Network parameters quantify loss of assemblage structure in human-impacted lake ecosystems

2.1 Network conceptualization

We conceptualize an ecological network as a biological assemblage living within an abiotic environment (Figure 1), in which the species and their microhabitats represent separate groups of nodes (Figure 1a). The hypothetical network comprises a large and diverse assemblage of species inhabiting a lake with multiple microhabitats. Species degree is defined for each species by the number of other species with which it shares a microhabitat (Figure 1b). Species are thus associated by virtue of their co-occurrences, which are assumed to issue from sharing microhabitats (Strogatz, 2001). We model a network of non-random associations amongst species that share different microhabitats within lakes.

The basic model makes no assumptions about the existence of direct interactions between species, such as competition or facilitation, which are known to be poorly represented by association networks (Barner, Coblentz, Hacker, & Menge, 2018). Although species co-occurrence may be influenced by direct competition, mutualism or commensalism, most non-random co-occurrence in diatoms are likely to arise from co-dependency on environmental features (e.g. water depth) and habitat characteristics (e.g. presence/absence of macrophytes). Thus, diverse microhabitats support a heterogeneous assemblage (Figure 1c) and similar microhabitats support an assemblage occupying similar resource niches (Figure 1d). The network in a heterogeneous environment is characterized by an assemblage of weakly connected species (Figure 1c). A homogenous environment, in contrast, supports clustering of highly connected species (Figure 1d). A shift in boundary conditions caused by exogenous impacts, such as nutrient loading, drives the introduction of new mechanisms, such as positive feedback controls, which modify the clustering of species and consequently the pattern of their degree distribution.

2.2 Outline of methodology

We analyse spatio-temporal changes in the frequency distribution of species associations on a large database of 452 diatom species existing in the surface (modern) sediments of 273 lakes distributed across China. The lakes occupy the Lower Yangtze river basin, the Yunnan-Sichuan Province region and the Tibetan plateau. The strongest impact of human activities across these regions is cultural eutrophication driven by nutrient enrichment from aquaculture, agriculture, deforestation and sewage disposal. As a result, many lakes suffer algal blooms giving rise to turbid water and degraded food webs. For all 452 diatom species occurring in near-modern samples across all 273 lakes, we defined the set of ‘connected pairings’ as the set of most frequently co-occurring species pairs. For each diatom in an assemblage of a single lake, we calculated its species degree from the sum of its connected pairings found in the lake (Figure S1).

For each lake, we calculated three network parameters from its constituent diatoms. Skewness in species degree, S, is the asymmetry in the frequency distribution of co-occurrences between species (showing in Figure 1c,d between positive and negative skew). The heterogeneity index, H, measures the uniformity in the species degree (Gao et al., 2016; May, 2006). The clustering coefficient, C, is the realized fraction of all possible co-occurrences between species linked to a common network (Watts & Strogatz, 1998). The diatom networks for the calculation of these parameters are unweighted, meaning that they account only for the presence/absence of linkages and ignore linkage strength. In order to evaluate the impacts of species abundance on network properties, we also calculated the weighted clustering coefficient (WC) of a hypothetical weighted network. In the weighted network, it is assumed that the pairs of diatoms are directly linked because of species competition for resources. Thus, the linkage from one species to the other linked species is weighted by the species relative abundances. We also developed a null simulation model to test whether similar results are produced through random associations, as opposed to being driven by cultural eutrophication.

Additionally, for five lakes that have contrasting levels of contemporary and historical human impact (Table 1, Supporting Information Appendix – Lake details and Figure S2), we reconstructed the species degree for diatoms preserved in sediment cores covering the past 100–200 years. Across China, the strongest impacts occurred in the 20th century with increasing use of artificial fertilizer and to a lesser extent untreated sewage, associated with a doubling of human population density (Le et al., 2010). We examine the responses of network parameters and diatom richness to the intensity of human impact using: (a) modern population density data; (b) measured concentrations of total phosphorus (TP) in lakes; and (c) the known historical records of disturbance in the five focal lakes (Supporting Information Appendix – lake details).

2.3 Lake sampling and survey

The 273 investigated lakes represent a wide range of climate zones (humid subtropical, semi-arid and high alpine), altitude (>5,000 m to <50 m above sea level), land use (high mountains, extensive grazing, intensive agriculture), lake type and geomorphological setting (freshwater, brackish, tectonic, oxbow and glacial lakes with some reservoirs) across China. At each lake site, we defined the main catchment land use from field observations and systematic analyses of land-cover maps as urban, intensive grazing or agriculture. Catchments with no significant areas of these land use types were classified as ‘no observable human impact’. Of the 273 lake catchments, 116 had clear evidence of human impact in the catchment, in the form of urban areas, intensive grazing and/or agriculture. Water samples were collected from the surface of each lake for chemical analysis. TP was determined in the laboratory using standard techniques (Institute of Hydrogeography & Engineering Geology, 1990). Short sediment cores were collected from the deepest parts of each lake using a Kajak gravity corer (Supporting Information lake details, Renberg, 1991). The uppermost 0.5 cm of each sediment core was taken as representative of the contemporary diatom assemblages in each lake. Five cores from the selected lakes (Table 1) were sliced in the field at 0.5 or 1 cm intervals (details in Supporting Information) and stored at <4°C until further analysis in the laboratory. Data for samples from the Lower Yangtze basin and the Tibetan plateau are taken from Yang, Anderson, Dong, and Shen (2008), Wang, Yang, Langdon, and Zhang (2011) and Yao (2011). Samples from the Yunnan-Sichuan Province were collected during summers in the period 2007–2012.

2.4 Sediment dating

Sediment ages on older core samples were determined (Appleby, 2002) from the activities of ²¹⁰Pb, ²²⁶Ra and ¹³⁷Cs radioisotopes measured with gamma spectrometry, using a well-type coaxial low background intrinsic germanium detector (Ortec HP Ge GWL series). Supported ²¹⁰Pb levels in each sample were assumed to be in equilibrium with the in situ ²²⁶Ra, and unsupported ²¹⁰Pb (²¹⁰Pb_exc) was estimated by subtracting ²²⁶Ra activity from the total ²¹⁰Pb activity. The onset of ¹³⁷Cs activity (~1954) and peak values (1963) derived from atmospheric nuclear weapon testing were used to validate the ²¹⁰Pb chronology. The dates of samples were calculated using the constant rate of supply model (Appleby, 2002).

2.5 Diatoms

All diatom sample preparations and identifications were made by authors from the State Key Laboratory of Lake Science and Environment, Nanjing Institute of Geography and Limnology, Chinese Academy of Sciences, using the same set of taxonomic references. Diatom sample preparations followed standard procedures (Battarbee et al., 2002). For surface samples, around 500 diatom valves were counted from each lake except for a few lakes with sparse diatom valves where the count was reduced to ~300 valves. For down-core fossil diatoms, at least 300 valves were counted from each sample. Nomenclature and taxonomy mainly followed Krammer and Lange-Bertalot (1988a, 1988b, 1991a, 1991b, 2000). Only diatoms identifiable to species level were counted, giving a total of 452 diatom species. All identified species, including rare species, were included in the network analyses.

2.6 Diatom network

Each lake diatom assemblage was defined in terms of the presence or absence in the lake of 452 species. The strength of interspecific associations was calculated on 2 × 2 contingency tables for all possible pairings in the data set. The association coefficient, , for each pair of species i and species j is quantified by Cramér's V:

(1)

where a is the number of lakes with both species i and species j, b is the number of lakes with species i but not species j, c is the number of lakes with species j but not species i, and d is the number of lakes with neither species i nor species j. Values of V range between +1 (strong association) and −1 (strong avoidance). The analysis of network parameters used only the set of most highly associating pairings (the ‘connected pairings’), given by the upper quartile (Q3) of positive V (V+). This set contained 23,443 pairings amongst all 101,926 pairings between the 452 species. We also checked the effects of alternative V+ criteria on network parameters by using Q2 and Q1 pairings.

2.7 Network parameters

Skewness (S) is measured on the frequency distribution of species degree. For each species i in the sediment sample of a lake, species degree is defined as the sum of its Q3 associations in the assemblage of that lake. For example, a species that shares a lake with three other species with which it frequently co-occurs in the national data set according to the Q3 criterion will have species degree x = 3. Each species in a sediment sample from a lake contributes a value of to the species degree vector for that lake. Network skewness is then calculated from the vector by the Fisher–Pearson coefficient (Joanes & Gill, 1998):

(2)

where is degree of species i for Q3 connected pairings in a lake, and and n are the mean degree and species richness in the lake. Species with no links were excluded from the network skewness calculations.

Heterogeneity (H) is measured on the frequency distribution of species degree. For an undirected diatom network in a lake, H = σ/<s>, where σ is the standard deviation of the frequency distribution of species degree, and <s> is the average species degree in the network. The calculation of H is modified from Gao et al. (2016) by taking the square root of the numerator to avoid the effects of richness across lakes.

As with the local clustering coefficient (CLC) developed by Watts and Strogatz (1998), the clustering coefficient (C) measures how close each node (species) and its neighbours are to a complete subgraph (i.e. a group of species all of which are linked to each other.) For each species in a lake, the local CLC is the proportion of actual links between its neighbours divided by the number of links that could possibly exist between them. For a whole lake, C is the average of the local CLC of all the species.

The calculation of WC follows Barrat, Coblentz, Hacker and Menge (2004). Given a species A that links to species B, and the relative abundance of A is a% in a lake and the relative abundance of B is b% in the lake, we assume that the link strength from A to B is weighted by a%, and the link strength from B to A is weighted by b%. Thus, for a single lake, we can have a direct network with weighted links by species relative abundance.

2.8 Simulated null distribution

The distribution of associations could feasibly be the product of random processes or naturally occurring species distributions, instead of human impact and nutrient loading. Accordingly, we hypothesize that the distribution of species degree is determined only by the sizes of species pools in the three regions containing sampled lakes. To test the spatial pattern of skewness in the diatom network under this null hypothesis, we randomly assigned species to the same 273 sampled lakes, each retaining its empirical richness. Within each region, all of the lakes had incidence probability for any given species set by the size of its regional species pool as a fraction of the total pool of 452 species. We repeated the simulation 1,000 times and compared the mean of the replications with human population density around each lake.

3.1 Trends in network parameters with human impacts

Skewness values of species associations for all lakes range from +1.8 to −3.0 (Table S1), with 53% of all lakes having positive S and 17% having S values ≥+0.5. Broadly, S shifts from positive in the west to negative in the east (Figure 2a). We found no correlation of S with gradients in altitude or temperature (Figure S3). Instead, S links strongly to population density (Figure 2b): being generally positive in sparsely populated regions of the Tibetan plateau and Yunnan-Sichuan Province, and negative in densely populated regions, especially in the Lower Yangtze basin. In a multiple linear regression model (Table S2), only population density influences S (p < .001), with no detectable effects of temperature, elevation or lake size. This result shows that for every 100 people/km² rise in population size, there is a 0.05 decline in S values. Of the lakes with no observable human impact (no urban, intensive grazing or agriculture) in the catchment, 66% show positive S and 25% show S values above 0.5. Of the lakes with observable human impact in the catchment, in contrast, 25% show positive S and only 5% show S values above 0.5. The tendency for S values to reflect human impact is borne out by a clear negative trend of S with rising human population at densities above 50 people/km² (Figure 2b).

The other network parameters, C and H, correlate strongly with S and each other, reflecting their measurement of the same overall network response (Figure S4). Skewness has a strong negative correlation with C and a strong positive correlation with heterogeneity H. Values for C and H mirror the spatial pattern of S, with low values for C and high values for H in the less impacted and low population density regions and high values for C and low values for H in the heavily impacted and high-density regions (Figure S5). To simplify presentation, we will focus principally on network skewness (S), and consider C and H as complementary network parameters.

3.2 Trends in network parameters with lake nutrient levels

In lakes with observable human impact, S values respond strongly to TP values with positively skewed lakes mainly located below a concentration of 0.031 mg/L (Figure 3). This is consistent with an OECD classification (http://www.eulakes-model.eu/models/oecd.html) of >0.035 mg/L for eutrophic lakes. Even within the higher range of TP values from 0.048 to 0.090 mg/L, lakes with human impacts tend to show lower S values than lakes with no human impact (Figure S6), suggesting that cultural, rather than natural, eutrophication causes low S values. For the five lakes with temporal data covering the past 100+ years, S values become more negative and more variable with higher cumulative levels of human impact (Figure 3b), a pattern repeated for H and C (Figure S7). The network parameters thus reveal more than differences between lakes in levels of modern human impact; they also reflect the accumulation of human impacts on the network structure of a single lake.

3.3 Trends of species richness with human impacts

Species richness is easy to measure and its loss has often been used to indicate environment degradation (McCann, 2000). Since richness measures the number of nodes in a network, it may seem reasonable to consider it as an indicator of network structure. In this study, however, the variation between lakes in diatom species richness, from 6 to 65 species, with >15 species in 87% of lakes, shows no correlation with the magnitude or direction of S (Figure 4a,b). Species richness, moreover, has no clear trend with TP, except at very high concentrations (Figure 4c), nor with the gradient of human impact in the five lakes (Figure 4d).

3.4 Other network parameters

Although network structure depends on additional network parameters including the first two moments of species degree: the mean and variance, they are less informative as measures of human impacts than S, C or H. The mean species degree correlates negatively with network skewness (Figure 5a), making it a candidate as an alternative network parameter. However, both the mean and variance of species degree are strongly influenced by species richness (Figure 5c,d), which is a poor predictor of human impact in this study (Figure 4). Hence, we also expect the same poor predictive power of the mean and variance of species degree. We do not have the same a priori low expectation of the network parameters because they vary independently of richness (Figure 4a). Moreover, S, C and H link explicitly to environmental heterogeneity by definition.

It is possible that environmental forcing could affect the relative abundances of species in an assemblage, which then biases the calculation of network parameters based on species association. However, we reject this possibility after comparing species degree to species relative abundance in the national data set and in each lake. We found no detectable relationship in the national data set (Figure S8), and only 19 of 273 lakes have detectable linear correlations (at p < .05, of which only 3 at p < .01). The network S in an assemblage additionally shows no strong evidence for a relationship with species evenness (Figure S9). We also found that WC shows the same patterns as S and C across the spectrum of human activities given by spatial variation in population size and lake TP level (Figure S10). These results provide evidence for the independence of network parameters as metrics for change in assemblage structure under environmental forcing.

3.5 The effects of V+ criterion on network parameters

We tested whether the observed patterns depend on our choice of Q3 of V+ for defining the threshold for most highly associating species. Alternative thresholds of Q1 and Q2 pairings still show strong correlation between TP and S (Figure S11a,b). Thus, a less stringent criterion for highly associating pairings gives similar spatial patterns. Q1 to Q2 thresholds generate larger sets of connected pairings where some otherwise moderately associated species become classified as strongly associated species. Thus, a Q3 threshold maximizes the discrimination of strongly associated species from moderately and weakly associated species.

3.6 Simulated skewness in the national data set

Compared to the empirical data set, the set of associations from the simulated null distributions gave a lower range of S values, from around +1.7 to −0.9, and a much less negative minimum value. Skewness in Q3 of V+ from this null model shows no positive correlation with the S decline observed in the empirical data set (Figure S11c). In contrast to the empirical data (Figure 2b), there was no trend from low population density area to high population density area (Figure 6b). We analysed the distribution of cosmopolitan and endemic species in each region, and found no detectable differences (Figure S12). We therefore find no evidence that the observed S pattern in diatoms across China was due to random species associations. Additionally, we recalculated S using regional, rather than national diatom data sets. For the Yunnan-Sichuan and Lower Yangtze River data sets, the negative relationship persists between S and TP (Figure S13). Taken together, these findings indicate that the network parameter patterns are neither random nor natural.

3.7 Network parameters changes over time

We can explore further the controls on the diatom network parameters by comparing the changes over time with known events in the catchments of the five lakes (Figure 7; Figure S14). Shade (Figure 7a) and Moon (Figure 7b) are isolated alpine lakes (Table 1; Figure S2) with consistently positive S values since the early 1900s (Figure S15a,b). For both lakes, the best-fitting model for the cumulative probability distribution of species degree is a truncated power law, rather than exponential or power law (Figure S16). Although the diatom assemblages may have been affected by distal human activities since the early 1900s, such as increased nitrogen deposition (Hu, Anderson, Yang, & McGowan, 2014), the catchments show no discernible human impacts (Supporting Information Appendix – Lake details). Both lakes had detectable species turnover since the 1800s (Figure S17), but with neither lake expressing a trend over time in network parameters (Figure S17), the changes in species composition appear not to have resulted in any substantial changes in functional composition sufficient to shift the assemblage structure.

In contrast, Lakes Longgan (Figure 7d) and Taibai (Figure 7e) have a long history of reclamation, fishing, polderization and intensive agriculture. Recent (2014) water analyses show Longgan having TP concentrations around 0.050–0.080 mg/L, and profuse macrophytes, while Taibai has higher TP at around 0.180–0.210 mg/L and few macrophytes. The partial recovery of Longgan since the mid-20th century is reflected in gains in low- and middle-degree species (<5) in the two most recent periods (Figure. S15d). The heavily impacted Taibai (Figure S15e) shows the most negatively skewed distributions overall, with fluctuations in negative S (Figure 7) caused by changes in relatively high- (>20) degree species rather than changes in lower degree species. These lakes have network parameters with considerably higher magnitudes, in negative S and two to three times larger values of C (Figure S14d,e), with strengthening trends in recent decades. The generally weaker values for Longgan than Taibai suggest that it might be the less impacted lake. This is borne out by its history of water regulation and protection measures since the 1940s, with no corresponding measures at Taibai (Supporting Information Appendix – Lake details). The higher human impacts in Taibai result in a homogenous diatom network (Figure S14j).

Lake Erhai bridges between the least and most impacted lake pairs (Figure 7c; Figure S14). Its skewness in the early 1900s took positive values commensurate with those of Shade and Moon lakes, and subsequently reduced to negative values similar to those of Longgan lake. The relatively low levels of low-degree (weakly connected) species throughout the 20th century suggests that the lake has been impacted throughout the whole period (Figure S15c). However, the sequence of frequency distributions (Figure S15c) suggests that increasing negative skewness was driven by losses of middle- and high- (>8) degree species, the latter especially after 2001, rather than progressive losses of low-degree species. Likewise, values of C increased through time, and since 2001 reached some of the largest values of any of the lakes. These changes reflect the intensification of human activities through the 20th century culminating in a critical transition from mesotrophic to eutrophic conditions in 2001 (Wang et al., 2012; see also Supporting Information Appendix – Lake details). Network parameters calculated on samples at nearly equal time intervals suggest that variable sediment accumulation rates (e.g. through compaction) do not bias these conclusions (Figure S18).

Network S has a strong negative correlation with C in the three heavily impacted lakes, and a much weaker correlation in the two pristine lakes (Figure S19). This strengthening of the negative correlation in highly impacted lakes is consistent with the general association across all 273 lakes of low and negative values of S with high values of C (Figure S3b). It is also consistent with the extremely high values of C in lake Erhai immediately following a regime shift (Figure S14c) caused by resilience loss (Wang et al., 2012).