Synthetic Cloud Height Prediction Using Stereo Matching and Deep Learning

Accurate cloud base height estimation holds paramount significance in aviation, weather applications, or solar irradiance nowcasting. Although effective, existing solutions, such as ceilometers, are cost-prohibitive and lack portability, limiting their utility for smaller airports, vertiports, and ad hoc operations. This paper introduces an innovative, cost-effective solution leveraging stereo camera setups combined with deep learning. Synthetic data is generated to test the model uti-lizing a virtual environment. Stereo-matching techniques obtained disparity maps from the captured image pairs. Building upon this data, we trained a model to transform these images into precise cloud height predictions. Our approach marries the affordability of stereo cameras with the computational prowess of deep learning, offering a promising low-cost alternative to traditional cloud height estimation instruments. Rather than 3D reconstruction methods, we focused on training a CNN model, using disparity maps of the captured images. Our CNN-based regression model predicts the cloud base height from the 500 m to 3500 m range, With the 77 test samples, we got an RMSE of 57.78 meters.