Deep-learning (DL) based classification methods have been successfully used for hyperspectral image classification in recent years. Among various DL-based methods, convolutional neural network (CNN) has attracted a lot of attention. However, limited number of samples restricts the DL-based methods for widespread application. To deal with this problem, we propose a classification framework that can be transfer-learned between hyperspectral data with different number of bands. First, band selection is conducted to retain same number of bands for imagery of different hyperspectral sensors. Second, we simplify the typical 1D-CNN architecture by removing max-pooling layers. Third, modified CNN is trained on a source data, and this pretrained CNN is then fine-tuned with the target data. In the experiment, we pretrain the proposed network using the Indian Pines scene, and then fine-tune parameters to classify pixels in the Botswana scene. According to classification results, this proposed method obtains the highest overall accuracy, compared to KNN, SVM and its corresponding original 1D-CNN model, and even spend the least time training. Therefore, it can be concluded that this proposed method indicate transfer learning can be used between different hyperspectral images, and be helpful to improve the classification efficiency.
|