t ha −1 yr −1 . The latter values are similar to soil erosion estimates for highly degraded Andean environments in southern Ecuador (Molina et al., 2008;Vanacker et al., 2014). Accelerated soil ero-20 sion has been shown to alter soil hydrological conditions, e.g. through a reduction of soil water infiltration rates and soil water retention capacity (Podwojewski et al., 2002;Molina et al., 2007). The effect of reduced soil water infiltration and retention after land cover change on the overall water balance can be inferred from the flow duration curves (Fig. 7). The increase of infrequent but very high flows combined with the strong de-

Conclusion
Land cover dynamics observed in the Pangor catchment are characteristic for the tropical Andes, with rapid deforestation of native forests and afforestation with exotic tree species in more recent decades. Given the nature of hydrometeorological data in tropical Andean basins, which often display an abrupt pattern of amplitude and frequency 5 modulation at different time scales, EEMD is an ideal method to extract physically meaningful signals. The EEMD analysis shows that the observed changes in streamflow  are not the result of long-term climate change. Despite increased precipitation, there is a remarkable decrease in streamflow that very likely results from direct anthropogenic disturbances after land cover change. Partial water budgets for 10 the montane cloud forest and páramo ecosystems suggest that the strongest changes in evaporative water losses are observed in páramo ecosystems, where progressive colonization and afforestation of high alpine grasslands leads to a strong increase in transpiration losses. This observation further points to the importance of land use planning, to minimize the potential impact of land cover change on freshwater flow regimes 15 in the tropical Andes. Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version