Ground-based cloud detection using mixture losses

Recently, neural networks dominate various aspects of the field of ground-based cloud detection, serving as an advanced and robust method for cloud observation. Since loss functions are crucial for optimizing neural networks, the existing networks with advanced loss functions are also a vital area. This paper explores the impact of three distinct loss functions and their mixture effects on the optimization of network performance and gives strategies for selecting loss functions. To ascertain the efficacy of various loss functions, we carried out a comprehensive set of experiments on a dataset dedicated to ground-based cloud detection. Our findings indicate that employing a mixture of loss functions significantly enhances the training process for models focused on ground-based cloud identification.