7 February, 2020 (Reichstein, Nature, Deep learning)
Reichstein, M., Camps-Valls, G., Stevens, B. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566, 195–204 (2019). https://doi.org/10.1038/s41586-019-0912-1
- How have machine learning algorithms, especially deep learning, been applied to Earth system science and what is the future direction of the application?
- Our ability to produce a deluge of data outpaces our ability to assimilate the information.
- The advance in computing system, statistical and machine learning modeling have combined with the plethora of Earth system data to create a new avenue of modeling: deep learning.
- Machine learning approach (such as deep learning) have successfully been applied to many problems in Earth system science.
- Machine learning and the physical models can complement each other, creating a new way of modeling: hybrid modeling.
- Strengths of the hybrid modeling
- Improving parameterizations: finding optimal values and pertinent patterns (e.g. using ML-derived parameters instead of the pre-defined PFTs).
- Replacing a physical submodel; many submodels are of semi-empirical, where the functional form has little theoretical basis.
- Analysis of model-observation mismatch: ML can help to identify, visualize, and understand the patterns of model error, which allows us also to correct model outputs.
- Constraining submodels
- Surrogate modeling: ML can be used for faster simulation.
- Deep learning and other ML algorithms have promising potential to be used for Earth system science.
- Hybrid modeling, which integrates physical- and data-driven modeling, is recommended.
- Bayesian framework is effective in quantifying uncertainties or in tracing the propagation of them.
- I think that Reichstein is a great writer.
- deluge, soothsaying, pertinent, notwithstanding, yet, tedious, exemplify, ad hoc, circumvent,