On the surface of our planet earth, there are numerous highly complex physical, chemical and biological processes, which are of great interest to both scientific research and socio-economic decision-making. Since the industrial revolution, researchers in various sectors have been accumulating data from the oceans and atmosphere. Effective data analytics is important to understand the complex processes and unveil unknown mechanisms, such as climate change, extreme atmospheric events e.g. tropical cyclones, variations of ocean circulation and mesoscale eddies, changes in the marine ecosystem and fishery resources and terrigenous materials transported across the continental shelf.
The last few decades have witnessed successful models to simulate and predict various processes in the climate system, including oceanic mesoscale eddies, ENSO or other climatological oscillators, as well as fluid hydrodynamic models for haze and marine pollutants. While the models are hypothesized based on known laws of physics, to tune/identify model parameters to fit practical observations is often difficult. At the same time, large volumes of observations accumulate rapidly and often from new measurement techniques. The amount of effort required to perform classical modeling and hypothesis tests in the era of big data becomes colossal.
The recent advances of deep neural networks-powered, machine intelligence techniques provide a hopeful solution with the promise of end-to-end data modeling, which allows researchers to avoid hypothesizing the possible low-level relationships and concentrate on higher-level trends. On the other hand, the most effective neural network models deal with complex signals, such as images or natural languages, by exploiting the underlying structures in the signals and assuming the structures are static and Euclidean. E.g. image pixels are organized as a regular 2D-grid. To effectively deal with data generated in real-world physical processes, the next-generation data models need to account for the dynamics in the underlying structures.
The aim of this Research Topic is three-fold. First, we encourage the development of deep neural networks (DNN) tailored for geology, ocean and atmosphere (GOA) data analytics. Second, we invite innovative interdisciplinary studies of effective DNN techniques in GOA data. Third, we want to motivate new tasks and problems related to GOA data analytics, which are of social significance and academic value. The three goals are specified in the scope of contribution below.
Novel neural network-based GOA data models
1. New neural network architectures and learning algorithms to deal with GOA signals of physical/chemical/biological processes.
Novel methods to address special aspects of GOA data applications
2. Reliable methods of learning large-scale models from limited observations.
3. Modelling and learning techniques specialized for interpretability to (1) identify meaningful structures in GOA processes (2) reveal causality and hidden mechanisms governing the processes (3) forecast the evolution of the processes.
4. Models of data with variable temporal scales and spatial resolutions.
5. Semi-supervised/self-learning methods that enable to encode domain knowledge.
New challenges
6. Proposal or review of challenges in GOA data analytics for which the new generation of neural network models provide potential solutions.
On the surface of our planet earth, there are numerous highly complex physical, chemical and biological processes, which are of great interest to both scientific research and socio-economic decision-making. Since the industrial revolution, researchers in various sectors have been accumulating data from the oceans and atmosphere. Effective data analytics is important to understand the complex processes and unveil unknown mechanisms, such as climate change, extreme atmospheric events e.g. tropical cyclones, variations of ocean circulation and mesoscale eddies, changes in the marine ecosystem and fishery resources and terrigenous materials transported across the continental shelf.
The last few decades have witnessed successful models to simulate and predict various processes in the climate system, including oceanic mesoscale eddies, ENSO or other climatological oscillators, as well as fluid hydrodynamic models for haze and marine pollutants. While the models are hypothesized based on known laws of physics, to tune/identify model parameters to fit practical observations is often difficult. At the same time, large volumes of observations accumulate rapidly and often from new measurement techniques. The amount of effort required to perform classical modeling and hypothesis tests in the era of big data becomes colossal.
The recent advances of deep neural networks-powered, machine intelligence techniques provide a hopeful solution with the promise of end-to-end data modeling, which allows researchers to avoid hypothesizing the possible low-level relationships and concentrate on higher-level trends. On the other hand, the most effective neural network models deal with complex signals, such as images or natural languages, by exploiting the underlying structures in the signals and assuming the structures are static and Euclidean. E.g. image pixels are organized as a regular 2D-grid. To effectively deal with data generated in real-world physical processes, the next-generation data models need to account for the dynamics in the underlying structures.
The aim of this Research Topic is three-fold. First, we encourage the development of deep neural networks (DNN) tailored for geology, ocean and atmosphere (GOA) data analytics. Second, we invite innovative interdisciplinary studies of effective DNN techniques in GOA data. Third, we want to motivate new tasks and problems related to GOA data analytics, which are of social significance and academic value. The three goals are specified in the scope of contribution below.
Novel neural network-based GOA data models
1. New neural network architectures and learning algorithms to deal with GOA signals of physical/chemical/biological processes.
Novel methods to address special aspects of GOA data applications
2. Reliable methods of learning large-scale models from limited observations.
3. Modelling and learning techniques specialized for interpretability to (1) identify meaningful structures in GOA processes (2) reveal causality and hidden mechanisms governing the processes (3) forecast the evolution of the processes.
4. Models of data with variable temporal scales and spatial resolutions.
5. Semi-supervised/self-learning methods that enable to encode domain knowledge.
New challenges
6. Proposal or review of challenges in GOA data analytics for which the new generation of neural network models provide potential solutions.