Development of a lightweight convolutional neural network-based visual model for sediment concentration prediction by incorporating the IoT concept

Flood damage is regarded as the most frequent natural disaster. Besides, the vast sedimentation is often carried by the enormous flooding due to broken watersheds and causes debris flow, reservoir siltation, river erosion and deposition, and water quality deterioration. This compound disaster frequently occurs in the current period because climate change contributes to dramatic flooding and sedimentation issues. Many studies have been conducted on sediment transportation during flood events. Still, several disturbing factors in field sites cause incomplete data collection, such as adverse weather conditions and the limitations of the measuring instrument. Despite developing early warning systems for decades, the accuracy has yet to be effectively improved due to insufficient sediment data. Above all, a fast, accurate, and reliable measurement technology needs to be proposed to reduce the uncertainty of the collected data and further supply sufficient data to improve the sedimentation prediction ability.

Traditional technology in sediment concentration (SC) measurement in the field presents challenges because the data representation often occurs during flooding. It is challenging to collect, and the above paragraph shows the reason. The measuring approaches for SC are divided into direct sampling and indirect measuring. Suction sampling is a commonly used direct method that is inexpensive and simple (Nielsen 1984; Bosman et al. 1987). The accuracy of the suction sampling method is high; however, it is time-consuming because the water samples need to be analyzed in the laboratory, thus being hard to apply in the field site for large-scale, real-time monitoring during flood periods. Indirect measuring methods have been developed in the past decades to address these limitations (Wren et al. 2000; Barua et al. 2001). The acoustic method for measuring SC is a contactless technique that uses the decay rate of sound frequency to estimate the concentration of suspended particles. However, the accuracy is limited by the need for prior calibration and a strong relationship between the reflection signal and concentration. In other words, it requires extensive laboratory work, and the established relationship may be difficult to transfer to other locations (Thorne et al. 1991; Lohrmann 2001; Landers et al. 2016; Sirabahenda et al. 2019; Lin et al. 2020). Laser diffraction is a technique for measuring SC that uses light. The laser emits a beam into the sample, scattered, reflected, and absorbed by suspended particles. The method determines the concentration by analyzing the volume of the particle. However, laser diffraction needs to be calibrated and limited to a single-point measurement (Agrawal & Pottsmith 1994; Guillén et al. 2000; Merten et al. 2014; Su et al. 2016; Coleman et al. 2020). Time domain reflectometry (TDR) is another technique for measuring SC. It uses the relationship of different dielectrics and sediment samples to estimate the concentration. TDR has been applied to sediment observations in vertical profiles, but the limitation of TDR is a significant error in low-concentration environments (Chung et al. 2013; Mishra et al. 2018; Miyata et al. 2020; Chung & Wang 2022). Except for the above measuring technique, the numerical model is considered convenient to solve the SC issue, and the improvement of the SC estimation was gradually obtained. However, the accurate concentration value is challenging to simulate in the numerical model due to parameter settings in the different field sites. Additionally, the mentioned approaches are difficult for real-time monitoring (Fang & Wang 2000; van Maren et al. 2015; Huang et al. 2019; Orseau et al. 2021). Note that a proper measuring technology needs to be developed, and the interdisciplinary artificial intelligence (AI) is an appropriate way to optimize the current water resources management method.

In hydraulic engineering, AI has been applied to estimate the time-space distribution, surface runoff, groundwater level, and water quality (Khan et al. 2019; Hasda et al. 2020; Ighalo et al. 2021; Iriany et al. 2022). Moreover, several studies mentioned the monthly, weekly, or daily SC value estimation by using different ML algorithms (AlDahoul et al. 2021, 2022; Ehteram et al. 2021; Fan et al. 2023; Latif et al. 2023). The time scale is the other important issue, and several papers applied to predict the short lead-time SC value and obtained reasonable results (Chang et al. 2020; Huang et al. 2021; Jhong et al. 2021). Computer vision is a field of AI that enables a computer to derive meaningful information from digital images, videos, and other visual inputs that may not be found with the naked eye. For example, convolutional neural network (CNN) convolves the data into a smaller area to extract features (Hwang 2017; Hwang & Chen 2017), e.g., colors, edges, textures, and so on. Via the combinations of these features, we can detect, recognize, or predict some further events. For instance, Chang et al. (2021) propose an architecture based on CNN to detect vehicles and predict the future position of each vehicle for collision avoidance. Next, we can train a machine to perform these ML functions in much less time than the procedure of retinas, optic nerves, and visual cortices. For example, CNN is also one of the fastest and most accurate algorithms for image applications, which can inspect a large number of images per minute (Hwang 2017; Hwang & Chen 2017; Chang et al. 2022). Obviously, it surpasses human capabilities. On the basis of the above discussions, we can realize that CNN suits the estimation of SC by using different images. More specifically, the variation of the SC needs to observe the details of colors and textures of water and inspect the sample fast, accurately, and stably, and both are the capabilities of CNN. Above all, the above references indicate that accurate SC estimation is a crucial issue, and the current technology is still under the improvement process. The ultimate goal should establish a robust SC measuring approach to apply to the real-time warning in the field site.

Traditional technology in SC measurement in the field presents challenges because data representation often occurs during heavy rainfall or typhoon events. It is difficult to collect and hard to be applied for large-scale and real-time monitoring during flood periods. Therefore, we designed a novel method based on CNN to properly deal with the SC issue estimation.

A neural network is an ML model which is inspired by the architectures of neurons. Owing to the properties of NN, it can be collocated with other techniques to cope with large and complicated issues, e.g., classifying images, speech recognition, optimization, and so on those references (Pedregosa et al. 2011; Chollet 2015; Chang et al. 2022; Géron 2022). Namely, it is versatile, scalable, and powerful.

The idea of convolution is inspired by the organization of the animal visual cortex and works well (Géron 2022). Briefly, individual neurons respond to the stimuli in a receptive field. Since the receptive fields of different neurons partially overlap, they finally tile the entire visual field. This can be approximated by the operation of convolution (Hwang 2017; Hwang & Chen 2017). On the other hand, the convolutional layer is generally paired with the pooling layer to reduce the dimensions by combining the outputs of neuron clusters in the previous layer into one single neuron in the next layer. Namely, the design can reduce the quantity of weights and complexity of calculations compared with traditional neural layers in the network. Hence, we can add some convolutional layer(s) as well as the pooling layer(s) in the first half of the neural network to introduce its advantages to the traditional neural network.

Technically speaking, a convolutional neural network is an architecture composed of at least some convolutional layer(s) as well as pooling layer(s) along with a neural network, and the two kinds of layers often pair in practice. Thus, we can use several simple equations to describe the computational cost via the change of neurons in successive layers. Assume that an image with n*n pixels is considered here. If w filters with m*m pixels and the pooling region with k*k are used in the convolutional layer and pooling layer, we need w*[(n − m + 1)/k]*[(n − m + 1)/k] neurons in the successive layer to receive such output.

In this section, we introduce three well-known classifiers that are adopted to compare with ours for the purpose of showing the superiority of our method, i.e., SVM (support vector machine), VGG19 (Visual Geometry Group 19), and ResNet50 (Residual Network 50).

Firstly, SVM is a classical ML model capable of performing classification for small- or medium-sized datasets (Géron 2022). It constructs a hyper-plane or set of hyper-planes in a high dimensional space, which can be utilized for classification. Generally, when the problem is not separable, the support vectors are the samples within the margin boundaries. Here is the brief mathematical formulation: given training samples x_i and a vector y, the goal is to find w and b such that the prediction given by is correct for most samples. Furthermore, if the number of classes is greater than 2, the strategy of OVR (One-Vs-Rest) is utilized to help us reach the goal (https://scikit-learn.org/stable/modules/svm.html#svc).

The next one is VGG19 from the VGG research laboratory at Oxford University, which also has a simple architecture (Simonyan & Zisserman 2015; Géron 2022). Briefly, there are a few convolutional layers with a pooling layer repeatedly until reaching a total of 16 or 19 convolutional layers that depend on different variants. The last one is ResNet50, which is based on the residual network architecture (He et al. 2015; Géron 2022). Noteworthily, like other variants with 101 and 152 layers, the ResNet architectures use an extremely deep design. Note that since the inputs and outputs of our work are different from the two originals, we have to modify them to fit our requirements. Next, for these two originals, using only our sediment samples is not enough to train them and causes very terrible performance. Hence, we incorporate the concept of transfer learning based on the concept of pre-training weight with ImageNet. We freeze several upper convolutional layers to keep the characters of the originals and only retrain the successive layers to fit our requirements such as 16 classes.

As a result, we choose a classical method, a shallower convolution neural network, and a deeper convolution neural network to compare with ours and give comprehensive discussions.

In this work, FUJIFILM X-S10 is used to take all experimental photos, which is an APS-C (Advanced Photo System Classic) camera. The settings are shown below:

• •

  Resolution: 6,240*4,160;

• •

  Aperture: f/8;

• •

  Shutter speed: 1/25 s;

• •

  ISO: 200;

• •

  Focal length: 35 mm;

• •

  Color temperature: 4,000;

• •

  Mode: manual mode.

On the other hand, the experimental scenario is illustrated in Figure 2. The camera is set horizontally to the sample in the film studio, where the distance is 100 cm from the former to the latter, which is shown in Figure 2(a). Then two fill lights are set horizontally on both sides of the film studio to reduce the contrast to match the dynamic range of the photo. Namely, it is for the purpose of recording the same number of details seen by humans (Figure 2(b)).

Next, the detailed settings of the learning procedure are given below:

• •

  Batch size: 128;

• •

  Dropout: 0.2;

• •

  Learning rate: 0.001;

• •

  Optimizer: Adam;

• •

  Beta_1: 0.9;

• •

  Beta_2: 0.999;

In this study, grouping was conducted with intervals of 1,000 ppm. The main reason for this choice was to ensure that each subgroup contained at least five or more data points for analysis. The comparison spacing is divided into eight parts, from 0 to 1,000, 1,000 to 2,000, 2,000 to 3,000, 3,000 to 4,000, 4,000 to 5,000, 5,000 to 6,000, 6,000 to 7,000, and 7,000 to 8,000 ppm. The values of MAE are 269.17, 241.19, 384.38, 296.37, 219.37, 281.91, 217.56, and 199.44, respectively. The maximum and minimum values are 384.38 and 199.44 at 2,000–3,000 and 7,000–8,000 intervals. The same trend is presented in RMSE; the maximum and minimum values are 389.56 and 242.59. Additionally, CC and NSE are adopted to evaluate the correlation and error. The range of CC is between 0.88 and 0.99; the range of NSE is from 0.72 to 0.81. Both CC and NSE show satisfactory results. All statistical data are shown in Table 1.

This study further investigates the meaning of statistical index to the prediction performance. The highest difference of MAE is at the interval 2,000–3,000 ppm, a value of 384.83. It is a sufficient accuracy in the field site application because the accuracy of the current instrument is 500–1,000 ppm. The maximum error of RMSE is in the interval 2,000–3,000 ppm, a value of 389.56. In addition, the minimum error value is 242.59, located at the interval 7,000–8,000 ppm. A Significant result is presented that the accuracy and the stability are acceptable because the value of MAE and the difference in different intervals of RMSE is small enough. CC is applied to assess the degree of correlation between the measurement and prediction. In addition, the ranges in value from −1 to 1. A CC closer to 1 or −1 indicates a stronger positive or negative linear correlation between the two variables, while a CC close to 0 indicates no linear correlation between the two variables. In this study, the range of CC is from 0.86 to 0.95 in different intervals, which means the measured and predicted values present a high correlation. NSE is commonly applied in the hydraulic engineering field to analyze the fitting of a prediction model to measured data. It is commonly applied in fields such as hydrology and water resources management. NSE ranges from negative infinity to 1, with a value closer to 1 indicating a better fitting. Conversely, a negative value signifies poor model performance. The range of NSE shows a good fitting in the proposed prediction model because all values are higher than 0.7, and the best performance is located at the interval 7,000–8,000 ppm with a value of 0.90.

Overall, three significant advantages are shown in the proposed model. The satisfied accuracy compared to the current instrument; the sufficient stability to apply in the field site; and last but not least, the high correlation of measurement and prediction is presented in all intervals, which means this model can be applied in 0–8,000 ppm. A practical application procedure could be formulated in the future.

In this subsection, via the same dataset, three well-known classifiers are adopted to compare with ours, i.e., SVM, VGG19, and ResNet50. The detailed results are given below. Firstly, Table 2 shows the number of parameters of each network. Since the number of parameters affects the time of the predicted procedure, we can observe that the methods need a similar ratio of time to finish their predicted procedure. This conforms with the characteristics of neural networks. Note that the specifications of the experimental computer are CPU: Intel Core i7-97,003.00GHz, RAM: 16 GB DDR43, 200MHz, and GPU: Nvidia GeForce RTX 2080 8 GB. Next, according to the results in Table 2, we can see that our method demonstrates its superiority in predicted accuracy as well.

This study firstly considers the SC issue as a classification topic for 16 classes with the lightweight CNN. Our design and experiments illustrate that the CNN-based method for SC prediction during flood periods is feasible and works well. The proposed CNN model can potentially develop a practical tool in the field site due to the following advantages.

Firstly, the proposed CNN-based model should be compared with the well-known classification algorithms instead of regression methods. This is why we chose SVM, VGG19, and ResNet50. This prediction model presents satisfied accuracy by comparing it with the current technology. The proposed CNN model performs better than the other well-known models, SVM, VGG19, and ResNet50. In addition, sufficient accuracy is presented by comparing it to the current technology (lower than 500ppm). Second, this model overcomes outlier trouble often occurring in the prediction process and shows sufficient stability. Third, it can be a simple network architecture and training procedure. Since our network only has a small number of trainable parameters, it is easier to be ported to embedded systems in practice. Most importantly, via our design, such a traditional but critical issue can be improved to a new state. This study proposes a practical concept by incorporating the Internet of Things (IoT) with our design, a contactless technology, the distributed computing system for large-scale environmental monitoring can be realized quickly and easily.

This study is worth continuing, such as challenging the higher concentration prediction, more accurate performance, and different particle sizes/shapes of the sediment. Furthermore, developing a real-time response system for the prediction of the turbidity water to improve the water resources management potential. The above vision will be presented in the future study.

The financial support for this study was provided by a grant from the National Science and Technology Council (project number: 111-2625-M-035-007-MY3).

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.