Introduction

During the past 50 years, rivers worldwide have received great quantities of nutrients such as nitrogen and phosphorus associated with the rapid population growth, energy consumption and food production (Mouri et al. 2011; Su et al. 2011a). Excessive nutrient loadings often cause eutrophication and subsequently accelerate hypoxia (Yin et al. 2004; Chen et al. 2007), leading to significant ecological and social problems (Mouri et al. 2011). Nutrients and dissolved oxygen in rivers therefore act as important indicators of aquatic life as well as ecological conditions of polluted watersheds (Sánchez et al. 2007; García et al. 2010). Consequently, analyzing the dynamics of river nutrients and dissolved oxygen has been applied often (Benson et al. 2006; Castillo 2010; Quiel et al. 2010; Mouri et al. 2011).

Previously, analyzing the dynamics of river nutrients or dissolved oxygen relied generally on measurement data from multiple sites along and within a river. Then, the relationships between these variables and some natural or anthropogenic determinants were identified. However, two issues were always ignored: spatial autocorrelation and scale effects. Spatial autocorrelation is the condition where similarity of attributes is higher among adjacent observations than those separated by greater distance (LeSage and Pace 2009). Nutrients should exhibit spatial autocorrelation to some extent since sites in closer proximity are affected by comparable human disturbances and similar natural characteristics (Ye et al. 2007; Chang 2008; Tu 2011). If spatial autocorrelation is not taken into consideration, determinants cannot be fully characterized or can be falsely interpreted.

The influential factors of nutrients or dissolved oxygen should be scale-dependent, given the scale-dependent characteristics of some determinants (e.g., land cover) (Buck et al. 2004; Chang 2008). However, most previous studies employed a relatively static approach and failed to examine the relationships between the transitions of input variables over time with dramatic changes at different spatial scales (whole sub-basin and buffer) (Buck et al. 2004; Chang 2008), offering limited understanding of the contributing determinants. In addition, most previous studies emphasized some influential factors, either natural or anthropological aspects, such as land use/land cover (LULC), topography, hydrology, etc. These studies did not investigate natural and anthropological determinants simultaneously, and social-economic factors were given less emphasis.

This study attempts to address the above concerns with a case study of Qiantang River in China. Our objectives are to (1) characterize the spatial autocorrelation for dissolved oxygen and nutrients, (2) determine the contributing anthropogenic and natural determinants at two different spatial scales and (3) provide general management recommendations.

Materials and methods

Study site

The Qiantang River, located in eastern coastal China (Fig. 1), is the largest hydrologic system in Zhejiang Province. Qiantang basin occupies approximately 40,000 km2 and has approximately 20 million inhabitants. Annual temperature averages 17.0 °C and mean rainfall is 1,800 mm. Following China’s market transition, this river basin has witnessed rapid population growth and social-economic development in recent decades. During this period, river water quality has declined. Specifically, domestic sewage and agricultural pollution, discharging a large amount of nutrients, contributed primarily to the deteriorated water quality (Huang et al. 2010; Su et al. 2011b). Scientific characterization and interpretation of the contributing determinants of reduced water quality is greatly needed.

Fig. 1
figure 1

Location of Qiantang River: main tributaries are indicated by letters (a), Land use/land cover of Qiantang watershed in 1996 and 2003 (b)

Full size image

Data sources and processing

Samples and chemical analysis

Monthly data of dissolved oxygen (DO) and two nutrient parameters (ammonia nitrogen, NH3-N; total phosphorus, TP) between 1996 and 2003 from 41 monitoring stations (Fig. 1) in Qiantang River were provided by the Environmental Monitoring Center of Zhejiang Province (EMCZP). These two nutrient parameters were identified and selected for study based on national river water monitoring reports and local water quality characteristics. All water samples were analyzed in the laboratory of EMCZP, following national water quality measurement standards (http://english.mep.gov.cn/standards_reports/standards/water_environment/quality_standard/200710/t20071024_111792.htm). The specific measurement method is briefly described as follows: DO, electrochemical probe method; NH3-N, spectrophotometric method with salicylic acid; and TP, ammonium molybdate spectrophotometric method.

Spatial data and processing

A 500-m riparian buffer was selected as our unit of spatial analysis since this width has been used successfully to assess localized determinants of river pollutants (Benson et al. 2006). This buffer width is effective for retaining relevant geospatial information and for reducing the variability of determinants, given that such determinants may become too homogeneous when the width is too small and, conversely, too heterogeneous when the buffer width is too large.

A 30-m resolution digital elevation model (DEM) was used to delineate the sub-basin drainage area and to calculate surface elevation and distance to river source for each monitoring site. This DEM then was used to calculate average slope gradient and standard deviation of slope gradient for land surfaces throughout the sub-basin and within a 500-m buffer of each sample site.

Rural-community population density and gross domestic product (GDP) data were subject to area-weighted average interpolation, and then corresponding values for each sub-basin/buffer in 1996 and 2003 were determined. Based on a digital soil map (1: 50,000) for Zhejiang Province, percentages of soil texture were calculated for each sub-basin/buffer. LULC was interpreted from Landsat Thematic Mapper (TM) images using multiple end-member spectral mixture analysis (MESMA) (Powell et al. 2007). The LULC maps (Fig. 1) included four dominant ecosystems in Qiantang basin: forest, water, built-ups and agricultural land. Supplemental visual interpretation was applied to correct the misclassified pixels by MESMA for each sub-basin and buffer.

Spatial dynamics

Spatial dynamics of nutrients and dissolved oxygen were characterized from two aspects: the average level and the temporal trend. Temporal trend was analyzed using the exploratory approach proposed by Su et al. (2011a), which used Kruskal–Wallis test and Mann–Whitney test to confirm the results of exponential smoothing. Trend changes were classified into five categories: no significant trend, significant gap (the value of certain interval year was higher/lower than its previous year or following year), sudden increase/decrease (the value of certain interval year was higher/lower than both its previous year and next year), and sudden changes (both increase and decrease). All statistical calculations were performed at 95 % confidence level using SPSS 16.0 (SPSS Inc., Chicago, IL).

Spatial autocorrelation

Moran’s I index (Moran 1948) was applied to measure the global spatial autocorrelation of nutrients and dissolved oxygen. Positive values of Moran’s I denote spatially clustered high or low values, while negative values imply non-clustered values among neighboring samples. Local indicator of spatial autocorrelation (LISA) was used to indicate the local spatial association. LISA generates four categories of statistically significant spatial clusters: high–high, low–low, low–high and high–low. High–high cluster indicates high values of DO, and nutrients are surrounded by high values of neighboring observations. Low–low cluster denotes clustering of low values that are below the average value. Low–high (a low value surrounded by high values) and high–low (a high value surrounded by low values) clusters exist within a mix of low and high values. GeoDa 0.9.5-i (Beta) was used for all calculations (Anselin et al. 2006).

Spatial regression

Theory

When considering spatial autocorrelation into model fitting with small data sets, spatial regression can be employed to characterize the spatial structure of the data set (Meng et al. 2009). Spatial regression models typically include three categories, named spatial error (Eq. 1), spatial lag (Eq. 2), and spatial Durbin regression (Eq. 3) (LeSage and Pace 2009).

$$ Y = {{\upeta}} + {{\upalpha}}X + e\left( {e = \lambda We + u} \right) $$
(1)
$$ Y = {{\upeta}} + {{\upalpha}}X + \rho WY + e $$
(2)
$$ Y = {{\upeta}} + {{\upalpha}}X + \rho WY + {{\upbeta}}WX + e $$
(3)

where Y is the annual value of a particular water quality variable in our case; X denotes a vector of determinants; e is the error term; We represents the spatial matrix for the error term; WY represents the spatial matrix for the dependent variable; WX represents the spatial matrix for independent variable(s); α is a vector of coefficients; λ, β and ρ are the spatial autoregressive parameters; u and η are scalar parameters.

The assumption of the spatial error model is that some spatially autocorrelated variables are omitted in the regression, and the model consequently generates spatially autocorrelated error. The spatial lag model assumes that, besides the explanatory variables, a water quality variable in one site is affected by the spatially weighted concentrations in its neighborhood. Spatial lag regression is an effective approach for water quality modeling in most cases, considering the degree of water exchanges and mixing in response to river flows. Spatial lag independent variables are also included in spatial Durbin regression. Spatial autocorrelation of independent variables can become particularly significant in large-scale watersheds given the underlying spatially varying landscape physical templates (Benson et al. 2006). Spatially structural influential factors can impact the dynamics of water quality variables through a variety of biogeochemical and physical processes. These dynamics then can lead to spatial autocorrelation in the patterns of these variables. Consequently, spatial Durbin regression should be more capable to reflect the dynamics of these water quality parameters, such as dissolved oxygen and nutrients used in this study.

Spatial weight matrix

Two categories of the weight matrix used in spatial regression exist, namely the distance matrix and the contiguity matrix. There is no widely accepted criterion for the specification of the spatial weight matrix. However, the basic rule for establishing spatial matrix is to use fewer neighbors instead of extra neighbors (Meng et al. 2009). This implied that the nearest neighbor distance matrix incorporate sufficient spatial autocorrelation information for spatial modeling (Meng et al. 2009). The number of nearest neighbors usually includes the nearest 25 % of total samples. This study thus used the nearest neighbor distance (n = 10 sites) as our spatial weight matrix.

Choice of independent variables

Three criteria facilitated the choice of input independent variables: (1) comparability with previous related research; (2) availability and quality of data sources; and (3) ability to characterize watershed ecological conditions. The final input variables included mean elevation (Elevation), deviation of slope gradient (Slope_std), average slope gradient (Slope_m), distance to river source (Distance), percentage of sand (Sand%), percentage of silt (Silt%), percentage of loam (Loam%), percentage of clay (Clay%), percentage of built-ups (Build%), percentage of water (Water%), percentage of forest (Forest%), percentage of farmland (Farmland%), population density (POP) and gross domestic product (GDP).

Application of spatial regression

Ordinary least square estimators, using inflated t test and F statistics, are not applicable for spatial regression models given the spatially autocorrelated error and spatial lag terms. Therefore, the maximum likelihood estimation (MLE) approach was used to generate consistent estimators of spatial regression. LeSage and Pace (2009) indicated that model specification normally can be chosen based on likelihood ratio tests. Robust Lagrange Multiplier (LM) diagnostics were, in particular, used to determine the form of spatial relationships (e.g., lag or error). Spatial dependence can be identified by robust LM test in the presence of another form (LeSage and Pace 2009). Empirical tests were employed further, in reference to strong priors and expert consulting, to specify the spatial lag or spatial Durbin model. Most statistics for spatial regression were facilitated by SpaceStatPack software (Pace 2003).

Before performing spatial regression, all the variables were standardized and normalized. The traditional variance-in-inflation method was utilized to select input variables, given potential multi-collinearity. Specifically, determinants of spatial patterns in 1996 and 2003 and determinants of spatial dynamics within this time period at the sub-basin and 500 m buffer scales were all explored by spatial regression. These 2 years were selected, because (1) they both were normal hydrological years (similar annual flows and precipitation in terms of the total either-year period) and (2) the Landsat TM images were available to produce LULC maps. Annual mean values of variables were subject first to suitable spatial regression, respectively, for the 2 years. We then calculated the changes of variables using Eq. (4) and performed spatial regression again to investigate the determinants of spatial dynamics between the 2 years.

$$ C = \frac{{R_{2} - R_{1} }}{{R_{1} }} \times 100\,\% $$
(4)

where C is the change of variable; R 1 is the value of variable in 1996; R 2 is the value of variable in 2003.

Results and discussion

Spatial dynamics

General descriptive statistics of the data set were presented in Table 1, and detailed annual statistics were shown in Fig. 2. Mean values of the three variables between 1996 and 2003 (Fig. 3) indicated high values of DO in the western upper basin which was mainly covered by forest and wetland. High DO values were also observed in less developed areas of the Fenshui tributary. Tributaries contributed to the low DO concentrations in the lower and middle part of the mainstream (e.g., Puyang and Dongyang). Obvious spatial variations in NH3-N were observed in the northeastern basin (low values) and in its counterpart basin (high values). Specifically, NH3-N had high values in farmland-dominated sub-basins of the Jiangshan tributary. Extremely high values of NH3-N were observed for the Dongyang tributary, which may be associated with the increasing discharge of industrial wastewater in this section of the basin. TP exhibited relatively complex spatial patterns; higher concentrations generally appeared in the industrial area of southeastern basin and in the estuary part of the river.

Table 1 Descriptive statistics of dissolved oxygen (DO), ammonia nitrogen (NH3-N) and total phosphorus (TP) between 1996 and 2003 in Qiantang River, China (unit: mg/L)
Full size table
Fig. 2
figure 2

Annual statistics of dissolved oxygen, ammonia nitrogen and total phosphorus in Qiantang River between 1996 and 2003 (unit mg/L)

Full size image
Fig. 3
figure 3

Spatial patterns of annual mean values of dissolved oxygen, ammonia nitrogen and total phosphorus in Qiantang River between 1996 and 2003 (unit mg/L)

Full size image

Temporal trends for DO and nutrients for five stations, located in the mainstream and two tributaries of the lower basin, exhibited sudden decreases in DO, denoting a deterioration of water quality over time (Fig. 4). Only one station showed signs of improvement and three sites displayed complex trends. NH3-N and TP remained stable for approximately 75 % of the stations (n = 31) as no significant changes in DO or nutrients were detected. Monitoring stations presenting gaps in NH3-N were mainly concentrated in the mainstream and its tributaries. Two other stations in the upper basin also exhibited similar trends. Changes in TP appeared to be linked to increasing use of detergents in residential areas and materials from wastewater in industrial areas.

Fig. 4
figure 4

Temporal trend of dissolved oxygen, a ammonia nitrogen and total phosphorus in Qiantang River between 1996 and 2003

Full size image

Spatial autocorrelation and temporal changes

Statistics of Moran’s I and patterns of LISA indicate DO and NH3-N generally have a low degree of spatial autocorrelation (Figs. 5, 6, 7). However, Moran’s I for TP was low in initial years and increased and exhibited a moderate degree of spatial autocorrelation in the 2,000 s. Such results suggest that nitrogen pollution remained a regionalized problem across the 8-year time period, while phosphorus pollution has become a more localized problem.

Fig. 5
figure 5

Spatial autocorrelation and spatial pattern of hotspots for dissolved oxygen between 1996 and 2003 in Qiantang River, China

Full size image
Fig. 6
figure 6

Spatial autocorrelation and spatial pattern of hotspots for ammonia nitrogen between 1996 and 2003 in Qiantang River, China

Full size image
Fig. 7
figure 7

Spatial autocorrelation and spatial pattern of hotspots for total phosphorus between 1996 and 2003 in Qiantang River, China

Full size image

Patterns of LISA also varied between years and between variables. Low–low clusters of DO appeared in middle stream in 1996 and 1997, but concentrated in downstream in 2,000 s. These results may be attributed to intensified human activities in downstream areas. Most high–high clusters for NH3-N were concentrated in the tributaries of Dongyang and Wuyi, which should be associated with large quantities of industrial wastewater (Su et al. 2011b). High–high clusters for TP before 1999 located downstream and those after 2002 appeared in Dongyang and Wuyi tributaries, and no high–high clusters were identified between this period.

Determinants of spatial patterns for DO and nutrients

Distance to river source and slope gradient acted as important natural determinants for the spatial patterns of DO and nutrients (Table 2). Water originating from the river source is generally of high quality, and thus the Distance variable is associated negatively with DO and positively with NH3-N and TP. Average slope gradient correlated positively with DO at 500 m buffer scale in 1996, given that fast-moving water at steeper slope gradients promotes water aeration and oxygen saturation (Chang 2008). Increased deviation of slope gradient also indicated more NH3-N (at buffer scale in 1996 and 2003) and TP (at buffer scale in 2003), because higher slope variability could accelerate erosion and subsequently increase rates of soil particles entering the riverine system (Chang 2008). The influence of soil texture was only pronounced for TP at the 500 m buffer scale. Higher percentage of clay, in particular, can lead to more concentrations of phosphorus in aquatic systems. Several properties of soils could account for these results. For example, clayey soils with fine particles, when wet, can become compacted easily and then accelerate runoff in red soils in close proximity to the river (Hilliard and Reedyk 2000). The runoff could contain high concentrations of phosphorus.

Table 2 Coefficients obtained by spatial regression for dissolved oxygen and nutrients patterns in 1996 and 2003 at different scales (n = 41)a
Full size table

Built-ups and farmland correlated negatively with DO and were positively correlated with NH3-N and TP at certain scales (Table 2). These findings were consistent with the point that farmland was a dominant predictor for declining water quality (Johnson et al. 1997; Castillo 2010; Tu 2011). Water bodies were correlated positively with NH3-N at the 500 m buffer scale in 2003, suggesting that water bodies could also be significant indictors of increased nitrogen pollution. The large portion of artificial aquaculture ponds could account for such results. The input of substantial quantities of feeding stuffs usually contained high concentrations of protein. In addition, these ponds served as important recipients of wastewater discharged by human activities which may become nitrogen pollution sources through surface runoff and ground water discharges. Forest areas were shown to protect river water quality since forests were associated positively with DO (at buffer scale, 2003) and negatively with TP (sub-basin scale, 1996; buffer scale, 2003). This may be linked to forest ecological functions of filtering pollutants during runoff events (Jarvie et al. 2002; Baker 2003; Tu 2011).

Population density and GDP were also incorporated in most models. Our results supported the argument that nitrate concentration and organic contamination levels are related to regional population and economic development (Peierls et al. 1991; Tu 2011). The quality of life for people in Qiantang basin has been greatly improved in the recent past. Demand for eggs, meat and milk increased greatly as population increased. Given the balance between feed and nitrogen inputs, livestock waste discharged more nitrogen loads accordingly. Crop production, energy production and human waste also increased simultaneously, which also contributed to more nitrogen loads. Increased organic loads from people and agriculture may result in the negative associations between DO and GDP and population density.

Determinants of spatial dynamics for DO and nutrients

Natural factors like topography and soil properties remained constant in such a short time period. Changes of anthropogenic determinants should account for the dynamics of DO and nutrients in our study. All regression models incorporated no more than two variables (Table 3). This can be attributed to the fact that changes of social-economic indicators always exhibit significant correlations with LULC change in most cases (Su et al. 2011c). Rapid social-economic development and the subsequent construction of built-ups were the main factors accounting for the dynamics of DO and nutrients. Comparing the coefficients of riparian forest for NH3-N and TP, lower percentage of riparian forest indicated higher NH3-N but lower TP. Such results were inconsistent with the point that riparian forest protected river water environment. The high proximity between forest land and livestock farms may account for this result. Polluted wastewater from farms, if not properly treated, may overflow into neighboring forests (Jung et al. 2008). Additional considerations should be given to forest species distribution and forest management options. Economic forests (e.g., nursery, chest nut trees and bamboo forest) occupied a large proportion of total forest area. Forest ecosystems may become highly stressed and dysfunctional as a result of excessive fertilization in seeking high profit.

Table 3 Determinants of dissolved oxygen and nutrients dynamics between 1996 and 2003 at different scales (n = 41)a
Full size table

Methodological discussion

Incorporation of spatial autocorrelation

Dissolved oxygen and nutrients were spatially autocorrelated in this study. This spatial autocorrelation leads to exaggerated effects of covariates and results in incorrect estimations of corresponding determinants when applying traditional linear regression. In addition, important impacts from neighboring sample sites will be ignored, thereby limiting our understanding of the spatial impact of these determinants. Specifically, spatial lag was incorporated for predicting the dynamics of NH3-N and TP at both scales and DO at buffer scale. These results indicated that the spatial dynamics of DO and nutrients depended not only on local independent variables but also on the dynamics and independents observed in neighboring sites. Conversely, spatial error regression was powerful in explaining the dynamics of DO at sub-basin scale. This implies that determinants of DO dynamics excluded from the model are spatially autocorrelated at sub-basin scale. Excluded factors may be climate, hydrology or management. We also note that the predictive ability of spatial regression for characterizing the dynamics of DO and nutrients was still relatively limited since the R 2 values were generally below .5. Other potential factors could account for unexplained variations.

Scale effects and management implications

This study highlighted scale effects from two perspectives: corresponding determinants and the predictive ability of spatial regression. Determinants of DO and nutrients differed by scale (Tables 2, 3). For example, built-ups had dominant control of spatial patterns of DO at sub-basin scale, while at the 500 m buffer scale the influence of farmland or forest was more significant. Slope gradient exerted a significant influence on the spatial patterns of TP at buffer scale rather than sub-basin scale. Spatial regression was more powerful at the buffer scale for predicting the spatial dynamics of DO and nutrients (Table 3). Whether sub-basin is a better spatial scale than buffer scale for predicting the spatial patterns of DO and NH3-N remains difficult to determine. Scale effects have been demonstrated in other studies. Guo et al. (2010) reported that the impact of LULC on TP changed with buffer width in a dynamic complex form. Zhang (2011) confirmed this observation by reporting a dynamic riparian width, within which LULC significantly governed riverine nitrogen loads. The influential distance for build-ups was much wider than that for farmland and forest (Zhang 2011). These findings illustrate the importance of multiple scale approaches for interpreting the spatial determinants of DO and nutrients.

The regression model developed in this study can be used to predict DO and nutrients where they are not monitored officially. In addition, the identified potential determinants can help to adjust current monitoring and management practices. For example, we can optimize land-use practices within buffer zones by reference to strong relationships between land-use type and nutrient concentrations. A watershed approach to water quality management has always been advocated, and the use of buffer zones can help improve the effectiveness of management practices (Guo et al. 2010; Zhang 2011). Therefore, we can integrate further a watershed approach with buffer zones to control or mitigate eutrophication by comparing the identified spatial determinants at sub-basin and buffer scales. The present methodology for the application of spatial regression to characterize and analyze determinants of DO and nutrients is designed for, but not limited to, its application for other water quality variables at other spatial scales.

Limitations discussion

Several limitations exist in the methodological approach of this study. First, we investigated only the determinants of DO and nutrients at two spatial scales. The relative importance of different determinants and their changes with spatial scale remained unknown. Secondly, seasonal dynamics were not analyzed given the unavailability of corresponding data for potential determinants. Thirdly, the number of samples is relatively limited, which could impact estimates of MLE. In addition, the limited number of samples also led to the shortcoming that we did not incorporate temporal and spatial dynamics simultaneously. A dynamic panel estimator can be applied, but only when the data are available. Fourthly, we did not compare the effects of different spatial matrices in spatial regression. Lastly, the resolution of satellite images as well as the method used for mapping population density and GDP may also affect the results. We plan to conduct additional studies that focus on these limitations.

Conclusions

This study analyzed the determinants of DO and nutrients at different spatial scales between 1996 and 2003 in Qiantang River, China. Our main conclusions are as follows:

  1. 1.

    Spatial patterns of DO and nutrients presented significant autocorrelation. Nitrogen pollution remained a regionalized problem, but phosphorus pollution has become a more localized problem.

  2. 2.

    Built-ups, farmland, water body, population density and GDP were positive indictors of nutrient pollution and hypoxia, while distance to river source and forest were negative indicators. Higher slope gradient variability can lead to more DO and nutrient pollution.

  3. 3.

    Population growth and economic development, and the subsequent built-up of urban land accounted for DO and nutrients dynamics to a large extent. Increases in farmland can result in more NH3-N pollution and hypoxia. Declines in riparian forests can lead to more NH3-N but less TP.

  4. 4.

    Incorporation of spatial autocorrelation into regression analysis was a methodological advantage, and scale effects should be given special attention. Application of spatial regression serves as a promising environmental management tool.