Response of Water Quality to Landscape Patterns in an Urbanized Watershed in Hangzhou, China

: Intense human activities and drastic land use changes in rapidly urbanized areas may cause serious water quality degradation. In this study, we explored the e ﬀ ects of land use on water quality from a landscape perspective. We took a rapidly urbanized area in Hangzhou City, China, as a case study, and collected stream water quality data and algae biomass in a ﬁeld campaign. The results showed that built-up lands had negative e ﬀ ects on water quality and were the primary cause of stream water pollution. The concentration of total phosphorus signiﬁcantly correlated with the areas of residential, industrial, road, and urban greenspace, and the concentration of chlorophyll a also signiﬁcantly correlated with the areas of these land uses, except residential land. At a landscape level, the correlation analysis showed that the landscape indices, e.g., dominance, shape complexity, fragmentation, aggregation, and diversity, all had signiﬁcant correlations with water quality parameters. From the perspective of land use, the redundancy analysis results showed that the percentages of variation in water quality explained by the built-up, forest and wetland, cropland, and bareland decreased in turn. The spatial composition of the built-up lands was the main factor causing stream water pollution, while the shape complexities of the forest and wetland patches were negatively correlated with stream water pollution. concentration Cyanophyta (Chla Cyan ), chlorophyll a concentration of Chlorophyta (Chla Chlo ), and chlorophyll a concentration of Bacillariophyta and Dinophyta (Chla Baci-Dino ), using a four-wavelength-excitation chlorophyll ﬂuorometer (PHYTO-PAM Fa. Walz, E ﬀ eltrich, The water samples were collected 0.3–0.5 m below the water surface from the middle of the stream with a 250 mL organic glass hydrophore, and three parallel samples were collected at each sampling site to avoid accidental errors. The samples were kept in iceboxes during transport to the laboratory. The TP and TN were determined using the alkaline potassium persulfate digestion UV spectrophotometric method (GB11893-89, China National Standards) and the ammonium molybdate spectrophotometric method (GB11894-89, China National Standards), respectively. According to the Environmental Quality Standards for Surface Water (EQSSW) (GB3838-2002, State Administration of China), the water quality could be classiﬁed into ﬁve classes by a series water quality parameters, including TN and TP, and the higher the class, the worse the water quality.


Introduction
Land use change is one of the key driving forces of global change and terrestrial environmental changes [1][2][3]. Land use may exert considerable influence on stream environments by altering multiple ecological processes, e.g., hydrological cycle, soil erosion, and nutrient migration and transformation [4,5]. Point and non-point source pollutants are the main contributors to stream water pollution. Non-point source pollution still affects stream water quality in developed regions where the point source pollutions are usually under effective control, and almost all non-point source pollutions are closely related to land use change [6][7][8][9][10][11][12]. At the watershed scale, the influence of land use on stream water quality is of particular interest [13,14]. The water quality in a particular watershed is determined by multiple environmental factors, as the migration and transformation of pollutants at the land surface are closely associated with the local environment (e.g., vegetation and soil) and human activities (e.g., urbanization). In this process, the landscape pattern is a critical factor to understand the spatial variation of stream water quality [15,16].
The land use information was extracted from one cloudless SPOT 6 image obtained on 26 October 2014. The multispectral and panchromatic bands of the SPOT image were fused to get a 1.5 m resolution image using the Gram-Schmidt method [43]. The SPOT image was radiometrically and atmospherically corrected, and the accuracy of geometric correction was under 0.5 pixels. The Level I land use classes were extracted by supervised classification method with a total accuracy of 91.49% and that kappa coefficient of 0.88, and then further subclassified into corresponding Level II classes using visual interpretation with the aid of ancillary data, e.g., ASTER DEM, field survey data, and Google Earth satellite images [44] (Figure 2). The classification accuracy was assessed by stratified random sampling with reference to ground-truth data collected through field survey. The assessment results showed that the overall accuracy of land use classification was 93.23% and that the kappa coefficient was 0.92.

Land Use Classification
A two-level classification system was adapted from the Chinese Standard of Land Use Classification [42]. We classified the land use in the study area into five Level I classes, i.e., Cropland (1), Forest (2), Water (3), Built-up (4), and Bareland (5), and 12 Level II classes, i.e., Cropland (11), Forest (21), River (31), Pond (32), Residential (41), Industrial (42), Commercial (43), Public management and service (44), Urban greenspace (45), Road (46), Under construction (47), and Bareland (51), respectively. The land use information was extracted from one cloudless SPOT 6 image obtained on October 26, 2014. The multispectral and panchromatic bands of the SPOT image were fused to get a 1.5 m resolution image using the Gram-Schmidt method [43]. The SPOT image was radiometrically and atmospherically corrected, and the accuracy of geometric correction was under 0.5 pixels. The Level I land use classes were extracted by supervised classification method with a total accuracy of 91.49% and that kappa coefficient of 0.88, and then further subclassified into corresponding Level II classes using visual interpretation with the aid of ancillary data, e.g., ASTER DEM, field survey data, and Google Earth satellite images [44] (Figure 2). The classification accuracy was assessed by stratified random sampling with reference to ground-truth data collected through field survey. The assessment results showed that the overall accuracy of land use classification was 93.23% and that the kappa coefficient was 0.92.

Land Use Classification
A two-level classification system was adapted from the Chinese Standard of Land Use Classification [42]. We classified the land use in the study area into five Level I classes, i.e., Cropland (1), Forest (2), Water (3), Built-up (4), and Bareland (5), and 12 Level II classes, i.e., Cropland (11), Forest (21), River (31), Pond (32), Residential (41), Industrial (42), Commercial (43), Public management and service (44), Urban greenspace (45), Road (46), Under construction (47), and Bareland (51), respectively. The land use information was extracted from one cloudless SPOT 6 image obtained on October 26, 2014. The multispectral and panchromatic bands of the SPOT image were fused to get a 1.5 m resolution image using the Gram-Schmidt method [43]. The SPOT image was radiometrically and atmospherically corrected, and the accuracy of geometric correction was under 0.5 pixels. The Level I land use classes were extracted by supervised classification method with a total accuracy of 91.49% and that kappa coefficient of 0.88, and then further subclassified into corresponding Level II classes using visual interpretation with the aid of ancillary data, e.g., ASTER DEM, field survey data, and Google Earth satellite images [44] (Figure 2). The classification accuracy was assessed by stratified random sampling with reference to ground-truth data collected through field survey. The assessment results showed that the overall accuracy of land use classification was 93.23% and that the kappa coefficient was 0.92.

Watershed Delineation
A fine delineation of the watershed is crucial in evaluating the influence of land use on water quality in relatively small areas. The commonly used watershed delineation algorithms, e.g., D8 [45], burn-in [46], DEMOM [47], and Dinf [48], work well in plain areas with low impact of human activities. However, the road partition, river embankments, and buildings in complex urban areas could alter the original runoff directions [49]. One feasible solution to improve the accuracy of watershed delineation, in this case, is to combine these terrain features into the Digital Elevation Model (DEM) [50]. Considering that the study area is composed of piedmont plains with dense stream networks and complex built-up land uses, we combined the land use map and adjusted the DEM before watershed delineation using the D8 algorithm ( Figure 3). The primary goal of this operation was to adjust the elevation values of the terrestrial features, e.g., buildings, ditch, and greenspace, to more closely reflect the urban surface morphology, and, especially, the surface runoff direction, as the buildings will block natural runoff while the ditches and greenspace could divert and absorb surface runoff [51][52][53].

Watershed Delineation
A fine delineation of the watershed is crucial in evaluating the influence of land use on water quality in relatively small areas. The commonly used watershed delineation algorithms, e.g., D8 [45], burn-in [46], DEMOM [47], and Dinf [48], work well in plain areas with low impact of human activities. However, the road partition, river embankments, and buildings in complex urban areas could alter the original runoff directions [49]. One feasible solution to improve the accuracy of watershed delineation, in this case, is to combine these terrain features into the Digital Elevation Model (DEM) [50]. Considering that the study area is composed of piedmont plains with dense stream networks and complex built-up land uses, we combined the land use map and adjusted the DEM before watershed delineation using the D8 algorithm ( Figure 3). The primary goal of this operation was to adjust the elevation values of the terrestrial features, e.g., buildings, ditch, and greenspace, to more closely reflect the urban surface morphology, and, especially, the surface runoff direction, as the buildings will block natural runoff while the ditches and greenspace could divert and absorb surface runoff [51][52][53].

Water Sampling
Based on the watershed division and the spatial distribution of the stream network and land use, sampling points were selected to evenly cover each sub-basin. In total 19 sampling points were selected near the outlet of each sub-basin. The water sampling was conducted within one week, when the stream flow was relatively stable during the local high water period in June 2014 ( Figure 3). Aside from the total nitrogen (TN) and total phosphorus (TP), we also measured the concentration of chlorophyll a (TChla), including chlorophyll a concentration of Cyanophyta (ChlaCyan), chlorophyll a concentration of Chlorophyta (ChlaChlo), and chlorophyll a concentration of Bacillariophyta and Dinophyta (ChlaBaci-Dino), using a four-wavelength-excitation chlorophyll fluorometer (PHYTO-PAM Fa. Walz, Effeltrich, Germany). The water samples were collected 0.3-0.5 m below the water surface from the middle of the stream with a 250 mL organic glass hydrophore, and three parallel samples were collected at each sampling site to avoid accidental errors. The samples were kept in iceboxes during transport to the laboratory. The TP and TN were determined using the alkaline potassium persulfate digestion UV spectrophotometric method (GB11893-89, China National Standards) and the ammonium molybdate spectrophotometric method (GB11894-89, China National Standards),

Water Sampling
Based on the watershed division and the spatial distribution of the stream network and land use, sampling points were selected to evenly cover each sub-basin. In total 19 sampling points were selected near the outlet of each sub-basin. The water sampling was conducted within one week, when the stream flow was relatively stable during the local high water period in June 2014 ( Figure 3). Aside from the total nitrogen (TN) and total phosphorus (TP), we also measured the concentration of chlorophyll a (TChla), including chlorophyll a concentration of Cyanophyta (Chla Cyan ), chlorophyll a concentration of Chlorophyta (Chla Chlo ), and chlorophyll a concentration of Bacillariophyta and Dinophyta (Chla Baci-Dino ), using a four-wavelength-excitation chlorophyll fluorometer (PHYTO-PAM Fa. Walz, Effeltrich, Germany). The water samples were collected 0.3-0.5 m below the water surface from the middle of the stream with a 250 mL organic glass hydrophore, and three parallel samples were collected at each sampling site to avoid accidental errors. The samples were kept in iceboxes during transport to the laboratory. The TP and TN were determined using the alkaline potassium persulfate digestion UV spectrophotometric method (GB11893-89, China National Standards) and the ammonium molybdate spectrophotometric method (GB11894-89, China National Standards), respectively. According to the Environmental Quality Standards for Surface Water (EQSSW) (GB3838-2002, State Environmental Protection Administration of China), the water quality could be classified into five classes by a series water quality parameters, including TN and TP, and the higher the class, the worse the water quality.

Landscape Metrics
There are numerous landscape metrics that quantitatively represent the spatial composition of a particular landscape pattern [20,22,23,[54][55][56]. In this study, we used five categories of metrics, i.e., area-edge, subdivision, contagion/intersection, diversity, and shape, to quantify the landscape pattern and explore the corresponding associations with water quality parameters ( Table 1). The landscape metrics at the landscape and class levels were analyzed by the software FRAGSTATS 4.1.

Statistical Analysis
The Spearman rank correlation was used to analyze the relationships between land use and water quality parameters. We used the Spearman rank correlation coefficient due to the nonparametric nature of the data, as the Shapiro-Wilk test of normality showed that most of the water quality and land use data did not follow normal distribution [57]. The multiple linear regression (MLR) was performed to evaluate the relationships between response (i.e., single water quality parameter) and predictors (i.e., land use metrics) after Spearman rank correlation analysis. To reduce the redundancy associated with the correlated variables, a stepwise regression approach based on the p values was chosen to eliminate insignificant predictor variables from the MLR models. The dependent and predictor variables were log-transformed to reduce the influence of the asymmetric distribution of the data [24,58]. The regression analyses were carried out with SPSS 16.0.
The constrained ordination methods, e.g., redundancy analysis (RDA) and canonical correspondence analysis (CCA), could well extract the variation in response variables that can be explained by a set of explanatory variables and have been widely used to analyze the relationships between land use pattern and water quality parameters [59][60][61][62]. The detrended correspondence analysis (DCA) showed that the maximum length of the gradient of the water quality data was less than three standard deviations, suggesting that the relationship between water quality and landscape metrics could be either linear or unimodal. Therefore, RDA was used to explore the relationship between water quality and landscape pattern parameters. The statistically significant landscape metrics in RDA were identified using a Monte Carlo permutation test (499 non-restrictive screening cycles) [63]. To simplify the interpretation of RDA results about the relationships between landscape pattern and water quality parameters, we further categorized the Level I land use classes (see Section 2.2) into four classes, i.e., cropland, forest, and wetland (by merging the forest and water classes), built-up, and bareland, from a functional perspective. The RDA was performed using the CANOCO 4.5 program.

Land Use Structure and Water Quality
Land use in each catchment is shown in Figure 4. The built-up land area accounted for nearly 50% of the whole study area. The areas of forest, cropland, and residential land covered 18.84%, 16.42%, and 16.42% of the study area, respectively. The areas of under construction, industrial use, and bareland occupied 9.42%, 7.19%, and 6.92% respectively, followed by the roads (5.53%), urban greenspaces (5.34%), ponds (4.99%), and the river (3.87%). The areas of public management and service, and commercial land, were the smallest with percentages of 3.20% and 1.86%, respectively. metrics in RDA were identified using a Monte Carlo permutation test (499 non-restrictive screening cycles) [63]. To simplify the interpretation of RDA results about the relationships between landscape pattern and water quality parameters, we further categorized the Level I land use classes (see Section 2.2) into four classes, i.e., cropland, forest, and wetland (by merging the forest and water classes), built-up, and bareland, from a functional perspective. The RDA was performed using the CANOCO 4.5 program.

Land Use Structure and Water Quality
Land use in each catchment is shown in Figure 4. The built-up land area accounted for nearly 50% of the whole study area. The areas of forest, cropland, and residential land covered 18.84%, 16.42%, and 16.42% of the study area, respectively. The areas of under construction, industrial use, and bareland occupied 9.42%, 7.19%, and 6.92% respectively, followed by the roads (5.53%), urban greenspaces (5.34%), ponds (4.99%), and the river (3.87%). The areas of public management and service, and commercial land, were the smallest with percentages of 3.20% and 1.86%, respectively. The statistics of the water quality and the corresponding spatial distributions are shown in Table  2 and Figure 5. The average concentrations of TN and TP were 6.54 mg/L and 0.62 mg/L, respectively, both exceeding the thresholds of TN and TP, i.e., 2.0 mg/L and 0.4 mg/L, of Class Ⅴ in EQSSW. The TN concentrations at 16% and 84% sampling sites were Class Ⅴ and beyond, respectively. The TP concentrations at 11%, 21%, 11%, and 53% sampling points were Class Ⅲ, Ⅳ, Ⅴ, and beyond, The statistics of the water quality and the corresponding spatial distributions are shown in Table 2 and Figure 5. The average concentrations of TN and TP were 6.54 mg/L and 0.62 mg/L, respectively, both exceeding the thresholds of TN and TP, i.e., 2.0 mg/L and 0.4 mg/L, of Class V in EQSSW. The TN concentrations at 16% and 84% sampling sites were Class V and beyond, respectively. The TP concentrations at 11%, 21%, 11%, and 53% sampling points were Class III, IV, V, and beyond, respectively; only No. 19 sampling point was Class II.   The average concentrations of ChlaCyan, ChlaChlo, and ChlaBaci-Dino were 9.90 μg/L, 68.44 μg/L, and 24.15 μg/L, respectively. In 63% of the sampling sites, no ChlaCyan were detected. The average TChla was 94.08 μg/L, much higher than the OECD eutrophication evaluation criteria [64].
The correlation analysis showed that TN was positively correlated with TP; TChla was positively correlated with TP; and ChlaChlo was positively correlated with TN, TP, and Tchla, respectively  The average concentrations of Chla Cyan , Chla Chlo , and Chla Baci-Dino were 9.90 µg/L, 68.44 µg/L, and 24.15 µg/L, respectively. In 63% of the sampling sites, no Chla Cyan were detected. The average TChla was 94.08 µg/L, much higher than the OECD eutrophication evaluation criteria [64].
The correlation analysis showed that TN was positively correlated with TP; TChla was positively correlated with TP; and ChlaChlo was positively correlated with TN, TP, and Tchla, respectively (Table 3).

Relationships Between Land Use Area and Water Quality
The correlation analysis between stream water quality and land use areas showed TP was positively correlated with the areas of residential, industrial, road, and urban greenspace ( Figure 6). The Chla Baci-Dino , Chla Chlo , and TChla were positively correlated with the areas of river, pond, and bareland; the Chla Chlo was positively correlated with the areas of road and urban greenspace; and the TChla was positively correlated with areas of road, urban greenspace, and industrial. Additionally, there were no statistically significant correlations between water quality parameters and the areas of cropland, commercial, public management and service, and under construction.
Sustainability 2020, 12, x FOR PEER REVIEW 10 of 19 Figure 6. Correlation coefficients between land use areas and water quality parameters based on Spearman's rank order correlation analysis. One asterisk indicates p-value < 0.05 and two asterisks pvalue < 0.01.
The stepwise multiple linear regression between water quality parameters and land use areas (Table 4) showed that industrial land was the dominant predictor of both TP and TChla concentrations, while bareland was the dominant predictor of algae biomass. Table 4. Regression of water quality parameters against land use area. Figure 6. Correlation coefficients between land use areas and water quality parameters based on Spearman's rank order correlation analysis. One asterisk indicates p-value < 0.05 and two asterisks p-value < 0.01. The stepwise multiple linear regression between water quality parameters and land use areas (Table 4) showed that industrial land was the dominant predictor of both TP and TChla concentrations, while bareland was the dominant predictor of algae biomass.

Land Use Pattern-Water Quality Relationships at Landscape Level
The correlation analysis between landscape pattern and water quality parameters showed that ( Figure 7) the landscape metrics, e.g., dominance (LPI), complexity (SHAPE, FRAC, PARA, and LSI), fragmentation (TE, NP, and DIVISION), aggregation and connectedness (CONTAG, IJI, and CONTIG), and diversity (PR and SHDI), all had significant correlations with water quality parameters. TP, TChla, and Chla Chlo , compared with TN, Chla Cyan , and Chla Baci-Dino , were much more sensitive to landscape metrics and negatively correlated with LPI, FRAC, and CONTAG, but positively correlated with SHDI. The Chla Baci-Dino were only negatively correlated with SHAPE and FRAC, two landscape metrics related to the shape of landscape patches. The Chla Cyan showed no significant correlations with the landscape metrics, so no further analysis of this parameter was carried out. The relationships between landscape pattern metrics and water quality parameters were also analyzed by the stepwise multiple linear regression modeling. As shown in Table 5, the PARA was the dominant predictor both for TN and TP; while SHDI was the dominant predictor both for TChla and ChlaChlo, and SHAPE was the dominant predictor for ChlaBaci-Dino. For the four land use classes, there were 2, 5, 7, and 5 explanatory variables with statistically significant relationships with water quality parameters, respectively ( Table 6). The total explained variations in water quality were 37.9%, 54.6%, 71.4%, and 35.9%, respectively, and more than 90% Figure 7. Spearman rank correlation between water quality parameters and landscape metrics. One asterisk indicates p-value < 0.05 and two asterisks p-value < 0.01.
The relationships between landscape pattern metrics and water quality parameters were also analyzed by the stepwise multiple linear regression modeling. As shown in Table 5, the PARA was the dominant predictor both for TN and TP; while SHDI was the dominant predictor both for TChla and Chla Chlo , and SHAPE was the dominant predictor for Chla Baci-Dino .

Land Use Pattern-Water Quality Relationships at Class Level
For the four land use classes, there were 2, 5, 7, and 5 explanatory variables with statistically significant relationships with water quality parameters, respectively ( Table 6). The total explained variations in water quality were 37.9%, 54.6%, 71.4%, and 35.9%, respectively, and more than 90% variations in water quality parameters could be explained by the first two axes. The built-up land use had the highest total explained variation. For all land use types, the total explained variations in water quality were 84.1%.  The RDA showed that the landscape patterns had different influences on water quality parameters ( Figure 8). The TN and TP were negatively correlated with the landscape shape metrics of the cropland and forest. In addition, the ordination also showed that TN and TP were affected by the landscape patterns of the built-up land uses. For example, TN and TP were positively correlated with PLANDs of the road and urban greenspace, PDs of the urban greenspace and under construction, and the ED of the road, as well as the COHE and CONTIG of the industrial land. The Chla Chlo and TChla showed similar responses to the landscape pattern; these two parameters were negatively correlated with SHAPE and FRAC metrics of cropland, and positively correlated with LSI, COHE, and CONTIG metrics of bareland and the COHE and CONTIG metrics of industrial land. The Chla Baci-Dino showed significant correlations with the landscape metrics related to the complexity of the forest and wetland patches; for example, the Chla Baci-Dino was positively correlated with the SHAPE and LSI metrics of the river, as well as the LSI of the pond and the SHAPE of the forest.
To be specific, water pollution, in terms of TN and TP, was positively correlated with the following landscape metrics, i.e., the connectedness of the industrial land, the dominance and fragmentation of the road and urban greenspace, and the fragmentation of the under construction land. On the contrary, the complexities of the cropland and forest patches were negatively correlated with TN and TP. The algae biomass, especially the Chla Baci-Dino , was positively correlated with the complexities of the river as well as the pond patches, and the connectedness of the river. The complexity of the bareland patch was also a positive indicator of the TChla.
By using both types of land use metrics, RDA showed that CONTIG and COHESION of the industrial land, COHESION and LSI of the bareland, COHESION and LSI of the river, LSI of the lake and pond, and SHAPE and FRAC of the cropland were the main factors dominating the variation in water quality.
negatively correlated with SHAPE and FRAC metrics of cropland, and positively correlated with LSI, COHE, and CONTIG metrics of bareland and the COHE and CONTIG metrics of industrial land. The ChlaBaci-Dino showed significant correlations with the landscape metrics related to the complexity of the forest and wetland patches; for example, the ChlaBaci-Dino was positively correlated with the SHAPE and LSI metrics of the river, as well as the LSI of the pond and the SHAPE of the forest. To be specific, water pollution, in terms of TN and TP, was positively correlated with the following landscape metrics, i.e., the connectedness of the industrial land, the dominance and fragmentation of the road and urban greenspace, and the fragmentation of the under construction land. On the contrary, the complexities of the cropland and forest patches were negatively correlated with TN and TP. The algae biomass, especially the ChlaBaci-Dino, was positively correlated with the

Water Pollution in Urbanized Watershed
The stream water in the study area was eutrophic with serious nitrogen and phosphorus pollution and high algae biomass ( Table 2). The percentage of impervious surface is an indicator that could well represent the intensity of urbanization, including its spatial expansion. Of all the influential factors in terms of water quality in urban area, stream water quality, and aquatic ecosystems could be damaged if the ratio of impervious surface reaches 10-15% in a specific watershed [65][66][67][68]. For our case, the percentage of the impervious surface was over 50%, much higher than the threshold mentioned above (Figure 4).
The built-up lands were important factors influencing water quality in the study area. For example, the TP was positively correlated with the areas of residential, industrial, road, and urban greenspace, and the TChla was also positively correlated with the areas of industrial, road, and urban greenspace. The TN was positively correlated with the areas of all the built-up lands, although none were statistically significant ( Figure 6). The increased built-up areas in urban area could result in a considerable increase of non-point source pollutants by precipitation runoff [69][70][71].
The biomass of Chla Chlo was positively correlated with TN and TP, but the correlations between Chla Baci-Dino and TN as well as TP were not statistically significant ( Table 3). The TChla, Chla Chlo , and Chla Baci-Dino positively correlated with the areas of the pond and river ( Figure 6). In addition to the absolute concentration of TN and TP, the growth of algae is also affected by factors like the ratio of nitrogen and phosphorus, water pH, and dissolved oxygen [72,73], and, therefore, the community structure of algae might not be simply explained by the nutrient concentrations in stream water alone. Urban greenspace, contrary to common sense, was positively correlated with Chla Chlo , TChla, as well as TP. We suggested that this phenomenon might be due to the urban greenspace in rapidly urbanized areas being not yet fully functional ecosystems and, additionally, the urban greenspace was mainly concentrated along with the road and was a major sink of vehicle emissions.

Influence of Landscape Patterns on Water Quality
During the urbanization, the landscape usually shows a trend of fragmentation, decentralization, and diversification with decreased complexity of the patch shape as the dominant landscape of the cropland or forest are replaced by various built-up land uses [25,[74][75][76]. We found that the dominance, fragmentation, complexity, and diversity of the landscape at the catchment scale were closely related to stream water quality (Figure 7).
It is hard to quantify the response of water quality to specific land use status without the corresponding landscape pattern of the target land use [56,77]. Indeed, different kinds of land uses had varied explanatory abilities to water quality. The landscape metrics of the built-up land had the highest explanatory ability, followed by the forest, water, cropland, and bareland ( Table 6). The results further showed that the spatial composition of the built-up land and the complexities of the forest, water, and cropland patches had close relationships with stream water quality. To be specific, the TN and TP were positively correlated with the connectedness of the industrial land, road area, and the fragmentations of urban greenspace and under construction lands were positive indicators of the TN and TP in the stream water in this study area.
There are numerous studies about the relationships between the dominance and fragmentation of the built-up land and water quality [22,24,78], but a more detailed classification of the built-up land is still needed to more accurately locate the problem. In this study, we classified the built-up land into residential, industrial, commercial, public management and service, urban greenspace, road, and under construction, and quantitatively analyzed the relationship between stream water quality and multiple landscape pattern metrics of all the built-up land uses using RDA. We found that a detailed classification of the built-up land could help more accurately identify the most influential land uses and the corresponding landscape patterns on water quality. For example, we found that the connectedness of the industrial land, dominances and fragmentations of the road and urban greenspace, and the fragmentation of the under construction land were correlated with the concentrations of TN and TP in the stream water. By identifying the influential land uses and the corresponding landscape pattern metrics, it might be possible to artificially influence water quality by optimizing the composition and spatial pattern of the specific land uses [79,80].
The forest may have a direct impact on the hydrological status of a specific catchment. The area, aggregation, and connectedness of the forest landscape were usually positively correlated with the abilities of fixation and retention of water pollutants, as well as the purification of surface runoff [81,82]. In this case, we found that the increased complexities of the forest and cropland landscapes were associated with the decreased TN and TP in the stream water, and this finding might be of interest to the forest or cropland planning practitioners.

Limits and Future Work
For studies about the relationship between water quality and land use/land cover changes at the medium or large watershed scales, series land use data and long-term water quality data are usually a prerequisite [14,71]. However, the hydrological stations and long-term water quality data at urban watersheds are usually quite limited. This study analyzed the response of stream water quality to the catchment landscape pattern using data collected during the high water period, but its applicability during the dry season or the flat water period still needs verification. Additionally, the land use status and water quality are constantly changing during the urbanization process [70,83], and it would be expected that regular monitoring of the land use change and water quality in urban areas could provide more effective guidance in urban land use planning and water environment management.
We showed the ability of landscape metrics to explain the variation of specific water quality parameters. However, many of the studies about the relationship between water quality and land use are case-specific and the selection of landscape metrics remains more or less subjective [84][85][86], so we suggested that it would be helpful to establish a common set of landscape metrics with clear ecological significance to improve the comparability across similar studies.

Conclusions
The feedback of water quality on landscape patterns was analyzed at the catchment scale in a typical rapidly urbanized area with a dense stream network. The stream water in the study area was characterized by eutrophic water with serious nitrogen and phosphorus pollution and high algae biomass. The built-up lands showed negative effects on water quality. The TP was positively correlated with areas of residential, industrial, road, and urban greenspace; except residential land, the TChla had similar statistical correlations with the areas of the other three land uses.
At the landscape level, the TP, Chla Chlo , and TChla were positively correlated with the decreased landscape dominance and shape complexity, and the increased landscape fragmentation and dispersity. At the class level, the percentages of the variation in water quality explained by the built-up land, forest and wetland, cropland, and bareland decreased in turn. The spatial composition of the built-up lands was the main factor causing stream water pollution, while the shape complexities of forest and wetland patches were negatively correlated with stream water pollution.
We suggested that, in order to improve water quality during rapid urbanization, it might be necessary to avoid the uncontrolled sprawl of construction land, especially of industrial land, and the excessive dispersion and fragmentation of farmland, and to improve the connectivity of rivers and the complexities of forests and wetlands.
Author Contributions: Y.S. and X.S. conceived and designed the framework of the research structure. Y.S., X.S. and G.S. contributed to the data analysis, interpretation, and manuscript writing. Y.S., X.S. and G.S. All authors have read and agreed to the published version of the manuscript.