Constructing Benthic Integrity Indices Based On Classification And Regression Trees And Random Forest Prediction Models

Freshwater ecological health has received widespread attention,and the multimetric index（MMI）for benthic fauna is an important evaluation method for water body health.However,the survival status of benthic fauna is affected by the natural environment,and especially at the watershed scale it is difficult to distinguish the information of natural variables from anthropogenic disturbance effects in biological parameters,thus unable to reflect the actual health level of lakes.Currently,predictive model-based methods can effectively control the natural variables of parameters in the process of river MMI construction,but they are less applied in lake basins.In this study,the water quality of Baiyangdian Lake in Hebei Province has been investigated by analyzing the relevant physicochemical parameters and benthic community structure characteristics.Two types of decision tree models,Classification and Regression Tree（CART）and Random Forest（RF）,have been used to construct the MMI of benthic animals for exploring the effects of predictive models on the performance of benthic animals’MMI by natural variables controlled.The effects of the two predictive models were compared on the performance of benthic animals’MMI.The key environmental factors affecting benthic animals’MMI were explored by correlation analysis between environmental factors and benthic animals’MMIs.The water quality monitoring results show that the comprehensive trophic state index（TLI）of the whole lake range from 22.08 to 35.92,68.67%of which has a poor trophic level,31.33%has a moderate trophic level while l eutrophication not having been occurred.Dissolved oxygen（DO）,total phosphorus（TP）and transparency（SD）in the water column are the key factors affecting the abundance of benthic communities.The abundance of seven benthic families is significantly correlated with DO,meanwhile 6 families with TP and 4 families with SD.Pielou evenness index is significantly and positively correlated with PO₄^3-and TP but negatively correlated with SD and WT.Margalef abundance index is significantly and positively correlated with SD.Simpson diversity index is significantly and positively correlated with TP.N-MMI（Normal-MMI）is used as the MMI constructed from the original values of biological parameters by constructing MMI using benthos.C-MMI（CART-MMI）is used as the MMI constructed from the parameters after controlling for differences in natural variables using the CART model.R-MMI（RF-MMI）is used as the MMI constructed from the parameters after controlling for differences in natural variables using the RF model.All three MMIs have good performance in evaluating the lake water body.The MMI constructed by the two prediction models has better performance in all aspects compared to the MMI constructed by the original values.In terms of precision,C-MMI has the smallest standard deviation of 7.99 at the reference point,which has the highest precision.R-MMI has the second highest standard deviation of 12.12 at the reference point,which also has high precision.N-MMI has the largest standard deviation of 17.95 at the reference point,which has the lowest precision.In terms of deviation,R-MMI has the smallest deviation of 0.C-MMI has the second highest deviation,and the variation in natural gradient explains 11.95%of the variation in C-MMI.The variation in N-MMI natural gradient has a higher degree of interpretation in MMI with a deviation of 13.25%.In terms of sensitivity,C-MMI and R-MMI have the same percentage of damaged points in non-reference state points,both of which are63.03%on higher sensitivity.N-MMI has the percentage of damaged points in non-reference state points,which is 56.30%with lower sensitivity.In terms of responsiveness,the absolute t-value of C-MMI is 7.47,which has the highest responsiveness between the reference and damaged points.The absolute t-value of R-MMI is 6.85 which has the second highest responsiveness between the reference and damaged points.The absolute t-value of N-MMI is 6.41 which has the lowest responsiveness between the reference and damaged points.In terms of the discriminative ability of human interference,C-MMI is relatively strong in discriminating human interference with an absolute t-value of 2.552.R-MMI was the second strongest in discriminating human interference with an absolute t-value of 2.537.N-MMI was relatively weak in discriminating human interference with an absolute t-value of 2.508.The correlations between the three MMIs and environmental factors are in high agreement and significantly positively correlated with both TP and PO₄^3-which having showed phosphorus being the key element affecting the benthic integrity index.In the fitted curves of the LOWESS function,there are turning points in the LOWESS curves of N-MMI and C-MMI with soluble phosphate PO₄^3-.But there is no significant difference in the MMI values turning.The remaining four LOWESS curves have smooth trends,and the conservation thresholds for TP and PO₄^3-could not be calculated.In summary,the method of using predictive models to control the natural variables of parameters can improve the performance of MMI in the lake basin,among which the CART model has the best performance in terms of precision,responsiveness and discriminating anthropogenic interference,and the RF model has better performance in eliminating bias.Meanwhile,phosphorus is a key element affecting the survival status of benthic fauna in Baiyangdian.