Deep learning and process understanding for data-driven Earth system science
Markus Reichstein et al 2019 in Nature
This is a review article that discusses the applications and challenges of machine learning in geosciences. The Earth system science has entered the era of big data. Earth system data is a typical big data, characterized by the four V’s: volume, velocity, variety, and veracity, such as various remote sensing, point observations, and model data. The current challenge is how to extract and interpret information from this big data, as the speed of information collection far exceeds our ability to digest it. The increase in data has not led to improved predictive capability of systems; we need to understand the data. In this context, machine learning presents a great opportunity.
The article elaborates on several aspects:
(1) Cutting-edge Machine Learning in Geosciences. Methods such as neural networks and random forests have long been applied to classification, change detection, and soil mapping problems in geosciences. However, these applications are spatially focused and relatively static over time, while the Earth is constantly changing. Machine learning regression methods have advantages in temporal dynamics, such as artificial neural networks with hidden layers, which can predict the changes in carbon flux over time and space. However, these applications also have issues that need attention, such as extrapolation capability, sampling or data bias, neglecting confounding factors, and the distinction between statistical correlation and causation. Classical machine learning methods require prior knowledge to determine certain spatiotemporal features and cannot automatically explore the spatiotemporal characteristics of the data. Some spatiotemporal dynamic features, such as ‘memory effects’, can be manually added as features to traditional machine learning, but the latest deep learning has no such limitations.
(2) Opportunities for Deep Learning in Earth System Science. Deep learning has been widely applied in other fields, but its application in geosciences is still in its infancy. Some studies have shown that deep learning can effectively extract spatiotemporal features, such as extreme weather events, with minimal human intervention. It can also be used for automatic extraction of urban change from remote sensing. Deep learning methods are typically divided into spatial learning (e.g., convolutional neural networks for object classification) and sequence learning (e.g., speech recognition), but these two are gradually merging and can be applied to video and action recognition problems. These problems are similar to the multidimensional structures that change over time in geosciences, such as organized precipitation convection and vegetation status. Although there is great potential for application, the application of deep learning to spatiotemporal changes in atmospheric and oceanic transport or vegetation dynamics still needs development.
(3) Challenges of Deep Learning in Earth System Science. Although the target objects of traditional deep learning and geoscience phenomena have many similarities, there are also significant differences. For example, hyperspectral and multi-band data are much more complex than RGB-based computer image recognition, and there are also noisy and missing satellite data. Additionally, the combination of spectral, temporal, and spatial dimensions can pose computational challenges. In computer images, there are abundant labeled training samples for recognizing ‘dogs’ and ‘cats’, but in geosciences, there are no similar large labeled training samples, such as for droughts. Externally, the authors summarize five major challenges: interpretability, physical consistency, complexity and determinism of data, lack of labeled samples, and computational demands. If these challenges can be addressed, deep learning will bring about significant changes in geosciences. The most promising recent application is ‘nowcasting’, with the future being long-term forecasting. The authors believe that deep learning will soon become the primary method for classification and spatiotemporal prediction problems in geosciences.
(4) Integration with Physical Modeling. Physical modeling (theory-driven) and machine learning modeling (data-driven) have often been considered two separate fields with different paradigms. However, the two methods can actually complement each other; the former has strong extrapolation capabilities, while the latter is more flexible and can discover new patterns. The authors propose several potential points for the combination of the two methods: improving parameterization, using machine learning to ‘replace’ sub-models in physical models, analyzing mismatches between models and observations, constraining sub-models, and substituting models or simulations.
(5) Promoting Scientific Development. Machine learning methods undoubtedly bring significant improvements to classification and prediction problems. Data-driven machine learning methods can also mine new information from data that was previously unknown, thereby promoting the emergence of new mechanisms and understandings.
(6) Conclusion. In the era of big data in Earth sciences, machine learning is undoubtedly useful, but there are also application challenges. The authors propose four recommendations: identify the specificity of the data, ensure the rationality and interpretability of inferences, estimate uncertainty, and validate against complex physical models. In the future, process models and machine learning will further integrate. Data-driven machine learning will not replace physical models, but will serve to complement and enrich them, ultimately achieving hybrid modeling.
The first author, Markus Reichstein, is a prominent figure, affiliated with 1 Department of Biogeochemical Integration, Max Planck Institute for Biogeochemistry, Jena, Germany. 2 Michael-Stifel-Center Jena for Data-driven and Simulation Science, Jena, Germany.