Search

This paper presents a dynamical neural network framework to understand how catchment systems respond to daily rainfall-runoff processes over time. We developed an interactive Python-based deep neural network (DNN) package called FlowDyn (presented through a JS-based web platform) to simulate and forecast daily streamflow data for >180 gauging stations across the globe. Several DNN models, including long short-term memory (LSTM), Gated Recurrent Unit (GRU), and a hybrid network of convolutional neural network and LSTM (ConvLSTM), as well as an auto encoder (AE) model were developed and integrated into the FlowDyn pipeline to analyze and forecast sequential daily streamflow values that are embedded within a web-based application for demonstration and visualization. Inputs were gathered from different web services, including the catchment attributes and meteorology for large-sample studies (CAMELS), the national climatic data center (NCDC), and the global runoff data center (GRDC). DNN configurations were trained and tested with an average accuracy rating of 0.83 across 183 river basins globally. FlowDyn simulation and performance demonstrated that different DNN models were able to learn both regionally consistent and location-specific hydrological behaviors. Through the findings of this paper, we advocate the merit of applying FlowDyn package in the field of daily rainfall-runoff prediction at both local and global scales. • A dynamic framework was developed to predict daily streamflow records of >180 gauging stations across the globe. • The FlowDyn pipeline offers functionalities, such as employing multiple DNNs, various performance metrics, and visualization. • FlowDyn has the ability to retrain and provide more accurate results in case of extending the datasets. [ABSTRACT FROM AUTHOR]

Research activity in the field of air pollution forecasting using artificial neural networks (ANNs) has increased dramatically in recent years. However, the development of ANN models entails levels of uncertainty given the black-box nature of ANNs. In this paper, a protocol by Maier et al. (2010) for ANN model development is presented and applied to assess journal papers dealing with air pollution forecasting using ANN models. The majority of the reviewed works are aimed at the long-term forecasting of outdoor PM 10 , PM 2.5 , and oxides of nitrogen, and ozone. The vast majority of the identified works utilised meteorological and source emissions predictors almost exclusively. Furthermore, ad-hoc approaches are found to be predominantly used for determining optimal model predictors, appropriate data subsets and the optimal model structure. Multilayer perceptron and ensemble-type models are predominantly implemented. Overall, the findings highlight the need for developing systematic protocols for developing powerful ANN models. • Research activity in ambient air pollution forecasting with ANNs continues to grow. • Forecasting of outdoor PM10, PM2.5, nitrogen oxides and ozone levels was widely done. • Feedforward and hybrid ANN model types were predominantly used. • Most of the identified model building steps were done in an ad-hoc manner. [ABSTRACT FROM AUTHOR]

Many environmental variables, in particular, related to air or water quality, are measured in a limited number of points and often for a limited time span. This forbids the development of accurate models for interesting locations with missing or insufficient data and poses the question of whether a model developed for another measurement site can be reliably applied. Such a question is particularly critical when the model is entirely data-driven, such as a neural network. In this context, the paper proposes a procedure to evaluate the expected performance of an existing neural network model applied to a new unmonitored station. This transferability assessment is exemplified by the problem of forecasting ozone concentrations in different environmental settings around the Alpine Arc. Long Short-Term Memory (LSTM) neural network models are applied for predicting hourly concentrations in 20 stations of different types (urban, rural, and mountain). The analysis of the results allows us to determine the expected performance of such models in new cases and reduce the transferability uncertainty when the existing models can be partitioned into clusters. The LSTM models demonstrate the possibility of high accuracy in ozone forecasting at all sites. Given the significant impacts of this gas on human health and the environment, this can contribute to better decision-making and mitigation strategies for air pollution control. • We investigate the possibility of training a model in a site and reuse it elsewhere. • We introduce the concept of transferability matrix and transferability indicators. • The transfer learning reliability is enhanced by clustering the sites. • Ozone concentration forecasting for 20 locations in the Alpine region is considered. [ABSTRACT FROM AUTHOR]

Abstract Data Mining (DM) is a fundamental component of the Data Science process. Over recent years a huge library of DM algorithms has been developed to tackle a variety of problems in fields such as medical imaging and traffic analysis. Many DM techniques are far more flexible than more classical numerial simulation or statistical modelling approaches. These could be usefully applied to data-rich environmental problems. Certain techniques such as artificial neural networks, clustering, case-based reasoning or Bayesian networks have been applied in environmental modelling, while other methods, like support vector machines among others, have yet to be taken up on a wide scale. There is greater scope for many lesser known techniques to be applied in environmental research, with the potential to contribute to addressing some of the current open environmental challenges. However, selecting the best DM technique for a given environmental problem is not a simple decision, and there is a lack of guidelines and criteria that helps the data scientist and environmental scientists to ensure effective knowledge extraction from data. This paper provides a broad introduction to the use of DM in Data Science processes for environmental researchers. Data Science contains three main steps (pre-processing, data mining and post-processing). This paper provides a conceptualization of Environmental Systems and a conceptualization of DM methods, which are in the core step of the Data Science process. These two elements define a conceptual framework that is on the basis of a new methodology proposed for relating the characteristics of a given environmental problem with a family of Data Mining methods. The paper provides a general overview and guidelines of DM techniques to a non-expert user, who can decide with this support which is the more suitable technique to solve their problem at hand. The decision is related to the bidimensional relationship between the type of environmental system and the type of DM method. An illustrative two way table containing references for each pair Environmental System-Data Mining method is presented and discussed. Some examples of how the proposed methodology is used to support DM method selection are also presented, and challenges and future trends are identified. Highlights • A methodology to support Data Mining (DM) method choice in environmental sciences. • DMMCM is a conceptual map providing an overview of common Data Mining methods. • DMMT provides templates of main DM families, to simplify use by practicioners. • Extensive real environmental applications and 5 detailed case studies are shown. • Existing and future challenges in DM for Data Science are postulated. [ABSTRACT FROM AUTHOR]

Time series analysis with explanatory variables encompasses methods to model and predict correlated data taking into account additional information, known as exogenous variables. A thorough search in literature returned a dearth of systematic literature reviews (SLR) on time series models with explanatory variables. The main objective is to fill this gap by applying a rigorous and reproducible SLR and a bibliometric analysis to study the evolution of this area over time. The study resulted in the identification of the main methods of time series that incorporate input variables per knowledge area and methodology. The largest number of papers belongs to environmental sciences, followed by economics and health. Regression model is the method with the highest number of applications, followed by Artificial Neural Networks and Support Vector Machines, which experienced rapid and recent growth. A research agenda in time series analysis with exogenous variables closes the paper. [ABSTRACT FROM AUTHOR]

The aim of this paper is to describe the state-of-the art computer-based techniques for data analysis to improve operation of wastewater treatment plants. A comprehensive review of peer-reviewed papers shows that European researchers have led academic computer-based method development during the last two decades. The most cited techniques are artificial neural networks, principal component analysis, fuzzy logic, clustering, independent component analysis and partial least squares regression. Even though there has been progress on techniques related to the development of environmental decision support systems, knowledge discovery and management, the research sector is still far from delivering systems that smoothly integrate several types of knowledge and different methods of reasoning. Several limitations that currently prevent the application of computer-based techniques in practice are highlighted. [ABSTRACT FROM AUTHOR]

Earth Observation (EO) technologies have played an increasingly important role in monitoring the Sustainable Development Goals (SDG). These technologies often combined with Machine Learning (ML) models provide efficient means for achieving the SDGs. The great progress of this combination is also demonstrated by the large number of software, web tools and packages that have been made available for free use. In this paper, we introduce a software architecture to facilitate the generation of EO information targeted towards soil moisture that derive several challenges regarding the facilitation of satellite data processing. Thus, this paper presents a web-based tool for Soil Moisture Estimation (SMETool), designed for the soil moisture estimation using Sentinel-1A and Sentinel-2A data based on Eo-learn library. SMETool implements several ML techniques such as (Artificial Neural Network (ANN), Random Forest (RF), Convolutional Neural Network (CNN), etc.). The SMETool could be very useful for decision makers in the region in assessing the effects of drought and desertification events. Experiments were carried out on two sites in Tunisia during the period from 2016 to 2017. Although the performance of the used models is very close, it is clear that CNN and RF outperformed other ML models. The achieved results reveal that the soil moisture, was highly correlated to the in-situ measurements with high Pearson's correlation coefficient r ( r RF = 0. 86 , r ANN = 0. 75 , r XGBoost = 0. 79 , r CNN = 0. 87) and low Root Mean Square Error (RMSE) (RMSE RF = 1.09%, RMSE ANN = 1.49%, RMSE XGBoost = 1.39%, RMSE CNN = 1.12%), respectively. • A new web-based tool for soil moisture estimation at the scale of arid regions. • Simple implementation using Eo-learn framework and machine learning methods. • A case study on arid region in southern Tunisia using Sentinel 1-A and Sentinel 2-A products. [ABSTRACT FROM AUTHOR]

Artificial Neural Networks (ANNs), sometimes also called models for deep learning, are used extensively for the prediction of a range of environmental variables. While the potential of ANNs is unquestioned, they are surrounded by an air of mystery and intrigue, leading to a lack of understanding of their inner workings. This has led to the perpetuation of a number of myths, resulting in the misconception that applying ANNs primarily involves "throwing" a large amount of data at "black-box" software packages. While this is a convenient way to side-step the principles applied to the development of other types of models, this comes at significant cost in terms of the usefulness of the resulting models. To address these issues, this inroductory overview paper explodes a number of the common myths surrounding the use of ANNs and outlines state-of-the-art approaches to developing ANNs that enable them to be applied with confidence in practice. [Display omitted] • Commonly held myths about artificial neural network (ANN) modelling are articulated and busted. • Good modelling practices applied to other types of predictive models are also applicable to ANNs. • State-of-the-art approaches to developing ANN models for prediction and forecasting are presented. • Size matters – ANNs with more inputs and hidden nodes are more likely to give non-sensical results. • Using an independent validation set is insufficient to ensure ANNs can be used with confidence in practice. [ABSTRACT FROM AUTHOR]

This paper reports the refinement of a published empirical model of salinity in the San Francisco Bay-Delta estuary by integration with a Bayesian artificial neural network (ANN) model and incorporation of additional inputs. Performance goals established for the resulting hybrid model are based on the quality of fit to observed data (replicative and predictive validation) as well as sensitivity when compared with a priori knowledge of system behavior (structural validation). ANN model parameters were constrained to provide plausible sensitivity to coastal water level, a key input introduced in the hybrid formulation. In addition to representing observed data better than the underlying empirical model while meeting structural validation goals, the hybrid model allows for characterization of prediction uncertainty. This work demonstrates a real-world application of a general approach— integration of a pre-existing model with a Bayesian ANN constrained by knowledge of system behavior—that has broad application for environmental modeling. [ABSTRACT FROM AUTHOR]

Validation is a critical component of any modelling process. In artificial neural network (ANN) modelling, validation generally consists of the assessment of model predictive performance on an independent validation set (predictive validity). However, this ignores other aspects of model validation considered to be good practice in other areas of environmental modelling, such as residual analysis (replicative validity) and checking the plausibility of the model in relation to a priori system understanding (structural validity). In order to address this shortcoming, a validation framework for ANNs is introduced in this paper that covers all of the above aspects of validation. In addition, the validann R-package is introduced that enables these validation methods to be implemented in a user-friendly and consistent fashion. The benefits of the framework and R-package are demonstrated for two environmental modelling case studies, highlighting the importance of considering replicative and structural validity in addition to predictive validity. [ABSTRACT FROM AUTHOR]

This paper presents a methodology that combines the power of an Artificial Neural Network and Information Theory to forecast variables describing the condition of a regional system. The novelty and strength of this approach is in the application of Fisher information, a key method in Information Theory, to preserve trends in the historical data and prevent over fitting projections. The methodology was applied to demographic, environmental, food and energy consumption, and agricultural production in the San Luis Basin regional system in Colorado, U.S.A. These variables are important for tracking conditions in human and natural systems. However, available data are often so far out of date that they limit the ability to manage these systems. Results indicate that the approaches developed provide viable tools for forecasting outcomes with the aim of assisting management toward sustainable trends. This methodology is also applicable for modeling different scenarios in other dynamic systems. [ABSTRACT FROM AUTHOR]

Input variable selection (IVS) is vital in the development of data-driven models. Among different IVS methods, partial mutual information (PMI) has shown significant promise, although its performance has been found to deteriorate for non-Gaussian and non-linear data. In this paper, the effectiveness of different approaches to improving PMI performance is investigated, focussing on boundary issues associated with bandwidth estimation. Boundary issues, associated with kernel-based density and residual computations within PMI, arise from the extension of symmetrical kernels beyond the feasible bounds of potential inputs, and result in an underestimation of kernel-based marginal and joint probability distribution functions in the PMI. In total, the effectiveness of 16 different approaches is tested on synthetically generated data and the results are used to develop preliminary guidelines for PMI IVS. By using the proposed guidelines, the correct inputs can be identified in 100% of trials, even if the data are highly non-linear or non-Gaussian. [ABSTRACT FROM AUTHOR]

Eutrophication represents an important ecological and environmental issue in coastal lagoons. This paper presents an extensive study of recurrent cell and network architectures to model eutrophication processes in the Venice lagoon, a very complex and fragile ecosystem that has been strongly altered by anthropic activities over years. Experimental results showed that recurrent models outperformed Random Forests (RF) significantly on two datasets, performing similarly to CNNs on one of the datasets, while outperforming CNNs on the other one. Additionally, the transferability potential of the trained models was investigated. The empirical analysis has shown that recurrent neural network models with lower computational complexity provide the highest eutrophication prediction accuracy when their trained models were tested on a new dataset. Designed models represent effective tools for early-warning eutrophication prediction that can support the implementation of relevant EU acquis (EU Marine Strategy and Water Framework Directives) and achievement of their environmental targets. • Extensive study of a wide variety of recurrent neural network methods for eutrophication prediction is conducted. • The transferability potential of the developed models was investigated. • Eutrophication Recall Rate (ERS), specific to the Venice Lagoon case study, was introduced. • Multivariate approach adopting five WQ parameters provided significant prediction performance. • Recurrent models with low computational complexity had higher transferability potential. [ABSTRACT FROM AUTHOR]

Metamodels have proven be very useful when it comes to reducing the computational requirements of Evolutionary Algorithm-based optimization by acting as quick-solving surrogates for slow-solving fitness functions. The relationship between metamodel scope and objective function varies between applications, that is, in some cases the metamodel acts as a surrogate for the whole fitness function, whereas in other cases it replaces only a component of the fitness function. This paper presents a formalized qualitative process to evaluate a fitness function to determine the most suitable metamodel scope so as to increase the likelihood of calibrating a high-fidelity metamodel and hence obtain good optimization results in a reasonable amount of time. The process is applied to the risk-based optimization of water distribution systems; a very computationally-intensive problem for real-world systems. The process is validated with a simple case study (modified New York Tunnels) and the power of metamodelling is demonstrated on a real-world case study (Pacific City) with a computational speed-up of several orders of magnitude. [ABSTRACT FROM AUTHOR]

Input Variable Selection (IVS) is an essential step in the development of data-driven models and is particularly relevant in environmental modelling. While new methods for identifying important model inputs continue to emerge, each has its own advantages and limitations and no single method is best suited to all datasets and modelling purposes. Rigorous evaluation of new and existing input variable selection methods would allow the effectiveness of these algorithms to be properly identified in various circumstances. However, such evaluations are largely neglected due to the lack of guidelines or precedent to facilitate consistent and standardised assessment. In this paper, a new framework is proposed for the evaluation and inter-comparison of IVS methods which takes into account: (1) a wide range of dataset properties that are relevant to real world environmental data, (2) assessment criteria selected to highlight algorithm suitability in different situations of interest, and (3) a website for sharing data, algorithms and results ( http://ivs4em.deib.polimi.it/ ). The framework is demonstrated on four IVS algorithms commonly used in environmental modelling studies and twenty-six datasets exhibiting different typical properties of environmental data. The main aim at this stage is to demonstrate the application of the proposed evaluation framework, rather than provide a definitive answer as to which of these algorithms has the best overall performance. Nevertheless, the results indicate interesting differences in the algorithms' performance that have not been identified previously. [ABSTRACT FROM AUTHOR]

This paper presents a novel methodology to perform adaptive Water Demand Forecasting (WDF) for up to 24 h ahead with the aim to support near real-time operational management of smart Water Distribution Systems (WDSs). The novel WDF methodology is exclusively based on the analysis of water demand time series (i.e., demand signals) and makes use of Evolutionary Artificial Neural Networks (EANNs). It is implemented in a fully automated, data-driven and self-learning Demand Forecasting System (DFS) that is readily transferable to practice. The main characteristics of the DFS are: (a) continuous adaptability to ever changing water demand patterns and (b) generic and seamless applicability to different demand signals. The DFS enables applying two alternative WDF approaches. In the first approach, multiple EANN models are used in parallel to separately forecast demands for different hours of the day. In the second approach, a single EANN model with a fixed forecast horizon (i.e., 1 h) is used in a recursive fashion to forecast demands. Both approaches have been tested and verified on a real-life WDS in the United Kingdom (UK). The results obtained illustrate that, regardless of the WDF approach used, the novel methodology allows accurate forecasts to be generated thereby demonstrating the potential to yield substantial improvements to the state-of-the-art in near real-time WDS management. The results obtained also demonstrate that the multiple-EANN-models approach slightly outperforms the single-EANN-model approach in terms of WDF accuracy. The single-EANN-model approach, however, still enables achieving good WDF performance and may be a preferred option in engineering practice as it is easier to setup/implement. [ABSTRACT FROM AUTHOR]

This paper compares one global parametric land use change model, the artificial neural network – based Land Transformation Model, with two local non-parametric models: a classification and regression tree and multivariate adaptive regression spline model. We parameterized these three models with identical data from different regions of the world; one region undergoing extensive agricultural expansion (East Africa), another region where forests are re-growing (Muskegon River Watershed in the United States), and a third region where urbanization is prominent (South-Eastern Wisconsin in the United States). Independent training data and testing data were used to train and calibrate each model, respectively. Comparisons of simulated maps from the three kinds of land use change patterns were made using conventional goodness-of-fit metrics in land use change science. The results of temporal and spatial comparison of the data mining models show that the artificial neural network outperformed all other models in a short-time interval (East Africa; 5 years) and for coarse resolution data (East Africa; 1 km); however, the three data mining models obtained similar accuracies in a long-time interval (Muskegon River Watershed; 20 years) and for fine resolution data with large numbers of cells (Muskegon River Watershed; 30 m). Furthermore, the results showed that the probability of agriculture gain was more likely in locations closer to towns and large cities in East Africa, urbanization was more likely in locations closer to roads and urban areas in South-Eastern Wisconsin and the probability of forest gain was more likely in locations closer to the forest and shrub land cover and farther away from roads in Muskegon River Watershed. [ABSTRACT FROM AUTHOR]

Recent breakthroughs in artificial intelligence (AI), and particularly in deep learning (DL), have created tremendous excitement and opportunities in the earth and environmental sciences communities. To leverage these new 'data-driven' technologies, however, one needs to understand the fundamental concepts that give rise to DL and how they differ from 'process-based', mechanistic modelling. This paper revisits those fundamentals and addresses 10 questions that might be posed by earth and environmental scientists, and with the aid of a real-world modelling experiment, it explains some critical, but often ignored, issues DL may face in practice. The overarching objective is to contribute to a future of AI-assisted earth and environmental sciences where AI models can (1) embrace the typically ignored knowledge base available, (2) function credibly in 'true' out-of-sample prediction, and (3) handle non-stationarity in earth and environmental systems. Comparing and contrasting earth and environmental problems with prominent AI applications, such as playing chess and trading in stock markets, provides critical insights for better directing future research in this field. • DL is rooted in connectionism, hyper-flexibility, and vigorous optimization , which are alien to process-based modelling. • A knowledge base is essential to enable predictions of complex, open , partially observable , and non-stationary systems. • Bridging DL and process-based modelling is embryonic but has great potential in an age of big data and computational power. [ABSTRACT FROM AUTHOR]