87 Can a Deep Learning Algorithm improve the Automated Classification of Cloud Imager Particle Data?

Monday, 9 July 2018
Regency A/B/C (Hyatt Regency Vancouver)
Annika Lauber, ETH, Z├╝rich, Switzerland; and G. Touloupas, J. Henneberger, A. Beck, A. Lucchi, and U. lohmann

Cloud particle imagers provide two-dimensional images of single cloud particles. These images are often sorted into multiple classes to sort out measurement artefacts, distinguish between circular liquid droplets and non-circular ice crystals or differentiate between ice crystal habits. To handle the millions of particles observed, an automated classification method is needed. Many of the automated classification methods use supervised machine learning approaches such as classification trees, which create a model from a large number of examples, hand-labelled by humans. Because of the large variety in particle shapes, weather conditions and instrument versions, such models usually do not generalize well to new data and the time-consuming hand-labelling step has to be repeated, sometimes for each measurement day. We found that a deep learning approach, using a deep convolutional neural network, which has significantly improved the state of the art in computer vision tasks, can achieve a more generalized classification of cloud particles.

We trained a convolutional neural network with particle images from the holographic imager HOLIMO using the TensorFlow deep learning framework. The image data was directly used as the input of the neural network, which makes feature extraction from the images (e.g. size, asphericity, signal-to-noise ratio) redundant compared to the baseline classification tree approach. In our case, we use the amplitude and phase images scaled to 32x32 pixels. The network consists of a standard convolutional architecture with added optimizations to classify the particles into circular water droplets, non-circular ice crystal, and instrument artefacts. Compared to the classification tree, the deep learning approach does not need feature engineering, increased the per-class and overall accuracies, and generalizes better to new datasets. The deep learning approach can be easily used for ice crystal habits analysis and has the potential to improve also the automated classification of other cloud imager probes.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner