Freshwater ecosystems are among the most threatened by human activities (Gatti 2016). The knowledge of the various components of these ecosystems is of paramount importance to the elaboration of public policies on conservation or recovery (Bunn et al., 2010). Biomonitoring data has been increasingly used in determining the ecological conditions of aquatic environments, in addition to the traditional physical and chemical indicators of water quality (Leese et al., 2018; Pardo et al., 2018; Gieswein et al., 2019). A comprehensive ecosystem integrity assessment should consider both structural and functional characteristics (Bunn & Davies 2000). While the structure of an ecosystem comprises physical and chemical attributes related to water quality, composition of biological assemblages and habitat conditions, its functioning is related to the processes regulating energy and matter fluxes (Tilman et al., 2014).

The most commonly used metrics to assess freshwater ecosystem integrity are those related to biological assemblages, such as species richness and diversity, abundance, the proportion of tolerant and sensitive taxa, organismal traits (e.g., feeding habits, body size, mobility), and indices of sensitivity to pollution (Hering et al., 2006). Macroinvertebrates, diatoms, macrophytes and fish are often used for that purpose (Son et al., 2018; Waite et al., 2019) as they are robust to the identification of several human disturbances and present particular features that facilitate such application (e.g., life cycle, habitat, size; Merritt & Cummins 1996; Kelly et al., 2008). Much less explored in the context of biomonitoring are aquatic fungi and bacteria, which are key decomposers of organic matter in streams. The responses of some ecosystem processes to stressors are fundamental to understanding the effects of human disturbances on ecosystem services that produce direct benefits to people. But despite this, there is still a lot of reluctance among managers and little use of functional indicators (e.g., litter decomposition) in monitoring programs (Schiller et al., 2017).

Although many studies have pointed out to the applicability of several ecological metrics for assessing freshwater ecosystem integrity, the main gap lies in the relative importance of different mechanisms that cause impacts and the interactions between them (Wenger et al., 2009). According to Sutherland et al. (2013), one solution is the use of modelling as a tool for measuring and monitoring systems. In the context of environmental management, most models used in monitoring programs consider biological assemblages, especially benthic invertebrates (AUSRIVAS, Smith et al., 1999; RIVPACS, Wright et al., 1984; USEPA, 2016), as indicators. Some studies suggest the use of ecosystem processes for this purpose (Gessner & Chauvet 2002; Feio et al., 2010; Woodward et al., 2012), and rare are those that present a multi-metric approach using structural and functional aspects (but see Castela et al., 2008; Clapcott et al., 2014).

In Brazil, as in many other tropical countries, monitoring programs focus on physical and chemical variables of water, with a substantial gap in knowledge about the structure and functioning of aquatic ecosystems. The Brazilian savanna (Cerrado) is a global biodiversity hotspot (Myers et al., 2000), and its headwaters are responsible for 70% of all water supply to other Brazilian regions (Lima & Silva 2007). However, the devastation of Cerrado has been taking place at levels proportional to its ecological and social relevance (Strassburg et al., 2017). It is urgent to know the behaviour of these threatened ecosystems through two valuable management tools: 1) the identification of variables that respond strongly to anthropogenic impacts than to natural variations, and 2) the identification of ecological thresholds that are adopted as standards in monitoring programs. We define the term ‘ecological threshold’ as a point along a stressor gradient where the relationship between the stressor and an ecological indicator shows an abrupt change in the response curve that can be ecologically explained and significantly relevant for management (Wagenhoff et al., 2017).

In this context, Boosted Regression Tree (BRT) models have been used as a robust tool to identify the influence of environmental variables, natural or those related to human activities, on ecological metrics, making it possible to evaluate the shape of the responses and to make forecasts by using new data (Clapcott et al., 2012; Waite et al., 2019). This approach allows the identification of gradients, from which is possible to detect non-linear responses, interactions among predictors and potential threshold zones. However, identification of potential thresholds alone is not helpful for management if it does not accompany by an analysis of their ecological consequences relevant to management decisions. Furthermore, it is important to consider a group of non-redundant variables so that the impacts of different stressors on the structure and function of ecosystems are detected. This approach would lead to more robust in-stream objectives and provide options for adopting goals that protect the aspects of ecosystems that people value most (Wagenhoff et al., 2017). Several studies have used BRT models to identify thresholds in a diversity of ecological areas, highlighting the potential of this tool to environmental management (Davis et al., 2019; Giri et al., 2019; Wherry et al., 2021).

This study evaluated how ecological metrics respond to natural environmental gradients and human-related stressors (local and catchment) at different spatial scales. We also identified the most suitable metrics to be used as indicators of stream integrity and assessed the response and potential thresholds of ecological metrics along environmental gradients to inform the ecological management of Cerrado freshwaters.

The study made it possible to identify the main predictors driving each ecological metric and how metrics responded to natural and human-related predictors, allowing the detection of potential indicators of stream integrity. While the percentage of EPT group abundance and the algal biomass would be good indicators of urbanization in the upstream catchment, the percentage of Eunotia abundance would indicate changes in water quality. In contrast, other metrics were poorly explained by the predictors or mainly influenced by natural predictors, making them inappropriate indicators of environmental disturbances for management purposes (Norris & Hawkins 2000). Like the Simpson Index for Diatoms showed a low variance explained (17%), the percentage of Plecoptera, which was influenced mainly by natural characteristics (elevation and drainage area), and decomposition primarily influenced by seasonality. Most of our models presented a unidirectional response for direct (land use) and indirect (water quality) human disturbances. Overall, increasing human disturbance (e.g., conductivity and changes in land use) led to a decrease in pollution-sensitive taxa (e.g., percentage of EPT group and Eunotia) and an increase in pollution-tolerant taxa (e.g., percentage of Oligochaeta and Hirudinea, the Trophic Diatom Index and the algal biomass production).

The nonlinear responses promoted insights into the subsidy-stress theory (too much of a good thing syndrome; Odum 1983), which predicts that the increase of limited resources (e.g., nutrients, light) in an environment may have an initial positive effect on biological communities and ecosystem functions. However, this effect rises to a certain threshold; after then, it can lead to adverse effects. In this context, considering that Brazilian savanna streams are poor in nutrients (Markewitz et al., 2006), nutrient inputs possibly promote the maintenance of more species/individuals. But at the other extreme of the gradient, intense disturbances are expected to reduce the number of species that can colonize or tolerate high impact levels (Odum 1983). The shape of the EPT curve (initial low value followed by a sharp rise, lately a decrease) indicates their sensitivity to disturbed environments face to the increase in conductivity, and catchment urbanization was an example of this (Ligeiro et al., 2013; Siegloch et al., 2017). The evaluation of the response curves from BRT models was also a good starting point for discussing thresholds for the considered predictors. Notable change points could be observed, such as conductivity, phosphate, and the percentage of urbanization in the upstream catchment.

Our study showed the most important predictors to explain the ecological metrics were physical and chemical variables often used to indicate human disturbances (Heathwaite 2010; Uriarte et al., 2011; Alvarez-Cabria et al., 2016), such as phosphorus and nitrate concentrations, but especially conductivity. For instance, we reported significant changes in ecological metrics when conductivity stood between 100 and 200 µS cm⁻¹, suggesting a potential threshold. Values above this threshold indicate loss of water quality, except when high conductivity is due to the natural background (Fravet & Cruz 2007; Fundação Nacional de Saúde, 2014; CETESB 2020). Conductivity was the main predictor for the studied metrics in terms of relative importance. Comparing water bodies of preserved and anthropogenic (especially those without vegetation protection) areas, the diffuse sources of pollution resulted in higher electrical conductivity (Gardiner et al., 2009; Rezende et al., 2014). Like in anthropogenic areas with inadequately treated effluents flowing to water bodies, increasing the nutrient concentrations of the water (Myrka et al., 2008).

Phosphorus increase is responsible for triggering the eutrophication of freshwaters (Figueredo et al., 2016; Zhang et al., 2017) coming from agricultural fields and urban effluents (Ockenden et al., 2016). Our results showed potential thresholds for phosphate around 0.5 mg L-1, and its contribution was especially relevant for metrics sensitive to pollution, such as %EPT, %Oli_Hir and %Eunotia, as shown elsewhere (Kelly. 1998; Salomoni et al., 2006; Ferreira et al., 2014; Pardo et al., 2020). Both conductivity and phosphorus were positively related to effluent discharge and deforestation. Anthropogenic areas (remarkably urbanized areas) strongly influence biological assemblages, and their effects are disproportionate to the size of the area used (Rezende et al., 2014; Campos et al., 2021).

Urban and agricultural cover in the upstream catchment was the most important land-use factor to explain the response metrics. The adverse effects of replacing native vegetation with urban or agricultural areas in the upstream catchment have been reported for the stream via complex pathways (Allan 2004) like changes in temperature, habitat diversity, hydromorphology, sunlight, and nutrient availability (Einheuser et al., 2013). These changes have translated into alterations in the structure and functioning of the stream ecosystem (Clapcott et al., 2012). We observed that values between 10 and 20% of urban cover in the upstream catchments led to a decrease in the abundance of the EPT group and an increase in algal biomass. Brito et al. (2020) reported abrupt changes in the composition of macroinvertebrates with the removal of 57–79% of native vegetation in the Amazon Forest, while Dala-Corte et al. (2020) reported threshold values between 3 and 40% of native vegetation removal across biomes in Brazil. Therefore, our results in the study region indicated more restrictive values suggesting that parts of the Brazilian savanna are more susceptible to the conversion of native areas. Additionally, the increase of algal biomass related to the urbanization process confirms a recent study that shows a 32% greater effect on stream functioning than in its structure in the tropics (Wiederkehr et al., 2020).

Changes in biological assemblages and ecosystem processes are commonly associated with alterations in the riparian plants (Encalada et al., 2010; Fierro et al., 2017), especially in headwaters that are light-limited systems and rely on plant litter inputs from surrounding vegetation (Bunn & Davies, 2000; Perona et al., 2009). However, we did not observe a robust relationship with macroinvertebrates. On the other hand, we found a consistent negative relationship among diatoms, urbanization and agriculture in the riparian zone, indicating a higher local than catchment-scale effect. This finding suggests reliable benefits of forested riparian buffers for stream biological diversity in urban environments, supported by previous studies (e.g., Mutinova et al., 2020).

The different responses to the set of predictors, including structural and functional ecosystem metrics, can lead to a comprehensive interpretation of river conditions (Feio et al., 2010). The prediction of ecological conditions is relevant from the management’s point of view since these are more complex data to be acquired but of extreme relevance for understanding the health of water bodies (Karr 2006). Knowledge about the importance of each predictor for the response metrics allows, for example, to predict some ecological conditions in places with limited availability of biological data.

Finally, the potential thresholds identified in the present study are important signs of significant changes in ecological responses. They should be employed in eventual review processes of guidelines to public policies for river health preservation and recovery (Huggett 2005). Brazilian national environmental guidelines do not consider, for example, conductivity (CONAMA n° 357, Brasil 2005), notwithstanding the importance of this variable as an ecosystem driver, as demonstrated in our study. In addition, further attention should be paid to the context of land use, especially to the urbanization processes in the upstream catchment. Currently, Brazil has increased awareness of riparian vegetation (Federal Law n° 12.651, Brasil 2012). However, for purposes of biodiversity conservation and maintenance of ecosystem processes, we also have shown it necessary to consider the entire context of the catchment in which the stream is located.

Our results demonstrated the importance of considering a set of ecological response metrics (structural and functional) and environmental factors (natural and disturbances), allowing a complete view of the freshwater ecosystem condition. The relative importance of predictors on ecological metrics pointed to metrics most affected by factors on a local scale (e.g., percentage of Eunotia abundance) and catchment scale (e.g., algal biomass). Also, the nonlinear responses permitted the detection of gradual or abrupt change curves, pointing out the existence of potential thresholds of important drivers, like the conductivity (100–200 µS cm-1), phosphate (0.5 mg L-1), and catchment-scale urbanization (10–20%). The potential bioindicators (considering the best performance models and the ability to respond more strongly to the human disturbances) were macroinvertebrates abundance, EPT abundance percentage, Oligochaeta and Hirudinea abundance percentage, percentage of Eunotia abundance, Trophic Diatom Index, and algal biomass. Although we have worked with many biotic and abiotic variables and the BRT model considered the interaction between them, models are simplified representations of a complex system, therefore presenting limitations. Nevertheless, the consistency and reasonableness of influential metrics within a given set of ecological metrics provide a weight of evidence in support of the models’ results. The BRT models approach proved to be powerful tools that can be effectively employed to enhance and give better direction to freshwater management, not only to the streams of the Brazilian savanna but also to water bodies in other regions.