Abstract
Since hyperspectral images (HSI) captured by different sensors often contain different number of bands, but most of the convolutional neural networks (CNN) require a fixed-size input, the generalization capability of deep CNNs to use heterogeneous input to achieve better classification performance has become a research focus. For classification tasks with limited labeled samples, the training strategy of feeding CNNs with sample-pairs instead of single sample has proven to be an efficient approach. Following this strategy, we propose a Siamese CNN with three-dimensional (3D) adaptive spatial-spectral pyramid pooling (ASSP) layer, called ASSP-SCNN, that takes as input 3D sample-pair with varying size and can easily be transferred to another HSI dataset regardless of the number of spectral bands. The 3D ASSP layer can also extract different levels of 3D information to improve the classification performance of the equipped CNN. To evaluate the classification and generalization performance of ASSP-SCNN, our experiments consist of two parts: the experiments of ASSP-SCNN without pre-training and the experiments of ASSP-SCNN-based transfer learning framework. Experimental results on three HSI datasets demonstrate that both ASSP-SCNN without pre-training and transfer learning based on ASSP-SCNN achieve higher classification accuracies than several state-of-the-art CNN-based methods. Moreover, we also compare the performance of ASSP-SCNN on different transfer learning tasks, which further verifies that ASSP-SCNN has a strong generalization capability.
Highlights
Hyperspectral sensors typically acquire images across hundreds of spectral bands with rich spatial information, which makes hyperspectral images (HSI) become essential data sources to deal with the heterogeneous and mixed landscape
In the experiment with pre-training, we propose a transfer learning framework based on adaptive spatial-spectral pyramid pooling (ASSP)-Siamese Convolutional Neural Network (SCNN) for verifying the classification performance of pre-trained ASSP-SCNN on a source dataset after transferring to a target dataset
In [20], we proposed an end-to-end Siamese convolutional neural network (ES-convolutional neural networks (CNN)) for HSI multi-classification tasks based on spectral information
Summary
Hyperspectral sensors typically acquire images across hundreds of spectral bands with rich spatial information, which makes hyperspectral images (HSI) become essential data sources to deal with the heterogeneous and mixed landscape. Deep convolutional neural networks (CNN) have attracted much attention due to their excellent performance in hyperspectral image classification tasks [9,10,11,12,13] Such methods usually exhibit two common issues: (1) How to develop a high-precision deep learning model with limited labeled samples. We propose a Siamese CNN with 3D adaptive spatial-spectral pyramid pooling, called ASSP-SCNN, which can be transferred to heterologous HSI datasets without unifying the number of bands of resource datasets and target dataset. Different from the first Siamese CNN [28] and the ES-CNN [20], the proposed network can explore the potential of both spatial and spectral information for HSI classification and allow us to put three-dimensional samples with varying sizes or scales into the network.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have