2.1 Study area

Poland is located in Central and Western Europe covering a total area of 311,888 km². The range of mean temperature varies from 6°C in the northeast to 8°C in the southwest. Poland receives about 600 mm of precipitation on an annual average basis, with higher values reported for southern mountain ranges (where average annual precipitation can increase up to 1,300 mm) compared to the central plains. Agriculture and forest constitute most land cover categories with 63% and 32% coverage, respectively (Piniewski, 2017). There is no strong spatial gradient in land cover, however, agricultural areas and high production intensity dominate in the central, lowland part of the country (Kluba et al., 2016).

The hydrological regime of rivers in Poland is diverse: Nival regimes with high flows occurring in the springtime (March–April) is prevalent in the lowland rivers, while a nivo-pluvial regime with a maximum flow in spring and secondary high flow in summer occurs in the mountain rivers in the south (Piniewski, 2017). Some of the rivers in the Carpathian region have the highest flow occurring in summer and the secondary one in spring. Hydromorphological state classes tend to be lower in rivers of central Poland; however, the pattern is rather noisy (Szoszkiewicz et al., 2020).

2.2 Biotic dataset

Sampling was conducted according to EFI + IBI_PL method used for monitoring of the environmental state of rivers in Poland (Adamczyk et al., 2017; Napiórkowska-Krzebietke et al., 2020) developed based on the new European Fish Index (EFI+) and Index of Biotic Integrity (IBI). The samplings were collected between late summer and early autumn, i.e., 15 August to 31 October 2011–2020. No sampling occurred in 2013 and 2015. The sampled river length was 10–20 times the width of the stream but no less than 100 m. Different teams across the country conducted the samplings. These data were later aggregated and are maintained by the Chief Inspectorate of Environmental Protection (GIOŚ), from which the whole biomonitoring dataset was acquired.

The electrofishing was conducted through wading, fishing from a boat or a mix of both methods (Kolada, 2020). When sampled by wading, a backpack electro-shocker unit was used (max. power 1.8 kW, half-wave rectified current voltage of 220 V and intensity 3–5 A) with two hand-hold electrodes: anode made of a round net (5-mm mesh diameter); and a cathode made of copper wire approximately 1-m long (the length varies depending on the water conductivity). When sampled from a boat, a stationary unit was used (min. power 1.8 kW, direct current voltage of 150–300 V and intensity 3–5 A) complemented with 1 or 2 anodes made of a round net (5-mm mesh diameter), and a cathode made of copper wire approximately 2-m long. A certified person conducted the electrofishing, and the fishing teams consisted of three to five people depending on the method applied. According to the methodology, it was attempted to catch at least 30 specimens per site.

Fishes were caught, examined, counted, identified to a specified taxonomic level and released unharmed into the water except for invasive species specimens (as specified in Dz. U. 2011, nr 210/1260), which were killed in a humane manner. The species identification was conducted by experienced staff, and in case of doubt, taxonomic manuals were followed (Brylińska, 2000; Gąsowska, 1962; Rolik & Rembiszewski, 1987).

Where possible, the samplings were complemented with measurements of water temperature, pH, conductivity and velocity. The qualitative features such as the riverbed heterogeneity, substrate type (silt/clay <0,2 mm; sand 0,2–2,0 mm; gravel/stones 2–200 mm; rocks >200 mm), and the characteristics of the immediate surroundings (e.g., land use, riparian vegetation and flow alteration) were assessed based on expert judgement. The fish dataset used counts 3,023 sampling events, collected over 2,107 distinct geographic sites.

2.3 Environmental variables

The influence of 14 environmental variables (hereafter: predictors) on the fish community structure was investigated: most of them originate from the biomonitoring database obtained from GIOŚ, while some were obtained from existing datasets (Table 1).

Field-observed information on the presence/absence of trees (0, absence; 1, presence), shrubs, floodplain and reinforcements were combined and defined as the sum of individual class scores to create a natural index (NI). Similarly, an anthropogenic index (AI) was developed based on data on presence/absence of flow alteration and upstream reservoirs (0, absence; 1, presence). Both NI and AI were defined as the sums of presence/absence binary values. Finally, the Hydromorphological index for rivers (HIR, Szoszkiewicz et al., 2020) was used to distinguish natural, heavily modified rivers and artificial channels under lowland, mid-altitude and highland settings for entire Poland. All environmental predictors were classified in two broad categories: those that represent a pressure due to human alterations (HIR, AI, NI, conductivity, land use) and those that represent broader charachteristics of the sampling sites and are less affected by human activities (e.g., elevation and slope).

2.4 Regionalization strategies

The biotic data were subsetted based on five different regionalization strategies: (i) river basin (Figure 1b); (ii) elevation zones (Figure 1c); (iii) ecoregion (Figure 1d); (iv) ichthyological types (Figure 1e); and (v) water regions (Figure 1f).

The Vistula and the Odra are the two major rivers in Poland and show distinct features. For instance, the Vistula River has twice as much carrying capacity of bed and suspended load as the Odra River (Bajkiewicz-Grabowska et al., 2020). The Odra and Vistula catchments were analysed separately as they drain almost 90% of Poland.

Three classes were used as elevation zones: below 200 m a.s.l. (predominating lowlands), 200–400 m a.s.l. (uplands) and above 400 m a.s.l. (mountainous regions in the south).

Freshwater ecoregions are large areas of land or water with uniform assemblages of natural communities and species (Abell et al., 2008). For this analysis, ecoregion 10 (the Carpathians), 14 (Central Plains) and 16 (Eastern Plains) (Illies, 1978) were considered.

The ichthyological typology is a classification developed for the WFD reporting of water bodies' ecological status aiming at clustering fish environments based on abiotic and biotic factors such as substrate type, geology, macroinvertebrate type and macrofauna. Ichthyological types considered in this study are as follows (Grela & Madej, 2019): Type I: mountainous, upland rivers and streams; Type II: flysch rivers (rivers with sedimentary rock as bed material); Type III: lowland streams; Type IV: lowland rivers; Type V: inter-lake salmon rivers; Type VI: inter-lake and estuary cyprinid rivers (excluded from analysis due to the low number of sampling sites).

Finally, the ‘water regions’ (subcatchments used in the WFD reporting) were used as the last subsetting strategy. Only 12 water regions containing a sufficient amount of data were used in the analysis (Figure 1f). They represent the highest level of spatial aggregation in this study.

2.5 Data analysis

We employed redundancy analysis (RDA) to relate environmental variables and species data using functions from the ‘vegan’ package (Oksanen, 2018 ) in the R computing environment (v4.1). We removed the years 2011 and 2012 from the analysis, due to the reduced geographic coverage of the samples collected over this years. Additionally, fish records were dropped where no corresponding field variables were available. This resulted in a smaller dataset (1,848 sampling events collected over 1,429 geographic sites). Raw abundance data were transformed to presence/absence to reduce the effects of the sampling effort bias.

To test whether different subsetting methods yield different fish community structure (objective i in the introduction), we first fitted an RDA without any environmental predictor (unconstrained RDA or principal component analysis [PCA]). Then, we extracted the first two axis (retaining most of the variance in the species dataset). We used analysis of variance (ANOVA) and Tukey's Honest Significant Differences (HSD) post hoc test to compare if different subsets would yield a different ordination along the first two PCA axis.

To reach the objective (ii), we used RDA with biotic and environmental data for each spatial subset. We modelled explicitly the spatial structure of the dataset using Moran's Eigenvalues Maps (MEMs; Borcard et al., 2004, Dray et al., 2006). First, we calculated a distance matrix using the ArcGIS 10.6 software (ESRI, Redlands, CA) with an ArcGIS Network Analyst Extension. Based on the distance matrix, we generated distance-based Moran's Eigenvalue Maps with the ‘adespatial’ package. We then used a forward model building procedure (function ‘ordistep’) to retain a parsimonious set of MEMs that best explains the spatial structure of the data (Blanchet et al., 2008). In the last step, we performed a partial RDA (pRDA) using the environmental predictors matrix as constraining matrix, and the retained MEMs matrix as a conditioning matrix, whose effects are ‘partialled out’ (Legendre & Legendre, 2012). pRDA allowed us to split the RDA model's variance into partialled (i.e., explained by the spatial predictors), constrained (i.e., explained by the environmental predictors) and unconstrained (i.e., unexplained).

Finally, in order to achieve the objective (iii), we used hierarchical variance partition (package ‘rdacca.hc’) to allocate the fraction of the explained variance (expressed as individual predictor's R²) to each retained predictor. We used the fraction of the variance explained exclusively by a predictor (i.e., no shared variance with other predictors) as a metric of the predictor's importance. We applied the same procedure to the full dataset and to the reduced datasets corresponding to each regionalization subset. We used the MEMs generated for the full dataset also for the subsets, to retain the full spatial variability embedded in the study area (Brind'Amour et al., 2018), but we used forward selection procedures multiple times to select the MEMs that best explained the spatial variance for each regionalization subset.

Although some of the sampling sites were sampled multiple times, we did not implement a repeated measures approach, because the largest majority of the sites (1,079) was sampled only once (288 sites were sampled twice, and 68 sites were sampled three times). Additionally, the time interval between two subsequent samplings was never smaller than 3 years, so we considered the repeated measures to be temporally independent.

3.1 Effect of the spatial subsetting on fish community structure

A total of 64 fish species belonging to 17 families were recorded in Poland (Figure 2). Cyprinidae was the most widespread family at the national level, followed by Percidae and Esocidae. Gobio gobio was the most abundant fish species followed by Perca fluviatalis. Some non-native, introduced species, like Carassius gibelio, were widespread, while others (Lepomis gibbomus, Neogobius gymnotrachelus and Neogobius fluviatilis) were less pervasive.

Among the subsetting units, Cyprinidae was consistently the most abundant fish family, following the national trend (Figure 2). Salmonidae were most abundant in the Carpathian ecoregion, while Centrarchidae and Siluridae were the least abundant families across all the ecoregions. The high elevation zone was found to have the least number of fish families recorded (eight), with brown trout (Salmo trutta) being the most abundant species in this class. Among the river basins, few differences were recorded (e.g., Barbatula barbatula was more abundant in the Vistula Basin than the Odra Basin). Cyprinidae and Salmonidae were the most abundant fish families in the ichthyological Type I rivers. The majority of water regions followed the national trend of Cyprinidae and Percidae as the two most abundant fish families, with the exception of Upper Vistula East and Upper Vistula West (Nemachellidae was the second most abundant family) and Narew (Gasterosteidae was second).

The first two axes of the unconstrained RDA ordination allowed for a clear separation of the fish communities for the ecoregion, the ichthyological type and the elevation subsetting methods, while yielding unsignificant or mixed results for the basin (the coarser resolution) and the water region (the finer resolution) subsetting methods (Figure 3b–f, Table 2). Therefore, the latter two subsetting methods have a limited potential to reduce the heterogeneity in the fish community data in the study area.

3.2 Effect of spatial subsetting on fish community variance partitioning

The RDA analysis on the national dataset explains 15% of the variance of the dataset, with the first two constrained axes contributing 9% and 3%, respectively (Figures 3 and 4). The explained variance was variable across different subsets, ranging between 5% and 25% for individual subsets. A high heterogeneity is observed for the water region subsets, where, for example, Lower Odra and Upper Odra show the lowest explanatory power in the analysis, <5%, while the results at the basin level are comparable to the national assessment. The fraction of variance explained by the spatial predictors (MEMs) is also variable and ranges between 10% and 45%, with the exception of the high elevation subset, for which it accounts for 85% of the variance. Two trends are visible across subsetting strategies: (i) variance explained by the spatial predictors increases with increasing elevation (e.g., high elevation subset, Type II ichthyological subset and Carpathian ecoregion) potentially indicating an increasing dispersal limitation in elevated areas; (ii) the higher the level of subsetting is, the higher is the variance explained by spatial variables. In the water region subsets, the variance explained by MEMs was generally highest indicating an increase in spatial pattern with increasing subsetting level.

3.3 Effect of spatial subsetting on the most relevant predictors for fish community structure

At the national level, the top-ranked predictors (Figure 5) were depth, slope, sediment type, velocity, geomorphological type (all natural predictors) and conductivity (the only pressure predictor). The ranking in the subsets was variable: Depth was ranked first for 13 subsets from all five subsetting strategies. Slope was ranked first for three subsets from three subsetting strategies. Sediment type was ranked first for five subsets. Velocity and geomorphological type respectively were ranked first only for one subset each. Pressure predictors were generally and consistently ranked lower than natural environmental predictors, with conductivity and AI having the highest ranks. Environmental predictors showed generally smaller gradients across subsets, more marked for the water regions (Table S1).

Comparison of predictor rankings (top five variables) for each subsetting strategy shows that for river basins and ecoregions, the differences in ranking between particular subsets were rather low (Figure 5). In contrast, water regions subsetting exhibited the highest differences in predictor ranking between its individual subsets. Predictor rankings were also quite diverse within the elevation and ichthyological type subsets.

Spearman's correlation coefficients calculated between the national ranking and the rankings at individual subset levels showed high variability (Table 3). River basins, low elevation zone, ichthyological classes III and IV, and Central and Eastern European ecoregion showed a significantly high alignment with the national ranking, while water regions did not. Fewer predictors were found to be significant for water regions (in some cases two or three) and for the high elevation class. Across 12 water regions, the first position was occupied by six different natural environmental variables. Depth, sediment type and slope were not significant for three, four and five out of 12 water regions, respectively. Sediment type, velocity and geomorphological type gained importance for several water regions compared to the national scale (Figure S2). For example, sediment type was ranked first in three water regions of southern Poland: the Middle Odra, Upper Odra and Upper Vistula West regions; velocity gained most importance in central and northern water regions, i.e., Lower Vistula (no. 1) and Middle Vistula, Warta and Notec (no. 2). Some pressure variables gained significant importance for specific water regions, e.g., conductivity (no. 3 in Upper Vistula West) and NI (no. 3 in Upper Vistula East and Notec).

4.1 Strengths and limitation of this study

Comparison of our results with previous similar research in Poland in the same spatial context is difficult, as the scope of previous studies linking fish communities with environmental variables was usually different from ours. No similar study performed at the country scale was detected, and, to our knowledge, no studies analysed such a wide array of potential predictors. However, a few studies analysed the effect of one or several factors on fish community structure in specific river basins. For example, Penczak et al. (2004) focused on determining changes in fish assemblage related to environmental variables in the Warta River and its oxbow lakes. The authors reported velocity, conductivity and water temperature to be significant environmental variables. The importance of velocity was consistent with our results that indicated substantial improvement of predictive power in velocity at the water region level, and for the Warta water region, velocity was the second most important predictor. Penczak (1995) showed that bankside vegetation had a strong effect on fish in a short section of the Warta River. Our results show a similar effect on a large scale, as the NI taking into account channel vegetation, was significant in the lowland subset as well as in the ichthyological Types III, IV and V, the spatial regions where pressures are most abundant. Similarly, our results capture the effect of impoundments (included in the AI) on fish assemblage structure, as demonstrated by Penczak et al. (2012) for the Warta River.

Our analysis could be further improved by the inclusion of several classes of predictors. The only indicator that could be used as a proxy for water quality was conductivity. However, other variables that do not correlate well with conductivity, e.g., total phosphorus, dissolved oxygen and metals, may have a substantial effect on fish communities (Bervoets et al., 2005; Sutela et al., 2010). Water quality has substantial variability across Polish rivers, with a direct effect on their ecological status (Solarczyk, 2021). Thus, we expect that analyses including water quality indicators would increase the explanatory power of our analyses. The use of upstream-accumulated predictors (e.g., cumulated land use and topology, Kukuła & Bylak, 2020) could further improve the explanatory power.

4.2 Factors affecting fish community for subsetting units

A clear pattern of drivers of fish assemblages across subsets and subsetting methods was lacking. At the ecoregion level, we observed a prevalence of morphological predictors, even though ecoregions should reflect bioclimatic diversity. Ecoregions with the lowest elevation gradient (Central and Eastern European and Eastern European) showed a higher diversity of families. On the contrary, the mountainous Carpathian ecoregion displayed the least number of families, with the intolerant cold-water fishes being dominant. The Carpathian ecoregion substantially differed from the others also based on the different rankings in the predictors. This is corroborated by the study of Wang et al. (2003), who found that in Canada in the ecoregion characterized with relatively narrow, high-gradient, cold streams, intolerant cold-water fishes were more abundant.

The effect of spatial structuring was evident also for ichthyological types. Streams and rivers in mountainous and upland areas belonging to ichthyological Types I and II had elevation or slope as the most important predictor, probably reflecting the changing climatic and temperature conditions with altitude (Gehrke & Harris, 2000). For Type III lowland streams, sediment type was an important driver for community composition probably due to the presence of lithophilic, psammophilic and phytophilic fish spawning guilds (Balon, 1975) strongly associated with the substrate conditions. In lowland rivers (Type IV), depth drives community composition as a proxy of river size (Griffiths, 2018). In the Type V inter-lake Salmon rivers, salmonids occur in relatively low numbers, but the fish assemblages comprise several rheophilic and lithophilic salmonid fish species (EFI+, Manual, 2009). Streambed and bank morphology, along with substrate type and water depth, can directly influence such species (Sullivan et al., 2006; van Treeck et al., 2020).

At the water region level, predictors rankings became highly diverse without a clearly interpretable pattern. The consistency in the rankings of predictors among Narew, Notec, Bug and Warta could be attributed to the fact that these subsets fall within the Central and Eastern European ecoregion. A similar explanation could yield for Upper Vistula West and Upper Odra basins. The coherence in ranking among neighbouring water regions might indicate the spatial extent of such subsets to be too small to effectively rule out potential confounding factors and at the same time preserve predictor gradients sufficiently large to detect significant relationships with the biotic data.

On the other hand, the variables relatable to anthropogenic pressures show clear trends in the data. The importance of pressure variables in the models decreases from lowland to upland regions, reflecting the increase of human pressures form upstream to downstream regions. This is a general pattern across Europe; in headwater streams, the proportion of unimpacted river stretches is higher than in lowland rivers where overall more than 90% of the rivers stretches are impacted (Schinegger et al., 2012).

Another clear pattern is the growing importance of spatial variables along the higher elevation gradient. This might indicate the increasing dispersal limitation in mountain regions caused by greater resistance of the river network to dispersal of organisms (Tonkin et al., 2017). There is not only a trend of increasing importance of spatial variables with increasing level of subsetting but also a trend of increasing heterogeneity, clearly showing the importance to account for spatial pattern when subsetting is applied in environmental studies.

Overall, we demonstrate how the analysis of fish community data for different regionalization strategies can show significantly different results in terms of rankings of important structuring factors and explanatory power. We conclude that in the analysis of biotic data, various regionalization matters should be carefully considered.

4.3 Management implications

The differentiation of fish assemblages on a regional scale has some implications for the accuracy of fish-based methods for environmental status assessment (Schmutz et al., 2007). In the EFI + IBI_PL method applied in Poland, some regionalization patterns are considered, such as ecoregion and abiotic river type (Adamczyk et al., 2017); however, it should be noted that the original IBI index was constructed for specific catchments (Karr, 1981; Karr et al., 1986). Thus, it can be assumed it has greater precision on catchment scale. The development of catchment- or water region-specific assessment methods may further improve fish-based tools for river ecological status assessment.

Knowledge of local fish fauna differentiation may be also of great importance for fisheries management practices. In Poland, the riverine fish populations are exploited mainly by recreational fishery (angling) and in some large rivers—commercial net fishery is still applied (Radtke et al., 2018). This creates a need for a stocking programme obligatory for all water users from the fisheries sector. The accurate determination of species composition for stocking and amounts of stocking material of each species stocked yearly is crucial for maintaining appropriate (close to natural) proportions in fish assemblage, which also corresponds to maintaining good ecological status (Kaczkowski & Grabowska, 2015). Understanding the factors that best explain the variation of fish communities can support the design and implementation of locally-tailored stocking programmes.

Environmental flows play a pivotal role in ecosystem protection, and their provision follows the natural flow pattern and enables the ecosystem processes to function throughout the year (Grela & Madej, 2019; Wurbs & Hoffpauir, 2017). Environmental flows are often set within comprehensive frameworks, such as ELOHA (Poff et al., 2010), consisting of hydrological classification on the one hand and defining flow–ecology relationships on the other. However, environmental flow assessment has often primarily centred around maintaining low flows as argued by Kuriqi et al. (2019). Streamflow-driven fluvial processes shape channel and floodplain morphology mainly during the high-flow season. Our results indicate a variable most correlated with streamflow, i.e., flow velocity, gains more importance at the water region scale than for other subsetting strategies. This suggests that water region could be more appropriate than other regionalizations for setting flow–ecology relationships for fish, but less appropriate to detect relevant factors for community composition. This is in accordance with the study of McManamay et al. (2015), who reported that ecoregion-type classifications are missing conceptual linkage between hydrological variation and fish communities and recommended using hydrological classes nested within larger units such as provinces or ecoregions to increase the variation explained in fish traits.