Mind the Gap: Bridging Prior Shift in Realistic Few-shot Crop-type Classification

The Eurocrops v2.0 dataset contains several years of farmer crop declarations from 16 Member States for the Common Agricultural Policy. These data were harmonized with a crop taxonomy and then linked to other classification systems including agricultural statistics and remotely sensed products. This dataset can support researchers and policy makers in tracking changes over time, calculating crop rotations, for national comparisons, for evaluating official statistics and for remote sensing.

We introduce a new global model that links how water and carbon move through land ecosystems. By combining process knowledge with artificial intelligence that learns from observations, we model daily changes in vegetation, water and carbon cycle processes. This model outperforms both purely data-driven and traditional process models, especially in dry and tropical regions. This advance could improve current understanding of water–carbon cycle relationships.

This study evaluates machine learning approaches for upscaling evapotranspiration from the site to the global scale. Sequential models capture temporal dynamics better, especially with precipitation data, but all models show biases in data-scarce regions. Improved upscaling requires richer training data, informed covariate selection, and physical constraints to enhance robustness and reduce extrapolation errors.

We use an innovative approach to studying the Earth's water cycle by integrating advanced machine learning techniques with a traditional water cycle model. Our model is designed to learn from observational data, with a particular emphasis on understanding the influence of vegetation on water movement. By closely aligning with real-world observations, our model offers new possibilities for enhancing our understanding of the water cycle and its interactions with vegetation.