Applications of Machine Learning in Hydroclimatology

Accurate prediction of different components of a hydrological cycle is of central interest to hydrologists for guiding water-resources-related policies. Over the years, various modelling frameworks have been developed and tested worldwide. These modelling frameworks, generally, are classified into three types: physics-based or process-based, black-box or data-driven and conceptual models. The physics-based models try to represent the catchment scale physical processes using the equations governed by mass, momentum and energy balance principles. The data-driven models, on the contrary, focus on mapping the input to the output by employing linear or non-linear transfer functions that seldom consider the underlying physical process. The conceptual modelling frameworks that lie in between the above two frameworks focus on representing the dominant hydrological processes in simplified mathematical forms. According to a few recent studies, deep-machine learning-based models that come under the category of data-driven models outperform the well-established conceptual hydrological models. These studies reported that the deep-learning models can better capture the information available in the model input over the traditional conceptual models. However, since the deep-learning-based models are physically inconsistent, their applicability under unforeseen scenarios may lead to significant ambiguity in the forecast. Taking into consideration the pros and cons of the data-driven models, a framework called physics-guided machine learning (PhyML) has been gaining popularity in recent years. It is built on the notion of incorporating the physical process understanding in the data-driven models. There are several ways of building the PhyML, such as providing an output of a physics-based model as input to an ML-based model, applying mass balance constrain in ML models, etc. The contrasting approach is also possible where a component of the conceptual hydrological model is simulated using ML. The current chapter discusses the recent applications of PhyML in the field of hydrology and compares the performance of a PhyML model with its simple ML variant for a case of dry-weather flow prediction.

Streamflow forecasting is one of the most crucial exercises for reservoir operations, hydropower generation, and irrigation scheduling management. Conventionally, it is done by developing a forecasting model based on historical records of various variables. Broadly, the models can be divided into either physically based mathematical models or data-driven “black box” models. To understand the hydrologic system, the development of physically based models necessitates a vast quantity of diverse data connected to numerous physical processes, which is typically challenging to gather under actual field conditions, especially in developing nations. On the other hand, data-driven models do not require system expertise, and as a result, they have gained popularity, particularly with improvements in artificial intelligence technology. Stochastic time series models have historically been employed to predict streamflow. Because Artificial Neural Network (ANN) models forecast streamflow more accurately than time series models did in the 1990s, they became increasingly popular. Artificial intelligence (AI), an offshoot of computer science, can analyse long-series and large-scale hydrological data. In recent years, applying AI technology to hydrological forecasting modelling has been one of the front-burner issues. Streamflow forecasting also uses various AI techniques like SVM, Fuzzy Logic, GP, GEP, etc. Later, by merging them with other models, the performance of the current AI-based data-driven models is improved. This chapter discusses various AI-based models, their basic concepts, types, and applications for forecasting streamflow.

Prediction of groundwater levels using machine learning techniques has gained substantial attention over the past few decades. Several researchers have reported the advances in this field and provide clear understanding of the state-of-the-art machine learning models implemented for GWL modelling. However, not many studies discuss the application of these results, which is of utmost importance to the practicing engineers and decision-makers for efficient planning and management of water resources infrastructure. The first part of this chapter provides a detailed review of the existing machine learning models in the domain of groundwater hydrology. This review provides a summary of all types of machine learning models developed for estimating groundwater levels over the past decade. The review also discusses the challenges associated with the use of machine learning in groundwater levels estimation. In addition, several studies from the recent past indicate the dominance of Ensemble Machine Learning in managing the sustainability of groundwater across the globe. So, the ability of ensemble machine learning models in estimating the groundwater level is discussed in the chapter. Furthermore, the chapter provides recommendations to enhance the knowledge in this domain and directions for the possible future research to improve the accuracy of the machine learning models in estimating groundwater levels.

This paper presents the applicability of deep learning techniques for river discharge forecasting. Deep learning techniques have been an area of interest for forecasting in engineering because of their forecasting accuracies. Two deep learning techniques long-short term memory (LSTM) and bidirectional long-short term memory (Bi-LSTM) have been applied and compared for forecasting river discharge data in this study. In LSTM, input flows in one direction (forward) as it has only one layer. But in Bi-LSTM, input flows in two directions (forward and backward) as it has an additional layer and the output of both layers is combined together for the final output. For the evaluation of this study, daily discharge data of the Salehbhata station in the Mahanadi River basin has been used. Two statistical measures which are Nash-Sutcliffe efficiency (NSE) and root mean square error (RMSE) were used for model evaluation. Results show that both models perform well but Bi-LSTM is slightly better than LSTM in terms of both statistical measures. These results of both models are with different values of hyperparameters. The performance of these models can be different with the same values of hyperparameters and comparison of both models with the same hyperparameters may not be ideal. From this study, it is concluded that Bi-LSTM can be an alternative approach for forecasting river discharge data.

Groundwater is considered as cleanest and widely used source of water throughout the world. Proper management of groundwater resources requires continuous monitoring of groundwater level. To make decisions and plans for future, decision-makers need reliable tools to predict the groundwater level by developing relationships between the factors/parameters that can affect the dynamics of groundwater. With the effects of climate change such as increasing heat, higher rainfall and frequently occurring extreme weather events such as floods and draughts, the need of tools and models for studying the effects of changing climate on groundwater level fluctuations has increased. Such tools and models help decision-makers in better decision-makings. With the recent development in the field of artificial intelligence, the reliable predictions of groundwater levels are possible using machine learning models. Machine learning models are data-driven models which can produce useful results utilizing the diverse data sets and computational infrastructure. This chapter focuses on types of machine learning models such as Artificial Neural Network (ANN), Adaptive Neuro Fuzzy Inference System (ANFIS), Support Vector Mechanism (SVM). Random Forest (RF) and Decision Trees. A detailed review of the algorithm of each model and their application in groundwater level prediction is discussed. A brief literature review of groundwater level prediction using machine learning approaches is presented for better understanding of the topic.

Critical shear stress (CSS) of sediment governs its transport in open channel flow. Earlier, mathematical models were developed to compute CSS of sediment under clay influence which was limited to their own experimental data. The present study aims to develop an ANN model to compute CSS of coarser sediment present in mobile channel bed made of cohesive sediment mixture. The proposed model was optimized using parent-selected operator GA (genetic algorithm), which helped to determine the optimal process parameters responsible for CSS. Initially, the mathematical model was implemented with the help of ANN (artificial neural network) and later on it was optimized by three well-known parent-selected operator genetic algorithms. Data from the literature along with current experimental data was used to develop ANN-based model to compute CSS for coarser particle under clay influence. It was found that linear ranked selected GA-ANN is the best-fitted model for both training and testing data. The CSS obtained for optimized input values for clay fraction by weight (CP), weighted geometric standard deviation of sediment mixture (SM), and dimensionless dry bulk unit weight of cohesive sediment (DB) are 0.104, 2.9523, and 1.6921 respectively which were quite satisfactory and performing better than the existed mathematical models.

The spatio-temporal variations in stream water quality are primarily due to the heat and mass exchanges occurring at the interfaces which are propagative and dilutive along the stretch. The streams are also susceptible for increased pollutant loading from various point and non-point sources causing an imminent depletion of dissolved oxygen (DO) concentration. Though water quality parameters are several, it is pertinent to identify the most critical factors for an easy, reliable and meaningful monitoring system. It is also necessary to have a simplified data-driven model for making necessary predictions for various decision-making processes. The present study investigates the characteristic relationship between various forms of oxygen demands (carbonaceous biochemical oxygen demand—cBOD and sediment oxygen demand—SOD) in a stream with the available DO for a selected stretch of River Bhavani in Tamil Nadu, India. The modelling framework consists of simulating a data-extensive water quality profile using QUAL2K and integrating the outputs for minimizing the required count of parameters using a multi-perceptron feed-forward artificial neural network (ANN) model. The proposed methodology suggested that a minimum of five parameters (inorganic suspended solids—ISS, DO, cBOD, SOD and total nitrogen—TN) are sufficient to simulate the concentration profiles with reasonable accuracy (R² varies from 0.88 to 0.95). For any local addition of organic-rich influent to the sediment load of the river, there is a corresponding change in the cBOD and DO within the selected stream profile. The results from the present study indicate the advantage of the combined modelling approach for prediction of water quality in case of complex interactions between various oxygen-demanding substrates.

This study aims to evaluate the accuracy of two deep learning models, gated recurrent unit (GRU) and long short-term memory (LSTM), for predicting daily suspended sediment load (SSL) in the Missouri River at Omaha, NE in the United States and compare the results with a traditional time series model, autoregressive integrated moving average (ARIMA) model. The study used daily data of SSL in the Missouri River at Omaha, NE from October 2008 to September 2018. The two deep learning models, GRU and LSTM, were applied to the data and their accuracy was compared with the ARIMA model. The models were trained and tested using a 70–30 split of the data and evaluated based on the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). The results showed that GRU had a better performance with a coefficient of determination of 0.871 and a root mean square error of 0.142, compared to LSTM’s coefficient of determination of 0.865 and root mean square error of 0.148. GRU also had a lower mean absolute error of 0.097 compared to LSTM’s mean absolute error of 0.101. The study concludes that both GRU and LSTM can be used effectively in SSL modeling. However, GRU requires fewer training constraints and has a faster performance with higher accuracy compared to LSTM, which may need additional data to improve accuracy. The results suggest that GRU can be a promising tool for daily SSL prediction in river basins.

Climate models play a pivotal role in understanding the complex dynamics of the Earth’s climate system. The C3S has contributed significantly to our knowledge, providing a wealth of rainfall output from diverse global climate models. However, the inherent spatial resolution of these models often limits their direct applicability for regional scale assessments. This study introduces a Hybrid Gamma Probability Distribution (HGPD) Model designed for comprehensive evaluation of C3S rainfall outputs at local level. The proposed hybrid method incorporates advanced statistical downscaling techniques to bridge the gap between the coarse resolutions of C3S models and the finer scales required for regional assessments. Model development involves the calibration and validation of the downscaling technique using observed rainfall data, ensuring robust performance across diverse climatic regions. The SD model is tailored to capture the intricate spatial patterns of rainfall, enhancing the accuracy of regional scale assessments. To showcase the applicability and effectiveness of the developed SD approach, a grid-level evaluation is conducted using rainfall forecasts. The evaluation encompasses various climate zones, providing a comprehensive overview of the downscaled model’s performance in different geographical and climatic contexts.

Key metrics such as spatial distribution, intensity, and variability are analyzed to assess the downscaling model’s ability to reproduce observed rainfall patterns.

The findings from this study contribute to advancing our understanding of the reliability and limitations of C3S rainfall outputs on a local scale. The developed downscaling approach offers a valuable tool for researchers and policymakers engaged in regional climate impact assessments, facilitating more accurate and informed decision-making in the face of climate variability and change. This research represents a significant step toward improving the utility of C3S data for regional applications, thereby enhancing our ability to address the complex challenges posed by a changing climate.

Machine learning techniques are now extensively being used for prediction of long-term and short-term rainfall time series datasets because the lack of data and availability of recorded daily rainfalls may affect the feasibility of a larger range of rainfall-based studies in light of repercussions from climate change and extreme hydro-meteorological phenomena. A daily gridded rainfall time series dataset (2007–2020) over Mizoram state is constructed using data machine learning methods such as Support Vector Regression and Random Forest. For this purpose, the open sources and observed gridded rainfall datasets CHIRPS, APHRODITE, IMDAA re-analysis, and IMD have been utilized. In this study, the statistical evaluation methods have been employed and tested over selected rainfall grids to enhance the accuracy of the datasets. Quantile-quantile (Q-Q) plots were also used to test the accuracy of each dataset in the case of extreme rainfalls with respect to IMDAA gridded rainfall (i.e., IMD Re-analysis), which has been taken as the observed/reference dataset. In this study, the quantile mapping and linear scaling bias correction methods are utilized to correct the rainfall datasets. The predicted rainfall datasets have been compared produced by SVR and RF, and the RF-based predicted rainfall datasets have performed superior than SVR datasets in terms of bias and extremity.