Application of a Hyper-Parameter Optimization Algorithm using Mars Surrogate for Deep Polsar Image Classification Models

Guangyuan Liu, Yangyang Li, Licheng Jiao

Abstract: Stacked auto-encoder with weight decay (SAE-WD) and convolutional neural network (CNN) have great performances in PolSAR image classification. But the performances of them highly depend on proper hyperparameter configurations. In this paper, we apply a hyperparameter algorithm (called MARSAOP previously proposed by us) to automatically find good hyper-parameter configurations for them. The obtained results on two real PolSAR images suggest MARSAOP could still perform well on deep learning models and save manpower to tune the hyper-parameters.