A Deep Multi-feature Based Transfer Learning Network For Hyperspectral Image Classification

Owing to the rapid development of remote sensing technology,hundreds of nearly continuous spectral bands and an enormous amount of spatial information can be captured simultaneously via the hyperspectral sensors.Simultaneously,hyperspectral image(HSI)has generated a new application field,HSI classification,which uses the spectral information of some pixels to train a classification model.Then the model can predict the labels of all the pixels on HSI.HSI classification has been widely utilized in diverse fields,such as precision agriculture,geological exploration,and environmental sciences.Traditional HSI classification methods used to apply a shallow classification model to label the unlabeled pixels.However,these artificial low-level features and classification parameters are more sensitive to local changes in the input data,which greatly reduces the classification accuracy.Deep neural network(DNN)has been proven to be able to automatically learn a hierarchical and abstract feature representation and recently attracted significant attention in the field of HSI classification.However,the existing DNN based HSI classification methods can not make full use of the spectral/spatial information,and these methods can only train and test on a specific HSI.Besides,the construction of an efficient DNN mostly relies on a large number of labeled samples being available.It is very time-consuming to label such a large number of samples manually in practice.To address above problems,this paper firstly proposes a deep active multi-feature fusion and classification network,which uses three hierarchical stacked sparse autoencoder(SSAE)networks to extract deep joint spectral-spatial feature on source HSI,which is fused by source deep spectral feature and spatial feature.Moreover,in order to increase the flexibility and scalability of the network with limited labeled samples,the active feature and sample transfer learning algorithm is proposed.Active learning is first exploited into the fine-tuning process of source deep active multi-feature fusion and classification network in order to use a few most informative samples to train the network effectively.Then,the algorithm transfers the pre-trained SSAE network and the limited training samples from the source domain to the target domain,where the target network is subsequently fine-tuned by the target training dataset which is updated by corresponding two active learning strategies.Experimental results on three popular datasets demonstrate that: 1)the learned deep joint spectral-spatial feature representation is more robust than many joint spectral-spatial feature representation;2)the network can be effectively trained using only limited labeled samples with the help of novel active learning strategies;3)the network is flexible and scalable enough to function across various transfer situations,including cross-dataset and intra-image.Besides,in order to further reduce the number of labeled samples required for training and transferring the network,an unsupervised deep feature based transfer learning algorithm is proposed.The pre-trained network that is transferred from the source domain to the target domain first learns the feature representation of the target training data,then the learned features are clustered by K-means clustering.Finally,the target training data and their clustering labels are used to fine tune the pre-trained network.The experimental results indicate that compared with the traditional clustering methods,the proposed unsupervised deep feature based transfer learning method can learn an improving feature representation for the target HSI without any target supervised information,and achieve a good classification accuracy.