Analysing the effects of water quality on the occurrence of freshwater macroinvertebrate taxa among tropical river basins from different continents

Abstract

Macroinvertebrates are globally used in environmental monitoring and assessment. However, due to environmental and biological evolution, local adaptations of species might occur. This can contribute to uncertainties in the extrapolation of family-specific ecological models developed from one region to another. Thus, we aimed to determine if models can be extrapolated to other regions with similar climatic conditions and if a reliable model can be developed from a pooled dataset (consisting of data from different regions). The occurrence of five families was modelled based on physical–chemical water quality variables with classification trees using the data from three tropical river basins (Chaguana in Ecuador, Gilgel Gibe in Ethiopia and Cau in Vietnam). The relevance of each model was tested on complementary data from both the same and other river basins, to test specificity and universality. Furthermore, models with a pooled dataset were developed and tested. Model reliability was assessed based on chance-corrected agreement (Cohen’s kappa, κ) and percent agreement (correctly classified instances, CCI). Values of higher than 0.4 (κ) and 70% (CCI) were used to classify models as good. Only the pollution sensitive taxon (Leptophlebiidae) resulted in reliable models for most cases. In general, responses of macroinvertebrates towards pollution were different among countries except for the pollution sensitive taxa. Thus, extrapolation of ecological models for sensitive taxa to another river basin with similar climatic and environmental conditions is possible. Nevertheless, this type of systematic analyses for all families is necessary to determine and minimize uncertainty in ecological assessment.

Keywords

Classification trees ecological modelling pollution decision trees ecological assessment

1. Introduction

Macroinvertebrates are globally used in environmental monitoring and assessment [35,37,45]. They integrate environmental changes in physical, chemical and ecological characteristics of their habitat over time and space [64]. Most aquatic invertebrates are relatively immobile and reside in the benthic habitat for at least part of their life. Due to their sensitivity, disturbances in the aquatic environment generally affect macroinvertebrate abundance, diversity and composition [56]. Thus, macroinvertebrates are a suitable bio-indicator of the ecological state of aquatic ecosystems.

Ecological modelling is an effective tool to investigate, describe and predict the ecological state of an aquatic ecosystem [40]. Techniques such as decision trees have been employed to analyse biological assemblages of macroinvertebrates in British and Slovenian rivers [24]. Decision trees can be classified into classification and regression trees. Classification trees are used when the predicted outcome is a categorical variable while the latter are used when the predicted outcome is a continuous variable. Everaert et al. [25] and Boets et al. [5] assessed the impact of alien species on the macroinvertebrate composition in Flanders, Belgium with classification trees. Furthermore, classification trees have been used to predict the distribution of an alien species, Azolla filiculoides (Lam.) at Selkeh Wildlife Refuge in Anzali wetland, northern Iran [68]. Decision trees accommodate the challenges posed by ecological data as these techniques are: flexible to handle continuous and discrete responses and explanatory variables; able to capture nonlinear relationships, complex interactions, and multiple structures; able to deal with missing values in response or explanatory variables; robust to outliers; invariant to monotonic transformation of the explanatory variables; and easy to construct and interpret [53].

According to some studies, responses of macroinvertebrates to changes in environmental conditions are similar in regions characterized by comparable climatic and environmental conditions. Bonada et al. [6] concluded that ecoregions with corresponding historical, climatic and environmental conditions are also inhabited by macroinvertebrates with an analogous assemblage structure. Furthermore, Dynesius et al. [23] reported that the responses of riparian plant species to river regulation were similar between the two different continents (North America and Europe) and suggested that a general guideline for rehabilitation of degraded boreal rivers is a realistic goal. However, some studies reported that relationship between taxonomic assemblage structure and land use intensity differed among ecoregions [18]. Taxonomic richness also varies greatly in response to environmental variability [62]. It remains unclear whether an aquatic macroinvertebrate taxon responds similarly towards changes in environmental conditions among the tropical region. This is due to the fact that related studies in the tropical-climate zone are minimal.

Although numerous ecological models were developed to analyse and predict the assemblage condition in the tropics [21,39,51], these models were never extrapolated to other regions with similar climatic and environmental condition. Furthermore, results from the development of ecological models with a pooled dataset (dataset consisting of data from regions with similar climatic condition) are limited. The question remains whether models developed for a specific region can be extrapolated to other regions with similar climatic condition or whether a reliable model can be developed from a pooled dataset. Thus, this study explores the relationship of indicator taxa and key environmental variables for water quality assessment in a broader regional context and provides an insight into whether a general guideline for bioassessment in tropical countries can be implemented.

This study aims to determine if models can be extrapolated to other regions with similar climatic conditions and if a reliable model can be developed from a pooled dataset. We used classification trees to model the relationship between physical–chemical variables and the occurrences of five macroinvertebrate families, ranging from pollution tolerant to pollution sensitive, occurring in Ecuador, Ethiopia and Vietnam.

2. Materials and methods

2.1. Study areas

Chaguana river basin, Gilgel Gibe river basin, Cau river basin is located southwest of Ecuador, south to west of Ethiopia, and north of Vietnam (Appendix Table 1). The three river basins are prone to similar pressures. The Vietnamese river basin is characterized by a high population density, industrialization and sand/gravel mining [57]. The Ecuadorian river basin is covered by 37% of near natural conditions (humid forests and bushes, mangroves and uncultivated land), while the remaining parts are operated by shrimp farming, banana plantations, human settlements [21]. Lastly, some streams of the Ethiopian river basin are still covered with natural vegetation, while the rest is altered by different activities such as grazing, ploughing, sand dredging, vegetation clearance and municipal waste discharge [4].

2.2. Data collection

Data were gathered from three different sampling campaigns in the Chaguana river basin, Gilgel Gibe river basin and Cau river basin (Appendix Figs 1–3). Multiple physical–chemical water quality parameters were measured and macroinvertebrate samples were collected in each survey (Appendix Table 1). The number of sampling sites varied for each river basin, but each sampling site was visited twice in each year: once in the wet season and once in the dry season. In total, 60, 104 and 306 samples were taken at 15, 29 and 47 sampling sites for 2, 2 and 6 years in the Vietnamese, Ecuadorian and Ethiopian river basin, respectively (Appendix Table 2) [26]. Water samples were analysed according to the ISO standards. Only the environmental variables that were monitored in all of the three river basins were retained for further analysis. These are conductivity (μS/cm), dissolved oxygen concentration (mg/L), pH (–), stream velocity (m/s) and water temperature (°C) (Appendix Table 2). Samples were collected during two different seasons (wet and dry); therefore seasonality was also included for further analysis. Benthic macroinvertebrates were sampled according to the method described in Gabriels et al. [32]. A standard hand net was used. The net was attached to a stick and it had a size and mesh size of 20 × 30 cm and 300 μm, respectively. A stretch of 10–20 m was sampled with kick sampling method for 5 minutes. Sampling effort was proportionally distributed over all aquatic habitats. The organisms were identified to family level according to the taxonomy in [19,29,67] for Ecuadorian taxa, [7] for Ethiopian taxa, and [15,22,50] for Vietnamese taxa. For a detailed overview of the sampling campaigns and descriptions of the locations, we refer to Dominguez-Granda et al. [21], Ambelu et al. [4], and Nguyen et al. [57].

2.3. Classification trees

Classification trees are simple nonlinear and nonparametric approach that is fitted by recursive partitioning of a multidimensional covariate space [42,71]. They were applied to model macroinvertebrate families occurrences based on physical–chemical water quality variables. Five macroinvertebrate families, present in the three river basins, were selected (Table 1). They ranged from pollution tolerant to pollution sensitive, according to the BMWP score list [12].

Table 1
Selected macroinvertebrate families

Macroinvertebrate family BMWP score [12]

Chironomidae 2

Baetidae 4

Hydroptilidae 6

Libellulidae 8

Leptophlebiidae 10

Macroinvertebrate family	BMWP score [12]
Chironomidae	2
Baetidae	4
Hydroptilidae	6
Libellulidae	8
Leptophlebiidae	10

Threefold cross validation was applied to indicate the robustness of the model. The original data were divided into three groups, stratified based on the presence or absence of the selected taxa. Thus, the proportion of presence/absence records of the selected taxa is the same in each group. Two of these groups were used to build the model (training sets); while the third subset was used to test the model (validation sets). The first series of classification tree models contained data from each country separately: Ecuador, Ethiopia, and Vietnam. The second series of classification tree models were developed based on pooled datasets. This resulted in 7 training and 7 validation sets as illustrated in Fig. 1.

Fig. 1.

Schematic outline of the model development and validation process. CT represents classification trees.

The models were developed based on tenfold cross validation to avoid model overfitting. The training set (two-thirds of the original dataset) was therefore divided into ten groups. Nine of these groups were used to build the model while the tenth subset was employed to test the model (referred to as the internal validation, which is part of the model development procedure) as illustrated in Fig. 2. Each developed model was validated with the 7 validation sets (referred to as the independent validation) as shown in Figs 1 and 2. The default settings from the machine learning package WEKA were applied using the J48 algorithm [36].

Fig. 2.

Schematic presentation of internal and independent validation.

Trained models were evaluated on the basis of two performance measures: the percentage of Correctly Classified Instances (CCI) [30] and the Cohen’s kappa statistic (κ), in which: $\begin{matrix} CCI = \frac{(TP + TN)}{(TP + FP + TN + FN)} \times 100 % \end{matrix}$ κ is derived statistics that measures the proportion of all possible cases of presence or absence that are predicted correctly by a model by accounting for chance prediction [13]. $\begin{array}{rcl} κ & = & [(TP + TN) \\ - ((TP + FN) (TP + FP) \\ + (FP + TN) (FN + TN) n^{- 1})] \\ \times [n - ((TP + FP) (TP + FP) \\ + (FP + TN) (FN + TN) n^{- 1})]^{- 1}, \end{array}$ where TP is the number of true positive predictions, TN is the number of true negative predictions, FP is the number of false positive predictions, FN is the number of false negative predictions for n data points. κ is a widely used performance metric which accounts for inter-rater agreements [4,43,51]. A model is considered as valid and reliable in case the CCI exceeds 70% and κ is larger than 0.4 [34].

3. Results

Based on kappa and CCI statistics, reliable models for Leptophlebiidae were developed from the training sets that included Ecuador data (Table 2). Models for Hydroptilidae with a κ higher than 0.4 and CCI higher than 70% were developed from the pooled training set of Ethiopia & Vietnam (Table 3). The J48 algorithm was not able to develop reliable models (based on performance criteria) for Baetidae, Chironomidae and Libellulidae (Appendix Table 3–12).

Table 2
Internal validation of Leptophlebiidae predictive models. Results in bold indicate datasets providing reliable models based on performance criteria. Mean and standard deviation of CCI and κ were derived from threefold cross validation

Training sets Leptophlebiidae

Mean CCI ± std Mean κ ± std

Ecuador 75.04 ± 5.6 0.50 ± 0.1

Ethiopia 98.30 ± 0.004 0.00 ± 0.0

Vietnam 75.80 ± 3.35 0.28 ± 0.04

Ecuador & Ethiopia 91.80 ± 0.6 0.60 ± 0.03

Ecuador & Vietnam 74.40 ± 5.3 0.45 ± 0.10

Ethiopia & Vietnam 94.70 ± 0.2 0.02 ± 0.06

Ecuador, Ethiopia & Vietnam 89.40 ± 1.5 0.50 ± 0.07

Training sets	Leptophlebiidae
Ecuador	75.04 ± 5.6	0.50 ± 0.1
Ethiopia	98.30 ± 0.004	0.00 ± 0.0
Vietnam	75.80 ± 3.35	0.28 ± 0.04
Ecuador & Ethiopia	91.80 ± 0.6	0.60 ± 0.03
Ecuador & Vietnam	74.40 ± 5.3	0.45 ± 0.10
Ethiopia & Vietnam	94.70 ± 0.2	0.02 ± 0.06
Ecuador, Ethiopia & Vietnam	89.40 ± 1.5	0.50 ± 0.07

Table 3

Internal validation of Hydroptilidae predictive models. Results in bold indicate datasets providing reliable models based on performance criteria. Mean and standard deviation of CCI and κ were derived from the threefold cross validation

Training sets	Hydroptilidae

	Mean CCI ± std	Mean κ ± std
Ecuador	82.70 ± 1.2	0.04 ± 0.06
Ethiopia	98.30 ± 0.004	0.01 ± 0.02
Vietnam	67.40 ± 11.7	0.34 ± 0.20
Ecuador & Ethiopia	94.90 ± 0.31	−0.002 ± 0.003
Ecuador & Vietnam	72.90 ± 1.06	0.16 ± 0.06
Ethiopia & Vietnam	94.10 ± 0.53	0.55 ± 0.05
Ecuador, Ethiopia & Vietnam	89.90 ± 0.14	0.31 ± 0.01

Based on independent validation, some models for Leptophlebiidae and Hydroptilidae were reliable for some validation sets (Tables 4–6). For instance, models for Leptophlebiidae developed based on the Ecuadorian training sets were valid in the pooled Ecuadorian & Vietnamese validation sets. These models almost exceed a CCI of 70% and κ of 0.4 in the Ecuadorian and pooled Ecuadorian & Ethiopian validation sets. Furthermore, Leptophlebiidae models developed from the pooled datasets of Ecuador & Ethiopia, Ecuador & Vietnam, and Ecuador & Ethiopia & Vietnam were all valid in their validation set counterparts. Hydroptilidae models developed from the pooled dataset of Ecuador & Vietnam were reliable in the validation set comprising the Vietnamese data. However, both the CCI and κ were unstable as reflected in their standard deviations. Conductivity and dissolved oxygen (DO) were the major variables determining the presence of Leptophlebiidae (Table 7; Appendix Figs 4–7) whereas season, pH and water temperature were the major variables determining the occurrence of Hydroptilidae (Table 8; Appendix Fig. 8).

Table 4

Independent validation of Leptophlebiidae’s single country models. Results in bold indicate datasets providing reliable models based on performance criteria. Mean and standard deviation of CCI and κ were derived from threefold cross validation

Validation sets	Model

	Ecuador		Ethiopia		Vietnam

	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std
Ecuador	69.36 ± 9.5	0.38 ± 0.2	61.3 ± 13.9	0.0 ± 0	55.7 ± 10	0.08 ± 0.2
Ethiopia	49.93 ± 9.2	0.02 ± 0.0	98.4 ± 0.5	0.0 ± 0	80.2 ± 19	−0.01 ± 0.0
Vietnam	64.56 ± 10.4	0.01 ± 0.2	81.7 ± 1.8	0.0 ± 0	74.8 ± 5.1	0.13 ± 0.2
Ecuador & Ethiopia	70.22 ± 5.5	0.36 ± 0.1	82.2 ± 0.2	0.0 ± 0	74.2 ± 8.1	0.07 ± 0.1
Ecuador & Vietnam	79.28 ± 1.6	0.56 ± 0.0	62.8 ± 0.8	0.0 ± 0	62.2 ± 1.1	0.10 ± 0.1
Ethiopia & Vietnam	51.07 ± 11	0.04 ± 0.0	98.4 ± 0.5	0.0 ± 0	80.4 ± 21	−0.02 ± 0.0
Ecuador, Ethiopia & Vietnam	65.25 ± 9.1	0.28 ± 0.1	82.1 ± 1.3	0.0 ± 0	0.00 ± 13	0.02 ± 0.07

Table 5

Independent validation Leptophlebiidae’s multiple country models. Results in bold indicate datasets providing reliable models based on performance criteria. Mean and standard deviation of CCI and κ were derived from threefold cross validation

Validation sets	Models with pooled dataset

	Ecuador & Ethiopia		Ecuador & Vietnam		Ethiopia & Vietnam		Ecuador, Ethiopia & Vietnam

	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std
Ecuador	79.7 ± 7.3	0.60 ± 0.14	84.7 ± 5.8	0.69 ± 0.1	51.9 ± 0.7	0.00 ± 0.0	91.3 ± 4.1	0.83 ± 0.08
Ethiopia	78.3 ± 2.3	0.00 ± 0.0	56.5 ± 13.2	0.03 ± 0.01	98.4 ± 0.5	0.00 ± 0.0	86.8 ± 9.1	0.01 ± 0.05
Vietnam	64.5 ± 12.0	−0.10 ± 0.1	81.8 ± 4.0	0.32 ± 0.1	81.7 ± 1.8	0.00 ± 0.0	72.4 ± 18	0.10 ± 0.3
Ecuador & Ethiopia	88.1 ± 2.3	0.60 ± 0.01	73.8 ± 5.7	0.39 ± 0.05	82.1 ± 0.2	0.00 ± 0.0	94.7 ± 1.3	0.81 ± 0.06
Ecuador & Vietnam	82.3 ± 5.4	0.61 ± 0.1	73.2 ± 2.5	0.42 ± 0.07	62.8 ± 0.8	0.00 ± 0.0	89.6 ± 1.7	0.76 ± 0.05
Ethiopia & Vietnam	75.6 ± 5.8	0.06 ± 0.01	55.7 ± 12.5	0.05 ± 0.03	98.4 ± 0.5	0.02 ± 0.0	86.2 ± 11.3	−0.02 ± 0.02
Ecuador, Ethiopia & Vietnam	94.0 ± 0.0	0.80 ± 0.02	72.1 ± 7.2	0.40 ± 0.09	82.1 ± 1.3	0.00 ± 0.0	92.7 ± 1.2	0.74 ± 0.01

Table 6

Independent validation of Hydroptilidae’s multiple country models. Results in bold indicate datasets providing reliable models based on performance criteria. Mean and standard deviations of CCI and κ were derived from threefold cross validation

Validation sets	Models with pooled dataset

	Ecuador & Ethiopia		Ecuador & Vietnam		Ethiopia & Vietnam		Ecuador, Ethiopia & Vietnam

	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std	Mean CCI ± std	Mean κ ± std
Ecuador	89.7 ± 5.9	0.00 ± 0.0	85.6 ± 2.2	0.21 ± 0.2	85.6 ± 0.2	0.08 ± 0.12	87.5 ± 2.6	0.38 ± 0.1
Ethiopia	93.0 ± 7.1	0.00 ± 0.0	85.8 ± 9.5	−0.01 ± 0.03	97.4 ± 0.9	−0.01 ± 0.01	96.7 ± 0.9	−0.01 ± 0.0
Vietnam	56.7 ± 1.2	0.00 ± 0.0	78.8 ± 16	0.52 ± 0.4	63.7 ± 10.8	0.18 ± 0.26	65.5 ± 13	0.23 ± 0.3
Ecuador & Ethiopia	94.4 ± 1.0	0.03 ± 0.0	80.8 ± 9.4	−0.05 ± 0.05	90.4 ± 4.5	0.07 ± 0.10	95.4 ± 0.5	0.37 ± 0.1
Ecuador & Vietnam	74.4 ± 0.2	0.00 ± 0.0	71.9 ± 5.1	0.16 ± 0.2	77.4 ± 4.4	0.16 ± 0.23	78.4 ± 3.8	0.24 ± 0.2
Ethiopia & Vietnam	98.0 ± 0.8	0.00 ± 0.0	87.1 ± 5.8	0.01 ± 0.04	97.7 ± 1.2	−0.01 ± 0.01	97.7 ± 0.9	0.01 ± 0.02
Ecuador, Ethiopia & Vietnam	94.0 ± 1.6	0.00 ± 0.0	82.1 ± 8.2	0.08 ± 0.2	94.3 ± 1.3	0.07 ± 0.1	90.7 ± 1.7	0.00 ± 0.06

Table 7

Major variables of classification trees for Leptophlebiidae and the times these variables were selected at each level in the tree. The variables selected in the topmost level of the tree represent a major variable influencing the occurrence of Leptophlebiidae

Fold	Level in the tree

	1st (topmost)	2nd	3rd	4th	5th	6th (bottommost)
Ecuador model 1	DO	Stream velocity (2×) *	Water temperature (2×) *	Stream velocity
Ecuador model 2	Conductivity	Season	Water temperature	Stream velocity
Ecuador model 2			Stream velocity
Ecuador model 3	Conductivity	Stream velocity	DO	Water temperature
Ecuador model 3			Season	Conductivity
Ecuador + Ethiopia model 1	Conductivity	DO (2×) *	Season	Stream velocity	Water temperature	Stream velocity
			pH	Season	pH
			Water temperature
Ecuador + Ethiopia model 2	Conductivity	Season	DO	pH	DO
Ecuador + Ethiopia model 2		DO	Water temperature (2×) *	Stream velocity
Ecuador + Ethiopia model 3	Conductivity	pH	DO (2×) *	Water temperature
		DO	Water temperature Stream velocity	DO
			Season
			Conductivity
Ecuador + Ethiopia + Vietnam model 1	Conductivity	DO (2×) *	Season	Stream velocity	Water temperature	Stream velocity
			pH	Season	pH
			Water temperature
Ecuador + Ethiopia + Vietnam model 2	Conductivity	Season	DO	pH	DO
Ecuador + Ethiopia + Vietnam model 2		DO	Water temperature (2×) *	Stream velocity
Ecuador + Ethiopia + Vietnam model 3	Conductivity	pH	DO (2×) *	Stream velocity	Water temperature
		DO	Water temperature	Season	DO
				Conductivity
Times selected
Eight	Conductivity		Water temperature
Seven				Stream velocity
Six		DO
Five			DO
Four			Season
Three		Season		Season	DO
Two		Stream velocity	Stream velocity	Conductivity	pH	Stream velocity
		pH	pH	pH	Water temperature
				Water temperature
One	DO		Conductivity	DO

Selected twice.

Table 8

Major variables of classification trees for Hydroptilidae and the times these variables were selected at each level in the tree. The variables selected in the topmost level of the tree represent a major variable influencing the occurrence of Hydroptilidae

Fold	Level in the tree

	1st (topmost)	2nd	3rd	4th	5th	6th	7th (bottommost)
Ecuador + Vietnam model 1	pH	Stream velocity (2×) *	Conductivity	Season	Water temperature	Conductivity	Conductivity
Ecuador + Vietnam model 1			Season	Stream velocity	Conductivity
Ecuador + Vietnam model 2	Season	Stream velocity	DO	Water temperature
Ecuador + Vietnam model 2			Conductivity
Ecuador + Vietnam model 3	Water temperature	Stream velocity	DO
Ecuador + Vietnam model 3		Season
Times selected
Three		Stream velocity
Two			Conductivity
Two			DO
One	pH		Season	Season	Water temperature	Conductivity	Conductivity
	Season			Stream velocity	Conductivity
	Water temperature			Water temperature

Selected twice.

4. Discussion

4.1. Model performance

Classification trees successfully modelled the occurrence of some macroinvertebrate families occurring in the three tropical countries. As the modelling technique is robust to outliers, all observations were included in the analysis. Furthermore, classification trees are nonparametric and nonlinear. Thus, there are no inherent assumptions that the underlying relationships between the predictor and target variable are linear [53]. Although classification trees can be quite unstable with respect to their predictive performances [33], some of our results were stable, particularly the models developed for Leptophlebiidae. This shows that the occurrence of Leptophlebiidae has consistent patterns in the datasets. Furthermore, classification trees are transparent and have the ability to deal with relatively small data sets. However, when large data sets are being processed, this can lead to large and complex trees, therefore adequate parameterisations are very crucial. Most decision trees are merely constructed in purely data-driven manner, wherein incorporation of expert knowledge is not integrated. In this study, the dataset is relatively small and incorporation of expert knowledge in the model would have been very hard, as the available ecological knowledge especially for tropical countries is minimal [60]. There are other modelling techniques such as fuzzy logic, Artificial Neural Networks (ANN), Generalised Linear Models (GLM), and Bayesian Belief Network (BBN). Each of these modelling techniques has its own advantages and drawbacks. This is elaborated by Van Echelpoel et al. [75]. Both fuzzy logic and BBN require knowledge-based rules and were therefore suboptimal for our study. GLM assumes that the underlying data are characterized by specific distributions. Some of our testing and validation sets do not comply with this distribution and therefore GLM cannot be applied in a standardized and efficient manner. Lastly, ANN can lead to highly reliable models, but the drawbacks and challenges of this method are lack of guidelines for optimal design, low ecological relevance and limited explanatory power.

Most values of Cohen’s kappa statistic (κ) were lower than 0.4 and CCIs were lower than 70%. This indicates that most models did not make reliable predictions. However, some classification tree models could reliably predict some macroinvertebrate taxa (i.e. Leptophlebiidae and Hydroptilidae). The incapability of models to yield reliable predictions is due to various reasons. For instance, there were only 5 environmental variables common among the three river basins. Other important variables (e.g. physical habitat degradation) than the one used in this study were not included during model development. In literature, it has been shown that macroinvertebrates distribution patterns are determined by conductivity, pH, flow, substrate and marginal vegetation [59]; substratum and historical flow characteristics [16]; pH [63]; nutrients, velocity and sediment type [3]. Environmental variables such as velocity, dissolved oxygen, chlorophyll, chemical oxygen demand, sediment type, land use and elevation [31] and oxygen level, temperature, ammonium concentration and conductivity [74] are known to influence the water quality index based on macroinvertebrates. Common variables reported in literature (velocity, conductivity, dissolved oxygen, temperature, pH) were included in the development of our models. The inclusion of only 5 environmental variables in the modelling exercise is sufficient as these variables cover a wide range of water quality issues. For instance, conductivity integrates minerals and inorganic pollutants [26]. Dissolved oxygen gives an indication of organic pollution whereas temperature may represent global warming [73].

Records for taxon absences may be due to various reasons. Taxa might not be present on the site due to the unsuitability of the habitat. However, incorrect records of absences could lead to incorrect model prediction. Sampling error could have caused the record of a taxon which is present as absent. The habitat might be suitable but the taxa could be temporarily absent. Furthermore, migration barriers could have caused the absences of these taxa despite the high suitability of the habitat [46]. These could have contributed to the lower accuracy of the models.

Models developed for most taxa were not reliable. Additionally, the model developed for Hydroptilidae based on the combined Ecuadorian and Vietnamese dataset was reliably validated with the Vietnamese dataset. The validation results were quite unstable as the standard deviation of κ was more than 75% of the mean κ. Moreover, the selected variables varied among models developed from threefold cross validation (Table 8). Aside from predictive performance, model stability is also important [31]. Thus, the validation results of Hydroptilidae cannot be considered valid.

Overall, it was not possible to develop a reliable and applicable model to predict the occurrences of Chironomidae and Baetidae. It is reported that Baetidae is influenced by the vegetation cover [51], pH and water temperature [9] and cobbles [2]. It can also be present on moderately disturbed rivers [11]. Chironomidae can occur at both disturbed and pristine conditions [47]. As both taxa can be present in a wide range of environmental conditions, developing reliable and valid models for these taxa can be challenging.

Models of Libellulidae could not be successfully developed, which could be due to the limited number of available input variables. Libellulidae need a continuous spectrum of stagnant, transitional and flowing water [38]. This variable is not incorporated by the input variables. Furthermore, Libellulidae are reported as a pollution-tolerant family in Vietnam [58] while it is moderately tolerant to pollution in South America [66]. Thus, Libellulidae can be present over a wide range of environmental conditions in some tropical regions. As there were no successful extrapolated-models developed and validated, these results indicate that Chiromidae, Baetidae, Hydroptilidae and Libellulidae may respond differently to pollution among the three river basins.

Literature has varying reports related to the dependence of Cohen’s kappa on prevalence. According to Feinstein and Cicchetti [27], κ is affected by prevalence and they recommended to design the research with well-balanced positive and negative constituents. However, Manel et al. [48] concluded that the effect of the frequency of occurrence of target organisms on Cohen’s Kappa is negligible. Our results revealed that the model of Hydroptilidae developed from the training set of Vietnam (prevalence of 43%) or Ethiopia (prevalence of 1.7%) was not reliable, whereas, the model of Hydroptilidae developed from the combined training set of Ethiopia and Vietnam (prevalence of 7%) resulted in better performing models. Furthermore, no valid and reliable Libellulidae models were developed from the training set comprising the Ecuadorian dataset (prevalence of 48%) or the Ethiopian dataset (prevalence of 49%). Prevalence of macroinvertebrate taxa may affect model’s performance. However, as illustrated in our study, it may not always be the case. Other reasons could have affected the poor performance of these models.

Valid models which consist of pooled datasets were developed for Leptophlebiidae. A model containing the Ecuadorian dataset generally resulted in a valid and reliable model. The prevalence of Leptophlebiidae in Ecuador is almost 50% and the data may contain clear patterns of the occurrence of the taxon. The low prevalence of Leptophlebiidae in the Ethiopian or Vietnamese dataset could have resulted in an unclear pattern of its occurrence. However, when the models were developed in combination with the Ecuadorian data, the model performances improved. This indicates that the patterns of the occurrence of Leptophlebiidae were magnified when the Ecuadorian dataset was combined with either the Ethiopian or Vietnamese dataset. Hence, similar patterns are possibly observed in the dataset between Ecuador and Vietnam and between Ecuador and Ethiopia. Leptophlebiidae occurring in these river basins possibly respond similarly to environmental pressures. On the other hand, no reliable models were developed from the training sets comprising the Ethiopian and Vietnamese data. The poor performance of these models could be due to both or either the low prevalence of Leptophlebiidae in these datasets and/or the differences in the patterns of occurrence of Leptophlebiidae between these countries. Thus, the response of Leptophlebiidae towards environmental pressures between the Ethiopian and Vietnamese river basins remains unclear.

4.2. Study design

The altitude of the Vietnamese river basin is lower compared to the other river basins. Ecuador shares common altitude with Ethiopia although some sites in Ethiopia are 1000 m lower than Ecuador. According to Rezende et al. [65] and Feio et al. [28], higher altitudes were likely the primary variables that increased taxonomic richness and density of macroinvertebrate communities. However, the change in taxonomic diversity in response to elevation is often related to the impacts of human activities [44]. Macroinvertebrate assemblage composition may be affected by altitude, although such effects are often confounded with anthropogenic disturbance, which is often most extensive at lowest altitudes in the tropics [55,76].

One of the strengths of our study design is the incorporation of seasonality in model development. Seasonal variations in subtropical countries may have an effect on macroinvertebrate assemblage [1,10]. The uncertainties due to seasonal variability were integrated into the models. This is confirmed in our results, wherein the occurrence of Leptophlebiidae and Hydroptilidae were influenced by seasonal changes. Leptophlebiidae favours wet season while Hydroptilidae can be present in both wet and dry season.

The sample size is different resulting in an unbalanced sample size between countries. This is one of the limitations of the study design. However the inclusion of all cases is important for an optimal visualization of patterns in each dataset. A large number of observations or a large sample size yields better model accuracy and prediction [70,77]. However, Yu and Abdel-Aty [78] concluded that smaller sample size enhances the model’s classification accuracy. Furthermore, large datasets have a disadvantage in classification trees models as they can result in complex models, which are hard to interpret. Despite the limited number of sampling sites (15–50), sampling occurred in a standardized way giving a relevant dataset that allows the extraction of possible ecological patterns.

A balanced prevalence of each taxon could have improved model’s performance and clarity of results. However, each sampling campaign was conducted independently. A dataset with balanced prevalence of target taxa is ideal however also very rare. Nevertheless, the modelling of representative taxa with varying degree of tolerance to pollution gives valuable ecological insights from different regions with similar climatic conditions.

4.3. Monitoring and modelling of ecological data in tropical regions

This modelling exercise explores the possibility of developing a reliable model from a combined dataset or the extrapolation of models to other river basins with similar environmental conditions and pressures. Among the taxa analysed, the most sensitive taxon (Leptophlebiidae) was successfully modelled. Leptophlebiidae seems to share similar patterns between Ecuador and Vietnam and between Ecuador and Ethiopia. Thus, the use of models developed in Ecuador can be extrapolated to either Vietnamese or Ethiopian river basins with similar environmental conditions and pressures. The extrapolation of models to other river basins allows the monitoring of a lesser number of sites. As a consequence, monitoring becomes cheaper which is beneficial in developing countries. Furthermore, the results of our study open the possibility for an implementation of a general guideline for bioassessment in tropical countries based on pollution sensitive taxa.

Results revealed that the major variables predicting Leptophlebiidae are conductivity and dissolved oxygen. Thus, a river basin manager can prioritize management activities that alleviate conductivity and dissolved oxygen issues in rivers. Information on the influence of environmental variables and distribution of this taxon is limited [60,69]. However, in a study, Leptophlebiidae was not found when water quality was reduced by use of agricultural chemicals [41]. Leptophlebiidae also disappears in a river stretch recipient of industrial effluent and domestic sewage [61]. Conductivity appeared to be an important driver of macroinverterbrate community structure within this river stretch. The variable conductivity can be related to effluents, sewage and agricultural chemicals, which are in line with the results of our study.

Although numerous data were collected, common variables among river basins were only limitedly available. The differences in biomonitoring programmes among the three countries resulted in few number of variables available for model development. Thus, the same standardized biomonitoring programme among countries is suggested to maximize the development of models on similar systems. We suggest the development of a guideline or framework for a standardized monitoring programme in tropical countries. As monitoring programmes may be limited by the financial and personnel resources, a cost-effective monitoring is recommended. The distribution of aquatic organisms is a result of the interaction among their ecological role, the physical conditions that characterize the habitat and food availability [52]. Thus, designing the biomonitoring programme including the basic environmental variables such as physical–chemical measurements and hydromorphological description is suggested. In addition, as unbalanced prevalence of the target species might influence the performance of the models, selecting sampling sites that would result in a more balanced prevalence of the target species is recommended to remove the bias on the effect of prevalence in modelling.

Three main conclusions could be drawn from this modelling study. First, pollution sensitive taxa are more easily modelled than pollution tolerant taxa. Secondly, In general, responses of macroinvertebrates towards pollution were different among countries except for the pollution sensitive taxa. Thus, extrapolation of ecological models for sensitive taxa to another river basin with similar climatic and environmental conditions is possible. Lastly, the implementation of a standardized biomonitoring programme in tropical countries is suggested. Nevertheless, this type of systematic analyses for all taxa is necessary to determine and minimize uncertainty in ecological assessment.

Footnotes

Acknowledgements

We would like to thank all the people who contributed to the sampling campaigns. Marie Anne Eurie Forio receives financial support from the special research fund of Ghent University to support the VLIR Ecuador Biodiversity Network. Luis Domiguez-Granda received financial support of the VLIR ESPOL IUC program in Ecuador and SENACYT. Argaw Ambelu was a recipient of an ICP-PhD scholarship from VLIR-UOS. Seid Tiku Mereta was a recipient of an IUC-PhD scholarship from VLIR-UOS (IUC JIMMA). Thu Huong Hoang received financial aid from the Belgian Technical Cooperation (BTC).

Appendix Table 12

Independent validation of Chironomidae’s multiple country models. Mean and standard deviation of CCI and κ were derived from threefold cross validation

Validation sets	Models with pooled dataset

	Ecuador & Ethiopia		Ecuador & Vietnam		Ethiopia & Vietnam		Ecuador, Ethiopia & Vietnam

	CCI	K	CCI	K	CCI	K	CCI	K
Ecuador	91.3 ± 0.1	0.00 ± 0.00	91.3 ± 0.1	0.00 ± 0.00	90.4 ± 1.5	−0.02 ± 0.02	91.3 ± 0.2	0.00 ± 0.00
Ethiopia	84.3 ± 0.7	0.00 ± 0.00	84.3 ± 0.7	0.00 ± 0.00	84.9 ± 0.4	0.06 ± 0.09	84.3 ± 0.7	0.00 ± 0.00
Vietnam	93.4 ± 2.1	0.00 ± 0.00	93.4 ± 2.1	0.00 ± 0.00	93.4 ± 2.1	0.00 ± 0.00	93.4 ± 2.1	0.00 ± 0.00
Ecuador & Ethiopia	86.1 ± 0.7	0.00 ± 0.00	86.1 ± 0.7	0.00 ± 0.00	86.8 ± 1.2	0.07 ± 0.11	86.1 ± 0.7	0.00 ± 0.00
Ecuador & Vietnam	92.1 ± 0.8	0.00 ± 0.00	90.6 ± 1.8	0.00 ± 0.00	91.5 ± 0.9	−0.01 ± 0.01	92.1 ± 0.8	0.00 ± 0.00
Ethiopia & Vietnam	84.1 ± 1.4	0.00 ± 0.00	84.1 ± 1.3	0.00 ± 0.00	84.4 ± 1.2	0.03 ± 0.04	84.1 ± 1.4	0.00 ± 0.00
Ecuador, Ethiopia & Vietnam	86.4 ± 1.4	0.00 ± 0.00	86.4 ± 1.4	0.00 ± 0.00	87.1 ± 2.3	0.09 ± 0.12	86.4 ± 1.4	0.00 ± 0.00

References

A.C.F.

Aguiar,

Gucker,

Brauns,

Hille and

I.G.

Boechat, Benthic invertebrate density, biomass, and instantaneous secondary production along a fifth-order human-impacted tropical river, Environmental Science and Pollution Research22 (2015), 9864–9876. doi:10.1007/s11356-015-4170-y.

A.K.

Ahmad,

A.S.

Hafizah and

Shuhaimi-Othman, Potential use of benthic macroinvertebrate as biological indicators in Sungai Ikan, Hulu Terengganu, Terengganu, Sains Malaysiana44 (2015), 663–670. doi:10.17576/jsm-2015-4405-04.

Alvarez-Mieles,

Irvine,

A.V.

Griensven,

Arias-Hidalgo,

Torres and

A.E.

Mynett, Relationships between aquatic biotic communities and water quality in a tropical river-wetland system (Ecuador), Environmental Science & Policy34 (2013), 115–127. doi:10.1016/j.envsci.2013.01.011.

Ambelu,

Lock and

Goethals, Comparison of modelling techniques to predict macroinvertebrate community composition in rivers of Ethiopia, Ecological Informatics5 (2010), 147–152. doi:10.1016/j.ecoinf.2009.12.004.

Boets,

Lock and

P.L.M.

Goethals, Modelling habitat preference, abundance and species richness of alien macrocrustaceans in surface waters in Flanders (Belgium) using decision trees, Ecological Informatics17 (2013), 73–81. doi:10.1016/j.ecoinf.2012.06.001.

Bonada,

Doledec and

Statzner, Taxonomic and biological trait differences of stream macroinvertebrate communities between Mediterranean and temperate regions: Implications for future climatic scenarios, Global Change Biology13 (2007), 1658–1671. doi:10.1111/j.1365-2486.2007.01375.x.

R.W.

Bouchard, Guide to Aquatic Macroinvertebrates of the Upper Midwest, Water Resources Center, University of Minnesota, St. Paul, MN, 2004.

Broothaerts,

Kissi,

Poesen,

Van Rompaey,

Getahun,

Van Ranstet al., Spatial patterns, causes and consequences of landslides in the Gilgel Gibe catchment, SW Ethiopia, Catena97 (2012), 127–136. doi:10.1016/j.catena.2012.05.011.

Caldararu,

M.C.

Topa,

Murariu,

Sangiorgio,

Basset and

L.P.

Georgescu, Ecological research for the future of our water – Interactions between biotope and biocoenosis, Journal of Environmental Protection and Ecology14 (2013), 856–863.

10.

Carrie,

Dobson and

Barlow, The influence of geology and season on macroinvertebrates in Belizean streams: Implications for tropical bioassessment, Freshwater Science34 (2015), 648–662. doi:10.1086/681541.

11.

M.M.

Castillo and

Perez, Varipes (Ephemeroptera: Baetidae) in Venezuela: Description of a new species, Revista Colombiana De Entomologia37 (2011), 346–349.

12.

R.K.

Chesters, Biological monitoring working party, The 1978 national testing exercise, Department of the Environment, UK, Water data unit technical memorandum 19, 1980.

13.

Cohen, A coefficient of agreement for nominal scales, Educational and Psychological Measurement20 (1960), 37–46. doi:10.1177/001316446002000104.

14.

Colombo and

Maran, Gilgel Gibe II hydroelectric project environmental impact assessment main report, Report A4511403 by Centro Electtrotecnico Sperimentale Italiano (CESI), 2004.

15.

N.T.

Dang, Freshwater Macroinvertebrates in Northern Vietnam, Publishing House of Science and Technology, Hanoi, Vietnam, 1980.

16.

A.M.

Davis,

R.G.

Pearson,

I.J.

Kneipp,

L.J.

Benson and

Fernandes, Spatiotemporal variability and environmental determinants of invertebrate assemblage structure in an Australian dry-tropical river, Freshwater Science34 (2015), 634–647. doi:10.1086/681303.

17.

T.A.

Demissie,

Saathoff,

Y.A.

Sileshi and

Gebissa, Climate change impacts on the stream flow and simulated flux to Gilgel Gibe I hydropower reservoir – Ethiopia, European International Journal of Science and Technology2 (2013), 64–77.

18.

Doledec,

Phillips and

Townsend, Invertebrate community responses to land use at a broad spatial scale: Trait and taxonomic measures compared in New Zealand rivers, Freshwater Biology56 (2011), 1670–1688. doi:10.1111/j.1365-2427.2011.02597.x.

19.

Domínguez,

M.D.

Hubbard and

M.L.

Pescador, Los ephemeroptera en Argentina, Fauna de agua dulce de la Republica Argentina33 (1994), 1–142.

20.

Dominguez-Granda, Indices based on macroinvertebrate communities for assessment of the quality of the Chaguana river in Ecuador, PhD Dissertation, Ghent University, Gent, Belgium, 2007.

21.

Dominguez-Granda,

Lock and

P.L.M.

Goethals, Using multi-target clustering trees as a tool to predict biological water quality indices based on benthic macroinvertebrates and environmental parameters in the Chaguana watershed (Ecuador), Ecological Informatics6 (2011), 303–308. doi:10.1016/j.ecoinf.2011.05.004.

22.

Dudgeon, Tropical Asian Streams. Zoobenthos, Ecology and Conservation, Hongkong University Press, Hongkong, 1999.

23.

Dynesius,

Jansson,

M.E.

Johansson and

Nilsson, Intercontinental similarities in riparian-plant diversity and sensitivity to river regulation, Ecological Applications14 (2004), 173–191. doi:10.1890/02-5127.

24.

Dzeroski,

Grbovic,

W.J.

Walley and

Kompare, Using machine learning techniques in the construction of models. 2. Data analysis with rule induction, Ecological Modelling95 (1997), 95–111. doi:10.1016/S0304-3800(96)00029-4.

25.

Everaert,

Boets,

Lock,

Dzeroski and

P.L.M.

Goethals, Using classification trees to analyze the impact of exotic species on the ecological assessment of polder lakes in Flanders, Belgium, Ecological Modelling222 (2011), 2202–2212. doi:10.1016/j.ecolmodel.2010.08.013.

26.

Everaert,

De Neve,

Boets,

Dominguez-Granda,

S.T.

Mereta,

Ambeluet al., Comparison of the abiotic preferences of macroinvertebrates in Tropical River Basins, Plos One9 (2014).

27.

A.R.

Feinstein and

D.V.

Cicchetti, High agreement but low kappa. 1. The problems of 2 paradoxes, Journal of Clinical Epidemiology43 (1990), 543–549. doi:10.1016/0895-4356(90)90158-L.

28.

M.J.

Feio,

W.R.

Ferreira,

D.R.

Macedo,

A.P.

Eller,

C.B.M.

Alves,

J.S.

Francaet al., Defining and testing targets for the recovery of tropical streams based on macroinvertebrate communities and abiotic conditions, River Research and Applications31 (2015), 70–84. doi:10.1002/rra.2716.

29.

H.R.

Fernández and

Dominguez, Guía Para la Determinación de Los Artrópodos Bentónicos Sudamericanos, Universidad Nacional de Tucumán, Tucumán, Argentina, 2001.

30.

A.H.

Fielding and

J.F.

Bell, A review of methods for the assessment of prediction errors in conservation presence/absence models, Environmental Conservation24 (1997), 38–49. doi:10.1017/S0376892997000088.

31.

M.A.E.

Forio,

Landuyt,

Bennetsen,

Lock,

T.H.T.

Nguyen,

M.N.D.

Ambaritaet al., Bayesian belief network models to analyse and predict ecological water quality in rivers, Ecological Modelling312 (2015), 222–238. doi:10.1016/j.ecolmodel.2015.05.025.

32.

Gabriels,

Lock,

De Pauw and

P.L.M.

Goethals, Multimetric Macroinvertebrate Index Flanders (MMIF) for biological assessment of rivers and lakes in Flanders (Belgium), Limnologica40 (2010), 199–207. doi:10.1016/j.limno.2009.10.001.

33.

P.L.M.

Goethals, Data driven development of predictive ecological models for benthic macroinvertebrates in rivers, PhD dissertation, Ghent University, Gent, Belgium, 2005.

34.

P.L.M.

Goethals,

A.P.

Dedecker,

Gabriels,

Lek and

Pauw, Applications of artificial neural networks predicting macroinvertebrates in freshwaters, Aquatic Ecology41 (2007), 491–508. doi:10.1007/s10452-007-9093-3.

35.

Gonzalo and

J.A.

Camargo, The impact of an industrial effluent on the water quality, submersed macrophytes and benthic macroinvertebrates in a dammed river of central Spain, Chemosphere93 (2013), 1117–1124. doi:10.1016/j.chemosphere.2013.06.032.

36.

Hall,

Frank,

Holmes,

Pfahringer,

Reutemann and

I.A.

Witten, The WEKA data mining software: An update, SIGKDD Explorations11(1) (2009), 10–18. doi:10.1145/1656274.1656278.

37.

He,

Cheng,

Zhou,

Keo and

Goethals, Monitoring and assessment of water quality and ecology, Sustainability of Water Quality and Ecology3–4 (2014), 1–2. doi:10.1016/j.swaqe.2015.02.001.

38.

Higashikawa,

Yoshimura,

Yagi and

Maeto, Microhabitat use by larvae of the endangered dragonfly Sympetrum pedemontanum elatum (Selys) in Japan, Journal of Insect Conservation20 (2016), 407–416. doi:10.1007/s10841-016-9874-x.

39.

T.H.

Hoang,

Lock,

Mouton and

P.L.M.

Goethals, Application of classification trees and support vector machines to model the presence of macroinvertebrates in rivers in Vietnam, Ecological Informatics5 (2010), 140–146. doi:10.1016/j.ecoinf.2009.12.001.

40.

J.E.

Holguin-Gonzalez,

Everaert,

Boets,

Galvis and

P.L.M.

Goethals, Development and application of an integrated ecological modelling framework to analyze the impact of wastewater discharges on the ecological water quality of rivers, Environmental Modelling & Software48 (2013), 27–36. doi:10.1016/j.envsoft.2013.06.004.

41.

Jandry,

Brulin,

Parinet and

Grandjean, Ephemeroptera communities as bioindicators of the suitability of headwater streams for restocking with white-clawed crayfish, Austropotamobius pallipes, Ecological Indicators46 (2014), 560–565.

42.

D.K.

Kajungu,

Selemani,

Masanja,

Baraka,

Njozi,

Khatibet al., Using classification tree modelling to investigate drug prescription practices at health facilities in rural Tanzania, Malaria Journal11 (2012), 311. doi:10.1186/1475-2875-11-311.

43.

A.P.

Kirilenko and

Stepchenkova, Inter-coder agreement in one-to-many classification: Fuzzy kappa, Plos One11 (2016).

44.

Lang and

Reymond, Empirical relationships between diversity of invertebrate communities and altitude in rivers – Application to biomonitoring, Aquatic Sciences55 (1993), 188–196. doi:10.1007/BF00877448.

45.

Lock,

Asenova and

P.L.M.

Goethals, Benthic macroinvertebrates as indicators of the water quality in Bulgaria: A case-study in the Iskar river basin, Limnologica – Ecology and Management of Inland Waters41 (2011), 334–338. doi:10.1016/j.limno.2011.03.002.

46.

Lütolf,

Kienast and

Guisan, The ghost of past species occurrence: Improving species distribution models for presence-only data, Journal of Applied Ecology43 (2006), 802–815. doi:10.1111/j.1365-2664.2006.01191.x.

47.

T.J.

Mackintosh,

J.A.

Davis and

R.M.

Thompson, The influence of urbanisation on macroinvertebrate biodiversity in constructed stormwater wetlands, Science of the Total Environment536 (2015), 527–537. doi:10.1016/j.scitotenv.2015.07.066.

48.

Manel,

H.C.

Williams and

S.J.

Ormerod, Evaluating presence-absence models in ecology: The need to account for prevalence, Journal of Applied Ecology38 (2001), 921–931. doi:10.1046/j.1365-2664.2001.00647.x.

49.

Matamoros, Predicting river concentrations of pesticides from banana plantations under data-poor conditions, PhD dissertation, Ghent University, Gent, Belgium, 2004.

50.

W.P.

McCafferty and

A.V.

Provonsha, Aquatic Entomology: The Fishermen’s and Ecologists’ Illustrated Guide to Insects and Their Relatives, Jones and Bartlett Publishers, Boston, USA, 1983.

51.

S.T.

Mereta,

Boets,

A.A.

Bayih,

Malu,

Ephrem,

Sisayet al., Analysis of environmental factors determining the abundance and diversity of macroinvertebrate taxa in natural wetlands of southwest Ethiopia, Ecological Informatics7 (2012), 52–61. doi:10.1016/j.ecoinf.2011.11.005.

52.

R.W.

Merrit and

K.W.

Cummins, An Introduction to the Aquatic Insects of North America, Kendal/Hunt Publishing Company, Dubuque, IA, 1984.

53.

G.G.

Moisen, Classification and regression trees A2, in: Encyclopedia of Ecology,

B.D.

Fath and

S.E.

Jørgensen, eds, Academic Press, Oxford, UK, 2008, pp. 582–588.

54.

MONRE, National Strategy for Water Resources up to 2020: Cultural and Information, Publish House, Hanoi, Vietnam, 2006.

55.

F.A.

Montano-Centellas and

Garitano-Zavala, Andean bird responses to human disturbances along an elevational gradient, Acta Oecologica – International Journal of Ecology65–66 (2015), 51–60. doi:10.1016/j.actao.2015.05.003.

56.

J.C.

Morse,

Y.J.

Bae,

Munkhjargal,

Sangpradub,

Tanida,

T.S.

Vshivkovaet al., Freshwater biomonitoring with macroinvertebrates in East Asia, Frontiers in Ecology and the Environment5 (2007), 33–42. doi:10.1890/1540-9295(2007)5[33:FBWMIE]2.0.CO;2.

57.

H.H.

Nguyen,

Everaert,

Gabriels,

T.H.

Hoang and

P.L.M.

Goethals, A multimetric macroinvertebrate index for assessing the water quality of the Cau river basin in Vietnam, Limnologica45 (2014), 16–23. doi:10.1016/j.limno.2013.10.001.

58.

X.Q.

Nguyen,

D.Y.

Mai,

Pinder and

Tillings, Biological surveillance of freshwaters: A practical manual and identification key for use, in: Vietnam, National University Publishers, Hanoi, Vietnam, 2004.

59.

A.S.

Niba and

S.P.

Mafereka, Benthic macroinvertebrate assemblage composition and distribution pattern in the upper Mthatha River, Eastern Cape, South African Journal of Aquatic Science40(2) (2015), 133–142. doi:10.2989/16085914.2015.1028323.

60.

Nolivos,

Villacís,

R.F.

Vázquez,

D.E.

Mora,

Domínguez-Granda,

Hampelet al., Challenges for a sustainable management of Ecuadorian water resources, Sustainability of Water Quality and Ecology6 (2015), 101–106. doi:10.1016/j.swaqe.2015.02.002.

61.

Ordonez,

V.L.

Lougheed,

J.L.

Gardea-Torresdey and

L.J.

Bain, Impact of metals on macroinvertebrate assemblages in the forgotten stretch of the Rio Grande, Archives of Environmental Contamination and Toxicology60 (2011), 426–436. doi:10.1007/s00244-010-9557-7.

62.

Peru and

Doledec, From compositional to functional biodiversity metrics in bioassessment: A case study using stream macroinvertebrate communities, Ecological Indicators10 (2010), 1025–1036. doi:10.1016/j.ecolind.2010.02.011.

63.

Pinto,

B.L.

Maheshwari and

E.C.

Morris, Understanding the relationships among phytoplankton, benthic macroinvertebrates, and water quality variables in Peri-Urban river systems, Water Environment Research86 (2014), 2279–2293. doi:10.2175/106143014X13975035526220.

64.

V.H.

Resh and

D.M.

Rosenberg, Recent trends in life-history research on benthic macroinvertebrates, Journal of the North American Benthological Society29 (2010), 207–219. doi:10.1899/08-082.1.

65.

R.S.

Rezende,

A.M.

Santos,

Henke-Oliveira and

J.F.

Goncalves, Effects of spatial and environmental factors on benthic a macroinvertebrate community, Zoologia31 (2014), 426–434. doi:10.1590/S1984-46702014005000001.

66.

Roldán , Bioindicacíon de la Calidad del Agua en Colombia. Uso del Método BMWP/Col, Universidad de Antioquia, Medellín, 2003.

67.

Roldán, Guía Para el Estudio de Los Macroinvertebrados Acuáticos del Departamento de Antioquia, Fondo FEN Colombia, Conciencias-Universidad de Antioquia, Santafé de Bogota, Colombia, 1988.

68.

Sadeghi,

Zarkami,

Sabetraftar and

Van Damme, Application of classification trees to model the distribution pattern of a new exotic species Azolla filiculoides (Lam.) at Selkeh Wildlife Refuge, Anzali wetland, Iran, Ecological Modelling243 (2012), 8–17. doi:10.1016/j.ecolmodel.2012.06.011.

69.

Shimano,

F.F.

Salles and

Juen, Study of the mayfly order Ephemeroptera (Insecta) in Brazil: A scienciometric review, Revista Brasileira De Entomologia57 (2013), 359–364. doi:10.1590/S0085-56262013005000043.

70.

D.R.B.

Stockwell and

A.T.

Peterson, Effects of sample size on accuracy of species distribution models, Ecological Modelling148 (2002), 1–13. doi:10.1016/S0304-3800(01)00388-X.

71.

Strobl,

Malley and

Tutz, An introduction to recursive partitioning: Rationale, application and characteristics of classification and regression trees, bagging and random forests, Psychological Methods14 (2009), 323–348. doi:10.1037/a0016973.

72.

Uhlenbrook,

Mohamed and

A.S.

Gragne, Analyzing catchment behavior through catchment modeling in the Gilgel Abay, Upper Blue Nile River Basin, Ethiopia, Hydrology and Earth System Sciences14 (2010), 2153–2165. doi:10.5194/hess-14-2153-2010.

73.

USEPA, (2016, April 12), Climate impacts on ecosystems.

74.

Van Ael,

De Cooman,

Blust and

Bervoets, Use of a macroinvertebrate based biotic index to estimate critical metal concentrations for good ecological water quality, Chemosphere119 (2015), 138–144. doi:10.1016/j.chemosphere.2014.06.001.

75.

Van Echelpoel,

Boets,

Landuyt,

Gobeyn,

Everaert,

Bennetsenet al., Chapter 6 – Species distribution models for sustainable ecosystem management, in: Developments in Environmental Modelling, Vol. 27,

S.L.C.B.

Young-Seuk Park and

J.S.

Erik, eds, Elsevier, 2015, pp. 115–134.

76.

D.L.A.

Vasquez,

Balslev and

Sklenar, Human impact on tropical-alpine plant diversity in the northern Andes, Biodiversity and Conservation24 (2015), 2673–2683. doi:10.1007/s10531-015-0954-0.

77.

M.S.

Wisz,

R.J.

Hijmans,

Li,

A.T.

Peterson,

C.H.

Graham,

Guisanet al., Effects of sample size on the performance of species distribution models, Diversity and Distributions14 (2008), 763–773. doi:10.1111/j.1472-4642.2008.00482.x.

78.

R.J.

Yu and

Abdel-Aty, Utilizing support vector machine in real-time crash risk evaluation, Accident Analysis and Prevention51 (2013), 252–259. doi:10.1016/j.aap.2012.11.027.

Analysing the effects of water quality on the occurrence of freshwater macroinvertebrate taxa among tropical river basins from different continents

Abstract

Keywords

1. Introduction

2. Materials and methods

2.1. Study areas

2.2. Data collection

2.3. Classification trees

Table 1 Selected macroinvertebrate families Macroinvertebrate family BMWP score [12] Chironomidae 2 Baetidae 4 Hydroptilidae 6 Libellulidae 8 Leptophlebiidae 10

4.1. Model performance

4.2. Study design

4.3. Monitoring and modelling of ecological data in tropical regions

Footnotes

Acknowledgements

References

Table 1
Selected macroinvertebrate families

Macroinvertebrate family BMWP score [12]

Chironomidae 2

Baetidae 4

Hydroptilidae 6

Libellulidae 8

Leptophlebiidae 10