Research on Strong Constraint Self-training Algorithm and Applied to Remote Sensing Image Classification

The remote sensing image classification is the key to remote sensing applications. Due to the large geographic area with a high temporal frequency of remote sensing image. It is difficult to use the conventional neural network to learn the representations of remote sensing images well under a condition of having an imbalance or few data. Meanwhile, the remote sensing image is easy to get but hard to label. The previous work for improving the accuracy of remote sensing image classification involved improving the structure of the Convolutional Neural Network (CNN). However, a semi-supervised learning algorithm proposed in this paper named Strong Constraint Self-training (SCS) considers the CNN classifier as a unit and ignores the detailed structure of the network. Initially, SCS uses a labeled dataset to train a base classifier. Secondly, the trained classifier to label unlabeled data and the data with pseudo labels join in the labeled dataset jointly train the next classifier. Lastly, the next classifier changes its role to be based classifier continues to train the next classifier. During this process, a threshold value is a gate to filtering the data with pseudo labels through the confidence coefficient that is given by the classifier. Furthermore, the value of the threshold decreasing as the process goes on. The key to this algorithm is to choose to transform learning to train a considerable base classifier. Extensive experiments are done in this paper. On a 50% ratio of AID training data and the NWPU-RESISC45 as an unlabeled dataset, the proposed algorithm achieves 96.01% on the rest of the AID data. On a 20% ratio of NWPU-RESISC45 data and the AID as an unlabeled dataset, this algorithm achieves 93.03% on the rest of NWPURESISC45 data.