Unsupervised deep learning models for aerosol layers segmentation

Abstract

Identification of atmospheric particles and aerosols is a very important topic in climatology. However, before classification, the various homogenous layers of aerosols need to be segmented. In this paper we present an initial work towards the development of an automated segmentation system for aerosols. Provided that there are no annotated datasets available for this task, we approach the problem using unsupervised machine learning techniques. Several machine learning (ML) models, previously used in other similar segmentation tasks, have been trained for the purpose of identifying various types of aerosols based on the input data. Initial model performance showed unsatisfactory results and thus several adjustments were made to fit our requirements. The ML models for aerosol segmentation have been evaluated objectively, only in terms of reconstruction efficiency, more precisely, how well does the model recreate the input data. Since there is no annotated dataset (neither for training, nor for evaluation), the segmentation efficiency of the models was not evaluated objectively. Consequently, the segmentation results have been evaluated by a human expert.