Remote sensing image classification algorithm based on ridge wave sparse collaborative representation convolutional neural network

Fengping An, Jun-e Liu

Abstract: Remote sensing image classification is a crucial link when processing remote sensing images. Through classification, remote sensing images are converted into classified features that can be understood and processed by computers running deep applications. However, the traditional remote sensing image classification methods do not meet the actual application requirements. In recent years, the rapid development of deep learning theory has provided a technical approach for solving remote sensing image classification. However, deep learning has the following problems when applied to remote sensing image classification: First, it is impossible to construct an activation function suitable for deep learning models that matches the characteristics of remote sensing images; second, the deep learning classification models have poor effects. In view of this, this paper first studies the activation function and constructs ridge waves with scale, displacement, and direction information. Because a ridge wave has good compact support characteristics, it can more closely approximate the high-dimensional nonlinear decision function and obtain a more effective activation function, thus solving the activation function problem in deep learning modeling. At the same time, to optimize the classifiers included in deep learning models, this paper proposes a sparse collaborative representation classifier that more fully combines the advantages of sparse representation classifiers and collaborative representation classifiers. It can yield the relationship between the competition and the collaboration of the remote sensing image to be classified and better use the characteristics of the remote sensing image, achieving better classification effect. Based on the above ideas, this paper proposes a remote sensing image classification algorithm based on a ridge wave sparse collaborative representation convolutional neural network. Finally, the method in this paper is verified by experiments on “UC Merced Land Use Dataset” and “RSSCN7 Dataset”. The results show that the average accuracy of the method proposed in this paper is significantly higher than that of machine learning and other deep learning methods, and the method in this paper has better stability and robustness.

External IDs:dblp:journals/mta/AnL21a