Tree-based machine learning algorithms in the Internet of Things environment for multivariate flood status prediction

2.1 Flood data acquisition

The IoT technology has been recently implemented in a wide range of projects as a smart solution for data collection and processing, especially in dynamic and complex systems. Ghapar et al. [9] propose the use of IoT architecture for flood data management. They suggest the usage of different sensors for collecting flood-related data, such as sensors for measuring hydrological, geological, and meteorological data. The project confirms the ability of the IoT architecture to facilitate data collection, transmission, and management. However, this architecture remains conceptual and is never implemented.

Noymanee et al. [21] propose a conceptual framework based on an IoT platform for flood early-warning systems in an urban environment. A context-aware module supports the framework to characterize the flood conditions. The IoT architecture for the context-aware system consists of five layers. These layers are application, storage management, processing or reasoning, raw data retrieval, and sensors. They intend to make the framework to predict the flood situations from observations without using an explicit prediction method. However, this framework also remains conceptual and is never implemented. Similarly, Chen and Chen [22] propose the use of a context-oriented IoT platform for data acquisition and integration. The integration process entails converting the raw data to a semantic context for easy storage, understanding, and sharing.

Balakrishna et al. [23] propose a sensor data acquisition and analysis framework based on an IoT platform. The framework is implemented in a traffic monitoring system by using the ThingSpeak IoT Cloud platform that provides data analysis and visualization services. A median filter is used to improve the context of the data. The data processing is performed by the Gaussian mixture model and context construction method. The test results show that the framework achieves an accuracy of 84.56% on average for traffic condition description.

Fang et al. [24] propose an IoT-based integrated information system for snowmelt flood early warning. The IoT architecture is used for data acquisition, sharing, and management of multi-source information. The integrated information system has been developed as a web application. The test case study is the Quergou River Basin, which is located in Hutubi County, Xinjiang, China. The resulted warning recommendations are compared between the actual and presented water levels.

2.2 Flood status prediction

Various research articles investigate the use of ML algorithms in the prediction of river flooding situations, and some of them are presented in this section. Chen et al. [2] use three techniques for spatial forecasting of floods in the Quannan region of China. They conduct an evaluation and comparison of these techniques, the Naive Bayes Tree (NBTree), the Alternative Decision Tree (ADTree), and RF methods, to determine their ability to predict floods. Their approach is based on producing a flood inventory with 363 flood sites and dividing them into training and verification datasets with a 70/30 random selection. Thirteen factors are used to create the spatial flood database to explain and understand the floods. Their results show that RF is an effective and reliable model for assessing flood vulnerability.

Hong et al. [6] examine four DT-based ML models, namely Logistic Model Trees (LMT), Reduced Error Pruning Trees (REPT), NBT, and Alternating Decision Trees (ADT) for flash flood susceptibility mapping in Iran. They construct a spatial database with 201 present and past flood locations and 11 flood-influencing factors. The capability of these models for flood predictions is evaluated and compared using statistical evaluation measures, the receiver operating characteristic curve, and Freidman and Wilcoxon signed-rank tests. The results show that the ADT model has the highest prediction capability for flash flood susceptibility assessment, followed by the NBT, the LMT, and the REPT, respectively. These strategies have proven to be effective in the rapid determination of flood-prone areas.

Hong et al. [6] introduce a new approach to building a flood susceptibility map in China by applying fuzzy proof weight (fuzzy-WofE) and data-mining methods. The thing that distinguishes the proposed approach is the use of fuzzy-WofE, which creates a preliminary flood sensitivity map and determines the variables associated with floods. LR, RF, and SVMs are implemented, taking into account the 11 flood-related variables. They evaluate the efficiency of their approach using the area under the curve (AUC). Their results show that the fuzzy WofE-SVM model produces the highest predictive performance (AUC value, 0.9865), which also appears to yield statistically significant differences from the other predictive models.

Tehrany et al. [19] examine and validate the hypothesis that the accuracy of the final susceptibility mapping result improves by adding more conditioning variables to the dataset used in river flood modeling. In addition, this research assesses the effect of individual conditioning influences on flood susceptibility mapping and their significance in the construction of accurate mapping of possible flood regions. They use DT and SVM to test spatial correlations between flood conditioning factors and rate their degree of importance for flood-prone mapping. They assess the accuracy for two ML approaches, SVM and DT, using the AUC method. The results show that SVM and DT provide the highest predictive accuracy levels of 85.52 and 88.47%, respectively, using DS1 (LiDAR dataset). Finally, it is concluded that the use of additional variables in the simulation does not necessitate the achievement of higher accuracy. Suliman et al. [20] review the related works of flood forecasting. The review focuses on the two most popular ML algorithms, which are ANN and SVM.

Choubin et al. [25] use two new algorithms, namely Multivariate Discriminant Analysis (MDA) and Classification and Regression Trees (CART) combined with the SVM algorithm in flood susceptibility analysis. They use these models with a flood inventory map and many factors of flood conditioning to develop a flood susceptibility map. A new framework for flood susceptibility assessment is proposed to ensure a more accurate ensemble model where only those models with an accuracy of 80% are permissible for use in ensemble modeling. The results show that the MDA model produces the highest predictive accuracy (89%), followed by the SVM (88%) and CART (83%) models. The ensemble modeling approach indicates that areas with a high population density are more vulnerable to floods, and therefore these areas should be given priority to flood prevention and treatment.

Liu et al. [26] combine Stacked Autoencoders (SAE) and Backpropagation Neural Networks (BPNN) to implement a new deep learning approach for floods prediction. To further develop the ability to model nonlinearity, their architecture performs two processes. First, K-means clustering is applied for data classification into various categories. Then, they represent their related data categories by using multiple SAE-BP modules. The results of the comparison between their approach and other approaches, which are SVM, BPNN, Radial-based Functions, and Extreme Learning Machine, show that their approach performs much better.

Widiasari et al. [27] provide a definition of the main model of the ANN that is useful in time series forecasting and a basic procedure for the practical implementation of the ANN in this form of mission. The model analyzed is the Multilayer Perceptron (MLP). To assess the degree of precision of the flood prediction, Mean Absolute Percentage Error (MAPE) is used in which the system predicts the lower the MAPE value, the more accurate the results. By using this, MLP achieves a MAPE value of 3.64%, which means that the error caused by the built-in device is 3.64% compared to the actual value used for testing. Also, MLP has a greater effect on the expected water level than multiple linear regression.

Widiasari et al. [28] use the Long Short-Term Memory (LSTM) algorithm, which is common and powerful at managing long-run periods of temporal dependencies for complex time-series data like precipitation and water elevation degree that can be sensed by using sensors. The model produces a MAPE value of 3.6% through the LSTM algorithm, which indicates that the error that is produced to predict the water level within the downstream river is 3.6% compared with the real water elevation value. LSTM produces more correct predictions of water elevation level inside the downstream if compared to MLR models that produce a MAPE prediction value of 10.55%.

From the previous contributions, we observe that ML algorithms are mainly used to predict floods. In Chen et al. [2], the RF model achieves the highest accuracy of 91.5% in predicting the flood status of five classes. The work of Khosravi et al. [5] achieves the highest accuracy of 94.3% in predicting the flood status of four classes through the ADT model. The work of Hong et al. [6] achieves the highest accuracy of 92.2% in predicting the flood status of five classes through the Fuzzy WofE-SVM model. However, none of these models introduces a complete architectural design of the flood alert system. They only focus on evaluating the ML ability to predict flood status from a maximum of five classes and neglect flood data acquisition.

3.1 Anglian river basin district (RBD) dataset

The dataset that is obtained from a specific source, such as the Internet, might be incomplete and inconsistent. Hence, selecting appropriate data is an important research issue [29,30]. The Anglian RBD includes 27,900 km² in which a total of 7.1 million people settled in this area. It includes the cities of Northampton, Lincoln, Chelmsford, and Milton Keynes as shown in Figure 1. The RBD dataset that is used in the research for flood prediction is an open dataset taken from the environmental agency [31]. The dataset includes a collection of datasets related to flood measurement, classification, environmental effect, and protection. The last update of the RBD dataset is on September 17, 2020. This dataset has 19 attributes and 149,676 instances, and the multivariate target classes of the dataset are 13 river states, as displayed in Table 1.

Figure 1

The RBD map [31].

3.2 ML algorithms

Data mining is a process of classification, audit, and semi-automatic analysis of very large amounts of data to obtain useful information and explore patterns and links [32]. Data mining is used in several different fields as a method in predicting data appropriately [33]. This section provides a summary of some data mining algorithms, which are DT, RF, and DJ algorithms.

DT is a structure of a branching tree that is used to determine the course of work or show the possibilities of a solution. Each internal node represents a test on an attribute, and each branch represents a potential DT [34]. Usually, an entropy function is deployed to control the DT splitting the data. The entropy affects the boundaries of a solution that the DT draws. The entropy formula is given in (1).

The DT provides a transparent tree-like structure with effective, easy rules that are easily interpreted and understood [35]. DT is used in flood forecasting of high-water levels and water flow [36]. Compared with other methods of classification, DT can be built quickly [34]. It has also been widely used for both continuous and discrete datasets. Variables screening and features selection are good enough in DT [37]. Regarding its performance, nonlinearity does not impact any of the DT parameters [38]. Hence, it is good to solve problems with multiple alternatives or situations that are related to risks and uncertainties. The basic rule in building a DT is to find the best question for each branch of the tree so that these questions divide the data into two sections. The first section applies to the question, and the second section does not apply so that through a series of questions, the DT is built with its chain of branches. Although the DT is used for exploration and data preparation for statistical operations, it is also used more often to predict the values of other cases not found in the training group [39].

RF is considered as an ensemble learning algorithm. It is utilized mainly for classification and regression tasks [40]. The RF architecture is represented as a multitude of DTs in which the output of the RF is the class that has an average or a majority selection among the DTs [41]. The prediction of a new case x after the training phase by using the average function is represented in (2).

For many classification problems, RF usually outperforms the DT for linear and nonlinear problems with an efficient mapping of input into forecasted spaces [42]. The design of the RF allows it to overcome the main problem of the DT, which is its tendency of training overfitting [43]. However, the RF performance is highlighted and affected by the characteristics of the data [44]. RF is currently considered as a popular classifier and has been integrated into many applications due to its stable performance in a wide range of classification problems [45].

DJ algorithm is an ensemble learning approach for classification. The algorithm works by utilizing and building different selection trees and then voting on the maximum famous output magnificence [46]. The trees that have high prediction have a greater weight in the final decision of the ensemble. A large number of applications have been developed using decision forests and trees in data science, although these approaches have certain drawbacks, such as given a large amount of data, the number of nodes in DTs may grow exponentially in size [47]. The DJ differs from DTs by the directed acyclic graph (DAG) group that permits multiple paths from a root to the leaves, whereas traditional DTs allow only one path per node [48]. The DJ method has a strong contrast between the two algorithms as it combines two modern nodes, thereby improving the function and structure of the DAG.

3.3 Evaluation metrics

To compare the performance of each of the classifications, preprocessing is carried out based on all the values of the taken 19 attributes in which no feature selection has been performed. A comparative study of DT, DJ, and RF classification results is performed with a three-fold validation method for training and testing [49,50,51].

• Accuracy: It is the most commonly used metric to judge a model and is not a clear indicator of performance. The worse happens when classes are imbalanced.

  (3) Accuracy = ( TP + TN ) / ( TP + FP + TN + FN ) .

• Precision or positive predictive value: It is the ratio of correctly classified attack flows (TP) in front of all the classified flows (TP + FP).

  (4) Precision = TP / ( TP + FP ) .

• Recall or sensitivity: It is the ratio of correctly classified attack flows (TP), in front of all generated flows (TP + FN).

where TP, TN, FP, and FN have their usual meaning.