A learning rate optimization method for model convergence of hyperspectral image classification

In recent years, many convolutional neural network (CNN) based works have demonstrated the great potential and more possibilities of CNN in improving HCIS performance. In addition to numerous studies on network models, many data preprocessing methods have been proposed. But few people think about the problem from other perspectives than the two mentioned above. Deep learning methods have the following basic steps: data preprocessing, building neural network models, and training and evaluation of the models. There are often drastic fluctuations in loss and accuracy during actual model training, which affects the training efficiency of the model. To mitigate this problem, a new Adam-based learning rate optimization method is proposed: the discount factor method abbreviated as Adam-DF, which takes inspiration from game theory and corrects the learning rate to some extent by the effect of the previous model training, so that the model can be trained better in the next parameter update using the gradient descent algorithm. A comparison with some other methods in the experiments confirms that this method can make the fluctuation of loss and accuracy in the training of the model significantly alleviate and reach the fitting state faster.