Optimal sampling of spatial patterns improves deep learning-based early warning signals of critical transitions

Area of Science:

• Ecology
• Environmental Science
• Complex Systems Science

Background:

• Complex spatio-temporal systems, such as lakes and forests, can exhibit alternative stable states and undergo critical transitions.
• Identifying early warning signals for these transitions is crucial, but forecasting the transition type (sudden vs. smooth) remains a challenge.

Purpose of the Study:

• To develop an advanced machine learning (ML) toolkit, the Spatial Early Warning Signal Network (S-EWSNet), for early detection and type forecasting of spatio-temporal critical transitions.
• To overcome the "black box" nature of ML models by enabling the deep neural network to learn latent features characteristic of transitions.

Main Methods:

• Developed S-EWSNet, an ML toolkit utilizing a deep neural network trained with an optimal sampling strategy (OSS).
• Employed a stochastic cellular automata model for generating training data, incorporating OSS for spatial pattern analysis.
• Validated the toolkit on simulated and empirical datasets.

Main Results:

• S-EWSNet successfully provides early warning indicators for critical transitions in spatio-temporal systems.
• The toolkit can accurately detect the type of transition (sudden or smooth).
• The integration of OSS allows for better interpretability of the ML model's learned features.

Conclusions:

• S-EWSNet offers a novel approach to forecasting critical transitions in complex systems, addressing a significant gap in early warning signal research.
• The developed method enhances the understanding of transition dynamics by learning characteristic spatial patterns.
• This toolkit has potential applications in environmental management and ecological forecasting.