Keywords

1 Introduction

In China, the ecological states of waterways are receiving increasingly prevalent attention. Especially in recent years, emphasis has been placed on incorporating the concept of ecological development into waterway training projects in the Yangtze River and the Beijing-Hangzhou Canal. During the waterway regulation of the Jingjiang Reach in the middle Yangtze River, the concept of ecological and environmental development is a key component of the project. Waterway health changes in the Jingjiang Reach were assessed by Li et al. (2018) in the context of a waterway regulation project. The results highlight potential mutual improvement between ecosystem health and navigation functions. In the Huai’an section of the Beijing-Hangzhou Canal in Jiangsu Province, the development of ecological waterways has been not just the protection of ecology, but also the integration of cultural heritage into the canal (Wang 2008). Based on the natural and socioeconomic attributes, Duan et al. (2021) constructed a theoretical framework and integrated index system to evaluate the ecological sensitivity, development potential, and current development patterns of riverfront resources. At present, a large number of projects focusing on the ecological development of waterways have been carried out, but much less research has been conducted related to the evaluation of the ecological situations of waterways.

Existing studies related to the ecological evaluation of waterways can be divided into two categories (Cai and Hu 2008): the first group focuses on specific ecological aspects and most of them relate to environmental factors such as pollutants, algae, invertebrates, and fishes. Methods in this group include the fish biotic integrity index proposed by Karr (1981) and the Riparian, Channel, and Environmental (RCE) Inventory developed by Prtersen (1992). The second group tries to carry out comprehensive evaluations. Li et al. (2018) took the Jingjiang section of the Yangtze River as an example, established an evaluation system, and used the hierarchical analysis method to evaluate the ecological situations of the Jingjiang reach in terms of the river navigation function, water quality, ecology, sand transport, and landscape recreation function.

In the last decade, neural network models have been widely used in various evaluation models (Chu et al. 2013; Hamida et al. 2020), but it has rarely been applied to evaluate ecological waterways. In addition, the quantitative analysis of the relative weights of each evaluation index is seldom reported. Therefore, the purpose of this study is to build ecological waterway evaluation indexes, evaluate ecological waterways by BP neural network model and PSO-BP neural network model, and then conduct a quantitative analysis of the relative weights of the indexes based on the ANN models.

2 Methodology

2.1 Data Collection

In China, eighty percent of waterways are distributed in five rivers or canals in terms of length of waterways: the Beijing-Hangzhou Canal, the Yangtze River, the Huai River, the Pearl River, and the Heilongjiang River. In addition, 97% of the throughput of China’s waterways is contributed by the Yangtze River system and some southern waterways. Therefore, the dataset in this study was mainly collected from the four major waterway systems, namely, Yangtze River, Pearl River, Huai River, and Beijing-Hangzhou Grand Canal. Specifically, collected data includes the number of vessels, maximum vessel tonnage, water system connectivity, bank slope, the total number of vessel accidents, river width, coverage of ecological berms, bank vegetation coverage, bank plant richness, functional area water quality, riparian buffer zone width index, bank park area, and GDP per capita. Normalization is applied to all the collected data using the maximum-minimum method to convert all data to the range of (0, 1). Then the expert scoring method was used to determine the ecological rates of each waterway sample. A total of 30 experts provided their evaluation results and the final ecological rates are obtained by averaging these evaluation results. Finally, a dataset including 84 waterway samples was built for the following model establishment. This dataset can be found in Guo (2021) in which the measurement of each parameter was explained in detail.

2.2 Ecological Waterway Evaluation Model Based on Traditional BP Neural Network

The BP-ANN model for the evaluation of ecological waterways is established by MATLAB which contains a toolbox of neural networks. With the help of this toolbox, a BP-ANN model can be readily built by regulating relevant parameters according to specific requirements. Building the BP-ANN model includes the following steps: the construction of modeling structure, modeling training, and prediction. There are 13 input parameters and one output parameter for current BP-ANN model. The output parameter is the evaluation levels of waterways, including excellent, good, medium, poor, and bad, totally five grades. After testing different numbers of hidden units in the intermediate layer of the BP-ANN model, results show that the highest prediction accuracy obtained when the intermediate layer has 11 hidden units. To avoid overfitting, the dataset is divided into a training dataset (approximately 80% of the data) and a test dataset (approximately 20% of the data). The former is used to train the ANN model, and the latter is used to evaluate the predictive accuracy of the model.

2.3 Ecological Waterway Evaluation Model Based on PSO-BP Neural Network

The BP-ANN model randomly assigns connecting weights and thresholds during the model training process, which may fall into local optimal solutions and thus result in different outcomes every time. To address this issue, this study uses the PSO algorithm to improve the BP-ANN model which combines the particle swarm algorithm with the BP-ANN model. The specific steps of building PSO-BP-ANN model include: 1) use the particle swarm algorithm to obtain the position and velocity values of optimal particles; 2) assign the obtained values to the weights and thresholds in the BP-ANN model, respectively; 3) train the BP-ANN model based on the dataset and evaluate the neural network fitting effect by the decision coefficient R2. This method avoids the random assignment of weights and thresholds in BP-ANN model, leading to a greater likelihood of obtaining a globally optimal solution.

3 Results

3.1 Performance of the BP-ANN Model and the PSO-BP-ANN Model

Three performance indexes of the BP-ANN model and the PSO-BP-ANN model are compared in Table 1. In this table, the R2 of the traditional BP-ANN model is 0.62, while that of the PSO-BP-ANN model reaches 0.97, which indicates that the improved PSO-BP-ANN model has a higher prediction accuracy. However, the improved accuracy is accompanied by an increase in training time which reaches 25.29 s in comparison to the 1.3 s of BP-ANN model. This is attributed to the continuous iteration process in the PSO. This improvement in accuracy at the expense of time is a common feature of advanced algorithms. The number of iterations in the PSO-BP-ANN model is 8, while it is 15 in the conventional BP-ANN model indicating a significant reduction of the number of iterations of the former model. The PSO produces better estimates of weights and thresholds, which enables the neural network to reach the optimal solution quickly.

Table 1. The performance of the two ANN models
Full size table

3.2 Influence of the Variation of Parameters on Grading Levels

The relative importance of the evaluation parameters is carried out with the help of the PSO-BP-ANN model. In order to simplify the analysis, the evaluation indexes were categorized into four groups: transportation function (including the number of vessels, maximum vessel tonnage, water system connectivity, and the total number of vessel accidents), ecological function (including bank slope, river width, the coverage of ecological berms, bank vegetation coverage, and bank plant richness, and functional area water quality), landscape function (including riparian buffer zone width index and bank park area), and economic function (GDP per capita), and analyzes the influence of these four functions on the evaluation results accordingly. The scores of the functions are the average of the corresponding parameters. The specific steps include: 1) fix one function with a certain value, 2) randomly change the values of the other three functions, 3) calculate the probability of the five grades of the ecological evaluation, 4) change the value of the fixed function and repeat the previous steps to obtain the probability distribution of the five grades under different values of this function.

Fig. 1.
figure 1

The influence of transportation function on evaluating grades

Full size image

As shown in Fig. 1, the abscissa indicates the score of the transportation function, and larger values indicate better conditions of navigation conditions. The vertical coordinate indicates the probability of the five grades. With the increase of the transportation function score (TFS), the probability of the excellent level increases first and then decreases when it is greater than 0.9. The overall probability of the excellent level is less than 0.1, which means the transportation function has limited influence on this grading level. When TFS is less than 0.3, the probability of the excellent level is 0. The probability of the good level varies in a limited range with the changing TFS, in an interval of 0 to 0.1. When the transportation function score is less than 0.1, the probability of the good level is 0. The probability of the poor level varies in a limited range as well, e.g., from 0.1 to 0.2. The possibility of the medium grade increases steadily from 0.1 to 0.5 with TFS increasing from 0 to 1. The bad grade changes most significantly compared with the other grades, and it decreases constantly for TFS increasing from 0 to 0.6 and remains stable thereafter. In summary, the enhancement of TFS can significantly reduce the probability of the bad level and increases the probability of the medium grade significantly. When the TFS score is higher than 0.6, the probability of all grading levels tends to be stable.

Fig. 2.
figure 2

The influence of the ecological function on the evaluating grades

Full size image

As shown in Fig. 2, the abscissa indicates the ecological function score (EFS). The larger the EFS, the better the ecological conditions of waterways. With the increase of EFS, the probability of excellent and good grades shows an increasing trend. The probability of poor and bad grades shows a decreasing trend with increasing the EFS, while the probability of medium waterways shows a non-monotonical trend that it first increases and then decreases with the increasing EFS. The probability of a good grade abruptly increases from 0 to 0.5 when the EFS increases from 0.35 to 0.5. When the EFS exceeds 0.5, the probability of the excellent grade starts to increase from 0 and stabilized at 0.06 after the EFS is greater than 0.65. In summary, increasing the EFS can significantly increase the probability of the good grade given the EFS being greater than 0.35. Unlike the TFS, the EFS does not exhibit a monotonic effect on the middle, poor, and bad ratings, but they fluctuate with the EFS to some extent. After the EFS reaches 0.65, the probability of all grades tends to be stable.

Fig. 3.
figure 3

The influence of landscape function on the evaluating grades

Full size image

Figure 3 shows the different grading levels as a function of the landscape function score (LFS). In this figure, the probability of the excellent grade is zero until the LFS reaches 0.6 and then increases steadily to 0.2 with a further increase in the LFS. Similarly, the probability of the good grade is zero until the LFS reaches 0.4 and gradually increases with the LFS thereafter. The probability of the good grade stabilizes when the LFS reaches 0.8. The most significant change belongs to the probability of the bad grade that with the LFS increasing from 0 to 1, the poor grade constantly decreases from 0.9 to 0.1. In contrast, the probability of the poor level is almost invariant with the LFS. In summary, with the increase of the LFS, the probability of excellent, good, and moderate grades gradually increase, and the probability of the bad grade significantly decreases.

Fig. 4.
figure 4

The influence of the economic function on the evaluating grades

Full size image

As shown in Fig. 4, the probability of the occurrence of the bad grade decreases constantly as the economic function score increases, i.e., it decreases from 0.6 to 0.25. The probability of the occurrence of the moderate grade increases substantially after the economic function score above 0.5, increasing from 0.28 to 0.48. The probability of occurrence of the excellent grade is 0 when the economic function score is below 0.4 and remains constant at 0.05 when the value lies between 0.6 and 0.9. It increases slightly after the score reaches 0.9. The probability of occurrence of poor and good levels remains relatively stable with the increase of the economic function score. In summary, the increase in the economic function score can greatly reduce the probability of the poor grade and increase the probability of the moderate grade, with no significant effect on the probability of remaining grades.

4 Conclusions

This study demonstrated the better performance of the PSO-BP-ANN model on the evaluation of the ecological waterway than the traditional BP-ANN model. Based on the former, the influence of four different function indicators including the transportation function, the ecological function, the landscape function, and the economic function on evaluating grades were analyzed. Among the four indicators, the landscape and transportation functions have the greatest influence on the probability of the bad level, and the increase of these two functions can significantly reduce the probability of this rating. The four indicators have an insignificant influence on the probability of the poor level. For the probability of the moderate level, the landscape function has the greatest impact. The ecological and landscape functions have a consequential influence on the probability of the good level. For the probability of the excellent level, all four indicators need to be greater than a certain value to reach this level.