Improved Metric Learning With the CNN for Very-High-Resolution Remote Sensing Image Classification

The number of labeled samples has a great impact on the classification results of a very-high-resolution (VHR) remote sensing image. However, the acquisition of available labeled samples is difficult and time consuming. Faced with the limited labeled samples on a high-resolution remote sensing image, a semisupervised method becomes an effective way. In semisupervised learning, an accurate similarity prediction between unlabeled and labeled samples is very important. However, a reliable similarity prediction between high-dimensional features is difficult. For more reliable similarity prediction for the high-dimensional feature, a novel semisupervised classification framework via improved metric learning with a convolutional neural network is proposed. In the proposed method, a novel trainable metric learning network is designed to accurately evaluate the similarity between high-dimensional features. The vector distance parameter solving problem is transformed into a neural network design problem, which can automatically calculate parameters by the back-propagation algorithm. Finally, the pixel constraint mechanism is introduced to select the unlabeled samples. Experimental results conducted on three VHR remote sensing images, including Aerial, Xi'an, and Pavia University, and the results present that the proposed method performs better than the compared state-of-the-art methods.

Automated summary

This study proposes a novel semisupervised classification framework utilizing improved metric learning and convolutional neural network to enhance the accuracy of similarity prediction between high-dimensional features. Experimental results demonstrate that the proposed method outperforms the current state-of-the-art methods on three very-high-resolution remote sensing images.