The data used in this study include monthly fish richness (number of species), fish abundance (number of individuals), water levels, water quality, climate and distance from the sea. Here we used both air and water temperature as model predictors to understand if the variability of fish richness and abundance is influenced by climate or water quality change. Fish richness and abundance, water level and water quality data were obtained from the long-term fish and environmental monitoring programs of the Mekong River Commission (MRC); climate data were derived from the Climatic Research Unit (Harris et al., 2020) at http://www.cru.uea.ac.uk/data, and distance from the sea measured from the point of zero-elevation to each fish sampling site was calculated using ArcGIS.

Fish data were collected using the MRC’s Fisheries Programme’s standard sampling protocols described in MRC (2007) and Ngor et al. (2016). We selected eight fish sampling sites in the LMB, including six sites in the lower Mekong mainstream, one site in the Tonle Sap River and one site in the Bassac River (Figure 1) from June 2007 to May 2014. The sampling sites were chosen because fish and environmental time-series datasets were regularly collected. At each site, three professional fishers participated in a daily fish sampling program supervised by the MRC National Line Fisheries Agencies with technical coordination by a regional fish monitoring specialist. Generally, the sampling sites were relatively unchanged over the study period, and gillnets (length: 120 ± 50 m, height: 2–3.5 m, mesh size: 3–12 cm, daily soak hours: 12 ± 2) were the most commonly used fishing gear to fish in riverine habitats and fish catches by species were recorded for the study (Ngor et al., 2018b). Captured fish were identified to species level and counted to estimate fish diversity and abundance in the LMB. Some 500 species (from 78 families) were recorded in the long-term MRC fish monitoring program, and among these, about 95% of the catch belonged to Cyprinidae (80%) and other common families included Pangasiidae, Cobitidae, Siluridae and Clupeidae (Ngor, 2018). We used a fish list containing ~900 species from the Mekong Fish Database (MFD, 2003) and fish species were further cross-checked with FishBase (Froese and Pauly, 2017) and other sources (Rainboth et al., 2012; Kottelat, 2013). Mean daily fish abundance data from three fishers at each site were aggregated to obtain the site’s total monthly fish abundance used in our study.

Water level and water quality data were monitored daily by the MRC in partnership with their member countries. We used hydrological and water quality monitoring stations that are closest to the eight fish monitoring sites and averaged daily data to monthly data. Water quality variables used in this study were water temperature, pH, dissolved oxygen, water conductivity, total nitrite and nitrate (hereafter referred to as total nitrate), ammonium, total phosphorus, alkalinity, total suspended solids and chemical oxygen demand. Protocols for water quality sampling, preservation protocols and data analysis were based on the 20th edition of the Standard Method for the Examination of Water and Wastewater (Clescerl et al., 1998) or the MRC standard (MRC, 2019). Mean air temperature was used in this study. All the predictive variables for the model are given in Table 1.

The study used six common single-SMs namely Generalized Linear Model (GLM), Classification and Regression Tree (CART), Artificial Neural Networks (ANN), k-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Random Forest (RF). Detailed information about each SM was described in Lek and Guegan (1999), Sor et al. (2017b), Van Echelpoel and Goethals (2018), and Guo et al. (2019). All of these models used the same dataset, and thus each model was affected by equal bias. Given that there were some missing values in our dataset, our final analysis included monthly data from 606 samples and 14 predictive variables, which we used to predict monthly fish richness and abundance in the LMB.

Prior to developing the models, response variables were normalized between 0 and 1 to meet the requirement of modeling methods used in this study. Predictive variables were standardized to homogenize the relative strength of the predictors. The selected single-SMs were developed and validated using a k-fold cross-validation (CV) procedure. This procedure randomly splits the whole dataset into k non-overlapping folds, and then uses k-1 folds for the model development process and the remaining 1 fold for the validation. This process is repeated k times so that at the end, each of the folds is used once as the validation set. The k-fold cross-validation is a robust method that is commonly used to estimate the accuracy of models by averaging the k results to produce a single estimate for a model. The replication of the k-fold CV (15 replicates in our case) is often conducted to assure model performance stability. We used the 15-fold CV because the majority of model performance estimates were almost unbiased when k is between 10 and 20 folds (Kohavi, 1995). We also made three replicates of the 15-fold CV cross-validation (Figure 2). Consequently, we developed and validated a total of 45 statistical models for each of the six selected single-SMs in predicting both fish species richness and abundance in the LMB. In all, we built a total of 540 different statistical models for the six selected single-SMs (45 models × 6 single-SMs × 2) to predict fish richness and abundance in the LMB.

The predictions of all single-SMs in this study were performed using the ‘trainControl’ function, ‘method = repeatedcv’ of the ‘caret’ package (Kuhn, 2008) in the R programming language version 4.0.5 (R Core Team, 2021). This package streamlines the process of building predictive models. The process of all single-SM development and validation on fish richness and abundance is summarized in Figure 2.

We used three model performance measures, namely Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Coefficient of Determination (R²) to evaluate and compare the overall performance of all single-SMs. MAE tells the magnitude of an error we can expect from the forecast on average; whereas, RMSE shows the average distances between modeled and measured data (Moriasi et al., 2007; Guisan et al., 2013). R² shows how well predictors explain the variation in response variables. The R² values range from 0 to 1, where 1 represents perfect model fit. The optimal model is the one with the smallest value of MAE and RMSE and the highest value of R². In addition, the ‘varImp’ function was applied on the fitted single-SMs to assess the relative importance of each predictor for each single-SM. Further, sensitivity analyses of GLM, ANN and RF response variables to key predictors were visualized using Lek’s profile method (Lek et al., 1995, 1996) computed with the ‘lekprofile’ function in the ‘NeuralNetTools’ package. Lek’s profile method explores the relationship between the outcome variable and a predictor of interest, while holding other predictors at constant values.

Differences in performance measures among the models were tested using a non-parametric multiple comparison Kruskal Wallis rank sum test. To compare the difference in performance measures between models, pairwise Wilcox tests were used, with the pairewise.wilcox.test function of ‘stats’ package.

The ensemble prediction is expected to provide robust results and improve the predictive performances when compared to any of the single-SM predictions (Grenouillet et al., 2011; Oppel et al., 2012; Guo et al., 2014). The approach is to average the outcome of multiple models by taking the mean. For this study, we therefore used the predicted results of the six single-SMs for both richness and abundance separately as the input for ensemble predictions. To perform the ensemble prediction, (i) ‘caretList’ function of ‘caretEnsemble’ package was applied to build a list of train objects (for all single-SMs) that is necessary to pass to the ensemble prediction; (ii) ‘caretStack’ function was employed to combine the outputs from all single-SMs to create the ensemble model from the list of train objects in (i). The values of MAE, RMSE, R² were then extracted from the results of the ensemble prediction to assess models’ performances. All predictions and statistical analyses were performed using R language program version 4.0.5 (R Core Team, 2021).

Overall, the six single-SMs performed well in predicting observed fish species richness using geo-hydrological, physicochemical and climatic data in the LMB, with median R² values ranging between 0.63 for GLM and 0.85 for RF. The single-SM with the best performance was RF (Figure 3). The second-best performing single-SM was CART and the ranking was followed by SVM, ANN and KNN. GLM was the worst-performing model, with the lowest median R² value and highest values of MAE and RMSE. The ensemble model representing the average of the six single-SMs performed best at predicting richness, with an R² of 0.90, which was significantly better than other models.

Among the 14 predictive variables, water level (wl), distance from the sea (dist.sea), dissolved oxygen (DO), total nitrate (NO3), water temperature (Temp) and alkalinity (ALK) were the key predictors explaining the variation of fish species richness in the LMB (Figure 4). Indeed, all six-study single-SMs consistently indicated that these six key predictors were among the important variables to predict species richness in the LMB (Supplementary Information S1). Predictive variables that contributed the least in predicting species richness were ammonium (NH4) and chemical oxygen demand (COD) (Supplementary Information S1).

Sensitivity analyses showed the overall pattern of the contribution profile of the six key predictors using linear GLM, non-linear ANN and ensemble-tree RF (Figure 5). The linear model (GLM) showed species richness increased linearly with increasing water level, dissolved oxygen and temperature. However, species richness decreased with the other three predictors namely alkalinity, distance from the sea and total nitrate. Non-linear models (ANN and RF) indicated clear non-linear patterns for species richness in response to the increase in the predictor values (Figures 5B,C), with peaks that then declined in the response of fish species richness to increasing dissolved oxygen, temperature, and water level. This was most apparent with the RF model and suggests optimal or near optimal ranges of predictor variables for species richness. For instance, the RF model showed a sharp increase in species richness with increasing water levels; however, this pattern was flattened when there was a further increase in flow. Likewise, the RF model showed an initial peak in species richness as the distance from the sea increased; however, this pattern sharply declined with a longer distance from the sea. Overall, there was more variability in species richness from increasing water levels with non-linear models (ANN, RF).

Overall, all six single-SMs performed well in predicting fish abundance using geo-hydrological, physicochemical and climatic data in the LMB, with median R² values ranging between 0.58 for GLM and 0.77 for RF. The second-best performance single-SM for fish abundance was SVM (R² = 0.73), followed by KNN (R² = 0.72), ANN (R² = 0.69) and CART (R² = 0.61). Consistently, the single-SM with the best performance for fish abundance was RF while the poorest single-SM was GLM, with a significantly lower R² value and higher MAE and RMSE (Figure 6). The ensemble model performed best to predict fish abundance, with the highest median R² value of 0.85, which was significantly higher than other models.

Among the 14 predictors, the four key predictive variables that explained the variation of fish abundance in the LMB were distance from the sea (dist.sea), water level (wl), alkalinity (ALK) and water conductivity (Cond) (Figure 7). Other predictors, notably ammonium (NH4) and precipitation (preci), contributed the least to explaining the variation in fish abundance. Single-SMs revealed major differences in terms of variable importance when predicting fish abundance. CART, KNN and SVM consistently indicated that distance from the sea, water level, alkalinity and conductivity were among the most important variables to predict fish abundance in the LMB. Water level and distance from the sea were the two most important variables to predict fish abundance for all single-SMs, except for ANN, where total nitrate and water temperature were more important than water level (Supplementary Information S2).

Figure 8 shows changes to fish abundance from changes in the four key predictive variables for the GLM, ANN, and RF models. The linear model (GLM) showed that fish abundance increased linearly with increasing water conductivity. However, fish abundance decreased with increasing alkalinity, distance from the sea and water level. Non-linear models (ANN and RF) indicated that fish abundance in response to the increase in the conductivity values (Figures 8B,C), with peaks that then declined sharply before the pattern was flattened. This was most apparent with the RF model and suggests optimal or near optimal ranges of predictor variables for fish abundance. Contrary to conductivity, in GLM, linear decrease in fish abundance was found to be connected with increasing distance from the sea, increasing water level and increasing alkalinity. However, it was observed that, in non-linear models especially RF, the initial increase in fish abundance was indicated when there was an increase in water level. The pattern was then followed by a sharp drop in abundance when there was a further rise in water level.

Our study is the first contribution from the LMB to the current literature on species richness and abundance prediction using SMs. It showed that the selection of single-SMs and modeling methods matters in predicting fish species richness and abundance of multi-species assemblages in one of the world’s biodiversity hotspot regions. Overall, the six single-SMs and the ensemble (average) model in this study demonstrated a varied ability (with R² values between 0.56 and 0.90 of the total variability depending on the model used) to predict the species richness and abundance using geo-hydrological, physicochemical and climatic data. Consistently, RF performed best among all single-SMs in its predictive ability, with R² values equal to 0.85 and 0.77 for richness and abundance, respectively. More importantly, we confirmed that the ensemble model performed better (R² = 0.90 and 0.85 for richness and abundance, respectively) than all single-SMs.

In addition, our results indicated that the performance of the RF model was also very high and almost comparable to that of the ensemble model. This may not be surprising because RF incorporates the notion of ensemble modeling technique with thousands of trees being generated and predictions aggregated by averaging (Araújo and New, 2006). With this process, each decision tree is randomly built from the input dataset, and finally, all single trees are merged to get a more accurate and stable prediction (Breiman, 1996). Additionally, the best predictive performance of the ensemble model could be due to the low-pass filtering ability or cleaning effect of the average function (Marmion et al., 2009). Our results were in agreement with those of the previous SM work discerning that the performance of RF was very close to the performance of the ensemble model (Grenouillet et al., 2011; Guo et al., 2014). Overall, some SM models (e.g., GLM) did not perform well in our study. This is because each model has its own approach to predicting different characteristics of response variables, e.g., common and rare species (Guisan and Zimmermann, 2000; Sor et al., 2017b). For instance, Sor et al. (2017b) found that ANN and GLM performed better in predicting rare species than other SMs. Our results strengthened the recommendations, e.g., (Araújo and New, 2006; Marmion et al., 2009; Grenouillet et al., 2011; Oppel et al., 2012; Guo et al., 2014) that the application of the ensemble modeling technique generally yields a more certain, robust and reliable result than those of the single-SMs, and may therefore be best for guiding biodiversity management and conservation policy decisions. In this regard, it is not impossible to predict continuous responses (i.e., observed fish species richness and abundance) in the LMB, a large tropical freshwater system that is being impacted by regional and global changes. Our findings here contradict that of Oppel et al. (2012) who proved that it was extremely difficult to predict bird abundance in the coast of Portugal and Spain. This could be due to the strong adaptability of aquatic organisms to the Mekong River flood pulse environmental conditions. Changes in the system’s environmental conditions, e.g., seasonality and predictability of hydrology would likely cause changes in fish diversity and community structure (Ngor et al., 2018a,b,d). Our modeling results may therefore contribute to the ecological impact assessment and the management of aquatic ecosystem health, particularly fisheries health in the context of regional development (e.g., increasing hydropower dam) and climate change in the Mekong and other large tropical flood pulse freshwater systems. For example, statistical models have been used to conduct a risk assessment to support wetland restoration and conservation in an Indian wetland (Ghosh and Das, 2020), and to evaluate the flood hazard risk and management options in the Dongjiang River Basin, China (Wang et al., 2015). In the Mekong River Basin, statistical models could be used to evaluate the benefits of restoration, or the trade-offs between flood hazard risk management and fisheries health.

We found that water levels and distance from the sea are the most important predictors of fish species richness variability. Water levels are widely known to be the key ecological trigger to fish migration in the lower Mekong system (Ngor, 2000; Baran et al., 2001a,b; Baran, 2006; Ngor et al., 2018c). Increasing water levels signal fish migration for reproduction or dispersal. Mekong fish often leave permanent water bodies during the early wet season, moving to floodplains with improved environmental conditions (e.g., often with favorable conditions of dissolved oxygen, temperature and access to food and habitat) primarily for rearing and feeding (Matthews, 1998; Poulsen et al., 2002; Ngor et al., 2018a,e; Pin et al., 2020). We also found that fish species richness increased with increasing dissolved oxygen (min: 2.14, max:10.81 mgL⁻¹) and water temperature (min: 19.2, max: 33.0°C); see Table 1 for the values of predictors. Indeed, increasing water levels (i.e., from dry to wet season) are significantly linked to the increase in dissolved oxygen in waters (Soum et al., 2021), and in habitat or water surface connectivity within the LMB’s river network and its surrounding floodplains, and thus more favorable water conditions and habitats available for fish (Poulsen et al., 2002; Ngor, 2018; Ngor et al., 2018a). Such conditions give fish species a colonizing advantage, which drives greater species richness and diversity (Henriques-Silva et al., 2013; Laske et al., 2016). The pattern of increasing water surface coverage was also described as one of the most significant variables explaining the increase in species richness in Venezuela’s Cinaruco River (Rodríguez and Lewis, 1997; Hoeinghaus et al., 2003) and in the Brazilian Pantanal River (Fernandes et al., 2010). Also, the positive association between richness and temperature is likely explained by the difference in the ecological gradient of our study sites. Sites situated in the upper reach of the LMB (with higher elevation and lower species richness) generally have lower temperatures than sites in the lower floodplain where there is higher species richness. Moreover, our results were in line with those of recent studies indicating that there was significantly higher species richness in the wet season than in the dry season, particularly in the Mekong River mainstream and its tributary habitats (Pin et al., 2020). Fish species composition (beta diversity) of the lower Mekong River system was also observed to be significantly unique during increasing water levels in the early part of the wet season (Ngor et al., 2018b).

Further, decreasing richness was predicted to be connected with increasing distance from the sea. In other words, richness increased along the ecological (upstream-downstream) gradient toward the lower floodplain of the LMB. The result was consistent with broad-scale patterns of fish diversity in the LMB with greater species richness in the lower floodplain, i.e., the Mekong delta closest to the sea (Chea et al., 2016b). The increase in species richness was also explained by the ‘addition’ concept where the addition of fish species took place in the downstream floodplain. The addition concept has been widely observed in many river systems in both tropical and temperate zones alike, due to an increase in habitat diversity (species-area relationship theory) and improved environmental conditions for rearing and feeding (Oberdorff et al., 1993; Bistoni and Hued, 2002; Suvarnaraksha et al., 2012). Interestingly, in the non-linear models, increasing distance from the sea initially showed a sharp increase in species richness, and then followed by a drastic decline in richness with the increase in distance from the sea. This may indicate the important area of the middle system in the LMB in Cambodia particularly the Mekongg-3S system as a fish biodiversity hotspot. This result is in agreement with previous studies indicating the high fish diversity in Cambodia and the Mekong-3S (Chea et al., 2016b; Ngor et al., 2018b; Pin et al., 2020). Moreover, our results revealed that species richness decreased with increasing alkalinity and total nitrate. Increasing concentration of the two variables by anthropogenic activities might cause water pollution and thus likely reduced species richness. They were found to have a significant negative relationship with hydrological parameters including water level and precipitation, while positively correlated with evaporation and temperature (Prathumratana et al., 2008). This may indicate that higher concentration of alkalinity and total nitrate takes place in the low flow period. Low flows, i.e., during the dry season or droughts, indeed could cause detrimental effects on fish biota and, thus decrease species richness (and possibly abundance), as a consequence of increased water pollution with high alkalinity and decreased dissolved oxygen, and increased biochemical oxygen demand by microbes and decreased dissolved organic carbon (Dahm et al., 2003; Prathumratana et al., 2008; van Vliet and Zwolsman, 2008).

Similar to fish species richness, higher fish abundance was linked to decreasing distance from the sea. In the LMB, the major flood zone where large fisheries yields occur is predominately located in the lower floodplain of Cambodia (55% of the total yield) and the Mekong Delta of Vietnam (25%) while only 9 and 11% of the total yield is observed in the upper reach of the LMB in Laos and Thailand, respectively (Hortle and Bamrungrach, 2015). Higher fish abundance in the lower floodplain closer to the sea may be due to availability of better and more habitat and food sources for fish, and especially young fish, to rear, feed and grow during high flow periods. This result further reinforces the importance of maintaining environmental health and ecological connectivity in the lower floodplain (Arias et al., 2019) in support of sustainable fishery yields that are necessary to feed millions in the region.

Opposite to species richness, we discovered that fish abundance increased with decreasing water levels. This can be due to fish seasonal migrations that are fine-tuned with seasonal hydrology of the Mekong. In the lower Mekong system, increasing water level facilitates fish dispersal from upstream to downstream inundated areas in the lower floodplains with increased habitat diversity during the wet season for fish rearing and feeding. During the period, fish are widely dispersed with seasonal floods over huge flooded plains in the basin. This may drive lower fish abundance as lower fish density is expected during the high flow which makes it difficult for fishers to capture them in large quantities. When water level recedes during the dry season, fish that have been reared and fed in the flooded plains migrate back to the deeper area of water bodies or main river channels such as lakes and deep pools for dry season refuge. Many fishes in large quantities are vulnerable to being captured during their dry season migrations and during their sedentary period in the dry season refuge since they tend to concentrate in high density in those habitats. This drives greater abundance of fish during lower flows. In fact, intensive fishing activities take place during the periods of decreasing flow in many parts of the LMB both for subsistence and commercial-scale fisheries, e.g., the century-old, commercial-scale stationary trawl bagnet fishery in the Tonle Sap River. The fishing activities particularly happen when the water levels drop to about the average flow level between the highest and lowest flow in December and January at the Tonle Sap River (Ngor, 2000; Halls et al., 2013; Sabo et al., 2017; Ngor et al., 2018c).

Interestingly, we predicted fish abundance would increase with increasing conductivity in water. This is likely reasonable because increased conductivity is generally accompanied by decreasing flow and higher temperature, and during the period, there is more concentration of dissolved substances, chemicals and minerals in water (Prathumratana et al., 2008; Izaguirre et al., 2018; Sensorex, 2021). Therefore, in the LMB, high water conductivity tends to occur during decreasing water levels concomitant with high fish abundance being generally harvested particularly during the first three to 4 months of the dry season (October–January) (Halls et al., 2013; Ngor et al., 2018a,c). This causal relationship may explain our result of increased fish abundance being linked to increased conductivity in water. However, toward the end of the dry season (March–April) which is among the driest period of the year, significantly high conductivity and alkalinity could occur in tandem due to high minerals and chemical pollutants entering the water, e.g., from agriculture, industries and urban sewages drained directly to rivers (Prathumratana et al., 2008; Sensorex, 2021). High conductivity, e.g., greater than 50 mSm⁻¹ could be harmful to stream fish (Zhang et al., 2019). In the LMB, maximum conductivity was recorded at 61.10 mSm⁻¹ in this study.

Finally, our study found that high fish abundance was related with decreased alkalinity. In polluted rivers, high alkalinity was associated with low oxygenation conditions, e.g., (Abril and Frankignoulle, 2001). In the LMB, high alkalinity was found to be connected with low precipitation and flow, but high evaporation and temperature (Prathumratana et al., 2008). This indicates that high alkalinity predominantly exists in the warmest months of the year (March–April), and could cause hypoxic conditions that could be harmful to aquatic biota. This may drive the inverse relationship between fish abundance and alkalinity in the LMB. Such relationship was also indicated for some stream fishes in eastern Thailand (Tongnunui and Beamish, 2009).