This paper proposes an image-based visibility estimation method with deep label distribution learning. To train an accurate model for visibility estimation, it is important to obtain the precise ground truth for every image. However, the ground-truth visibility is difficult to be labeled due to its high ambiguity. To solve this problem, we associate a label distribution to each image. The label distribution contains all the possible visibilities with their probabilities. To learn from such annotation, we employ a CNN-RNN model for visibility-aware feature extraction and a conditional probability neural network for distribution prediction. Our experiment shows that labeling the image with visibility distribution can not only overcome the inaccurate annotation problem, but also boost the learning performance without the increase of training examples.

Classification of weather systems provides a simple description of atmospheric circulations and bridges the gap between large-scale atmospheric conditions and local-scale environmental variables. However, the existing classification methods are challenged due to lack of labels and inaccurate similarity measures between data samples. In this letter, we propose a self-supervised Spatiotemporal Contrastive Learning (SCL) framework for the classification of weather systems without manual labels. In particular, we operate both spatial and temporal augmentation on multivariate meteorological data to fully explore temporal context information and spatial stability in accordance with synoptic nature. With the classification results, we apply a statistical downscaling method based on analog forecasting for the assessment and comparison of classification results. The experimental results demonstrate that the proposed SCL model outperforms traditional classification methods.