A Review on Flood Prediction Algorithms and A Deep Neural Network Model for Estimation of Flood Occurrence

.


Introduction
Flood visits most part of the country, nearly on a yearly premise. The floods may occur due to heavy rainfall and the overflow of the rivers in those rainy regions. The districts of those regions under the flood encounter precipitation, year after year. The precipitation contributes a great deal to the flooding. Areas situated in a plain, this does not improve the situation much. When it rains in the upper area, the lower fields experience the anger of the waterway. But due to recurrent and merciless occurrence of flood every year the people of those regions, the flora and fauna faces lots of loss and endangerment. To control the flooding, the experts can manufacture dykes on either side of the channel to confine the flood of water. Without such measures, the flooding will keep on unleashing ruin and decimation. The intermittent decimation can be ceased just if measures are taken to spare individuals and property.
In other to control this damage, it is necessary to propose a prediction algorithm to know about the occurrence of flood in the area beforehand so that proper measures can be taken to prevent damages and preserve the nature. However, the prediction of flood lead time and occurrence location is fundamentally complex due to the dynamic nature of climate condition. Therefore, nowadays major flood prediction models are mainly data-specific and involve various simplified assumptions. The neural networks play and important role in flood predication and the accuracy could be increased by applying suitable deep learning algorithms. This paper helps to analyze few prediction models and propose a model based on Recurrent Neural Network with LSTM and GRU.
The rest of the paper has been organized as follows. The section II shows various classification of the prediction algorithm and few methods in each classification. Section III shows the proposed model and its working functions, followed by section IV that concludes the paper.

Existing Techniques
In this, various existing techniques for flood prediction are discussed. Fig.1 shows few classifications of flood prediction algorithms.

A. Digital Image Processing
Image Processing is one of most prominent method to predict the floods with the help of digital image made up of finite number of components known as pixels [1].
Complex water flow caused by Local Rainfall was predicted by [2] infrared images captured at low cost. Linear Regression algorithm was applied to predict the flood. This is also integrated with sensors to read the data and to improve the accuracy [3]. Using the digital images obtained, the land area and the water level are extracted. Estimate the difference between the land area and the water-level for the selected region. If the difference between the two areas is close to zero it is highly possible that the area could be flooded. If the difference is negative then flood has already occurred [4][5][6][7].
[8] used image segmentation with semi-supervised learning algorithm to handle flood mapping. The algorithm helps to divide the satellite image into patches with graph-based approach to detect the water areas. Monitoring and prediction of flood will be efficient when we made the observation from the space. But the usage of optical image may be failed in the presence of cloud. The image capture from Synthetic Aperture Radar (SAR) [16] provide valuable information as they are independent of weather conditions.

B. Wireless Sensors
Wireless Sensor Network has a wide range of applications including weather prediction and a major impact in predicting the flood [5]. Wireless Sensor Network (WSN) is a component of information retrieval that provides more optimal results in obtaining data time series.
[5] predicts the water elevation level in data time series controlled by remote site and control center. Water level sensor system with real-time river-level monitoring helps in flood prediction. The water level sensor system defined by [2] uses Linear Regression Approach for data assimilation to the collected information from the sensor. This helps to improve the accuracy in predicting the flood.

C. Multiple Layer Perceptron
Multi-Layer Perceptron is a feed-forward artificial neural network. The network is integrated with numerous nodes through linking weights. The main purpose of the MLP network is to map the group of inputs into a set of desired output [6]. This MLP also plays a major role in flood prediction by processing the collected information.
[5] used Multilayer Perceptron a deep learning algorithm in time series data with neural network. Produce better results in water elevation on downstream canal. [6] The activation functions such as Sigmoid or logistic, tanH, Rectified Linear Unit (ReLU) are used in hidden layers to fetch the occurrence of flood in the future is predicted. The following equations helps to define the activation function as defined in [6] is given in equation 1 and 2.

D. NARX Neural Network
Nonlinear dynamic systems are modelled using dynamical neural architecture (NARX) [7]. It has been proven that utilizing NARX systems, instead of traditional recurrent systems helps in improving the prediction with no computational loss and that they are at any rate equal to Turing machines [9] helps to improve the prediction by incorporating machine learning algorithms to train the dataset generated. [10] shows gradient descent learning can be more successful in NARX systems than in other intermittent structures with "hidden states". [7] proposed an architectural recurrent neural network with embedded memory for flood prediction. This uses non-linear autoregressive model to model the input and the analysis is done with the chaotic time series prediction. [11] proposed flood water level modelling to overcome nonlinearity problem in flood prediction. The methods check the preciseness of the true model and the prediction model to predict the error. The best fit for this model is 87.0799% and RMSE is 0.0806m. [13] The traditional Soil Conservation Service model is low efficient for medium based flood forecast. To overcome this streaming big data processing is integrated with Soil Conservation Service in which basins are divided into sub-basins and are executed in parallel to predict the flood. The results show that this overall computation time is reduced by 73%. [12] an analytical model using Soil Conversation serviced is proposed for flood prediction. [14] proposed a survey of different soil conservation methods and its various applications and the various improvement to be done in the model for flood prediction. [15] shows an improvement in SCS model to adapt runoff calculation in large and small basins. [16] uses remote sensing and geographic information system to extract and analyse the spatial and attribute data. Soil Conservation Service curve number method (SCC-CN) is applied to calculate the runoff coefficients of different underlying surfaces characteristics. Then the amount of rainwater collected in the basin is estimated and it achieves good results.

F. Wavelet Neural Network
Wavelet neural network theory works on reconstruction theory of wavelet function. The combination between wavelet transformation and neural network is built by wavelet decomposition of consistent approaching. This helps in predicting the flood. [15] used wavelet neural networks in which water flux and water level are used as an input for which the time frequency feature was analysed. The accuracy of the result is increased by combing the low and high frequencies of signal which are decomposed into large scales.
[17] used scale vector, transfer vector, weight vector for the wavelet network. In which the weights and displacement and scale vectors are updated based on the activation function the main advantage of the prediction is the non-linear time series and non-linear function approach. [17] used temporal and spatial availability of data in which binarization and image differences are considered for pre-processing which is mapped with class of type water, urban, forest, agriculture and bare land respectively. Results show that it produces a fast and more appropriate result than existing methods. There are two groups of Bayesian approaches in document categorization: Naïve and non-naïve Bayesian approaches. The naïve part of the former is the assumption of word independence, meaning that the word order is irrelevant and consequently that the presence of one word does not affect the presence or absence of another one A disadvantage of Bayesian approaches in general is that they can only process binary feature vectors

Other Classifiers
Other classification techniques include random forest, recurrent neural networks, pattern-based classification Every method has its own advantage and disadvantages, few of the disadvantages, few of the disadvantages of various techniques have been discussed [16] applies backscatter temporal variability on the SAR image to classify open water body and the land. They done the analysis in the heritage site Kaziranga National Park, Natural Wonder of the World, Assam during the monsoon period of 2017 with six SAR images captured during the period of 21st march to 26th July 2017. This helps to predict the flood in the region. [19] proposed image fusion technique, a combination of two statistical threshold approaches to increase the accuracy of the prediction for optical and SAR images.

Novel Approach for Flood Prediction
From the above survey, it has been identified that Neural network plays an important role in predicting the flood. So, in this, a model for flood prediction based on Recurrent Neural Network is proposed. Two main issues faced by existing RNN are how to get the knowledge of the previous steps and vanishing gradient are solved in this model.
First problem can be solved by the introduction of the hidden layers [18] which is responsible for remembering the information about a sequence. For each input, the hidden layer uses the same set of parameters to perform the same task. Downside of RNN is that it has to put up with the vanishing gradient problem [19] with Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU).
This LSTM and GRU have an impact in analyzing the data using deep learning algorithms. The combination of these two methodologies will improve in the accuracy of the prediction and the occurrence of flood in the near future. The flow diagram of the model is shown in the Fig. 2. Input layer will have the parameters as pressure, precipitation amount, dew point, wind speed which are the factors considered for the proposed model. The next step is Max pooling, a sample-based discretization process that down sample an input representation by reducing its dimensionality with the features contained in the sub-regions. Next to overcome the vanishing gradient problem of recurrent neural network, Long Short-Term Memory (LSTM) is used. LSTM contains four gates namely the input gate, forget gate, cell gate and the output gate along with pointwise multiplication and pointwise addition. The trained data will be then fed to the Gated Recurrent Unit (GRU). The GRU has two gates namely reset gate and the update gate along with pointwise multiplication and pointwise addition. All these are concatenated and formulated into a single output, which gives us the desired output for predicting the occurrence of flood using the above-mentioned parameters as inputs. The conventional method uses sensors in major rivers to detect the water level of the rivers. These data are sent to a remote server where using data processing the occurrence of flood is predicted. Various other factors that causes flood are not taken into consideration. However, with this model the various other factors that lead to the occurrence of flood are taken into consideration which improves the accuracy for predicting flood over the conventional method.

Conclusion
The classification of the prediction algorithms is discussed and we modeled a prediction system based on neural network using a combination of LSTM and GRU to predict the occurrence of flood. In future, the model can be trained with the factors like temperature, pressure, precipitation amount, dew point, wind speed, etc. which plays an important role in the occurrence of flood. These factors will be used to train and test the neural network model and finally predict the occurrence of flood in the near future. .