A Supervised Approach for Estimating Layered Cloud Motion Vector and Clearsky Indices from Satellite Images

The rapid growth of solar energy rises challenges due to its inherent intermittency, significantly driven by cloud dynamics. While approaches based on satellite images and cloud motion vectors (CMV) are considered as a reference for intra-day forecasting, they do not account for the vertical structure of the atmosphere. Such approaches treat clouds as a single two-dimensional layer, leading to significant errors during events where multiple cloud layers move at different altitudes and speeds. Numerical Weather Prediction models, on the other hand, account for the three-dimensional structure of clouds, but their spatial and temporal resolution limits their performance.

In parallel, recent research has leveraged Deep Learning (DL) architectures for forecasting solar irradiance from satellite images. For instance, Convolutional Neural Networks, including the UNet-based approach [1], have been used to forecast satellite images. While such approaches yield promising results and can theoretically capture cloud formation and dissipation mechanics, their accuracy remains constrained when relying solely on satellite images without supplementary meteorological data, notably the vertical structure of clouds.

In this work, we propose a layer segmentation approach to improve the physical consistency of cloud motion estimation. Motivated by the distinct behaviors of driving parameters across different altitudes, we reconstruct three atmospheric layers from consecutive satellite observations (kc). This methodology follows the direction of recent studies [2], in which cloud phases were separated and overlapping CMV were calculated using cloud properties and infrared temperatures. 

Our approach leverage the apparent motion of clouds on image sequences for estimating the layered cloud motion vectors. The proposed method relies on a UNet-based architecture designed for the joint task of layer semgentation and layered cloud motion vector estimation. A sequence of four consecutive satellite images along with meteorological data, including kc forecasts and wind vectors at different pressure levels, is taken as input. From these variables, three distinct kc maps, each corresponding to a specific atmospheric layer (low, medium, and high altitude), are generated. The training process is supervised by comparing the reconstructed global kc against the ground truth from the subsequent satellite observation. This approach ensures that while the model learns the complex dynamics of individual layers, it remain constrained by satellite observation. 

 

REFERENCES

[1] Nils Straub, Steffen Karalus, Wiebke Herzberg, and Elke Lorenz. Satellite-based solar irradiance forecasting: Replacing cloud motion vectors by deep learning. Solar RRL, 8(24):2400475, 2024. doi: https://doi.org/10.1002/solr. 202400475. 
[2] Cuiping Liu, Wei Han, Feng Zhang, Jiaqi Jin, Qiong Wu, Wenwen Li, and Chloe Yuchao Gao. Deriving overlapped cloud motion vectors based on geostationary satellite and its application on monitoring typhoon mulan. Geophysical Research Letters, 52(13):e2025GL116397, 2025. doi: https://doi.org/10.1029/2025GL116397.