MULTI-TEMPORAL LAND COVER CLASSIFICATION WITH LONG SHORT-TERM MEMORY NEURAL NETWORKS

Global evapotranspiration (ET) can be estimated using machine learning (ML) models optimized on local data and applied to global data. This study explores whether sequential neural networks, which consider past data, perform better than models that do not. The findings show that sequential models struggle with global upscaling, likely due to their sensitivity to data shifts from local to global scales. To improve ML-based upscaling, additional data or integration of physical knowledge is needed.

We use an innovative approach to study the Earth's water cycle by blending advanced computer learning techniques with a traditional water cycle model. We developed a model that learns from meteorological data, with a special focus on understanding how vegetation influences water movement. Our model closely aligns with real-world observations, yet there are areas that need improvement. This study opens up new possibilities to better understand the water cycle and its interactions with vegetation.

We present a physics-aware machine learning model of the global hydrological cycle. As the model uses neural networks under the hood, the simulations of the water cycle are learned from data, and yet they are informed and constrained by physical knowledge. The simulated patterns lie within the range of existing hydrological models and are plausible. The hybrid modeling approach has the potential to tackle key environmental questions from a novel perspective.