Research On Deep-Learning Based Hyperspectral Image Classification Without Information Leakage

With the rapid development of sensor technology,hyperspectral sensors can collect hundreds of spectral bands.This abundant spectral and spatial information in hyperspectral remote sensing data has been widely used in a broad range of applications include agriculture,environmental science,physics and mineralogy.Most of these applications rely on the accurate classification of each pixel.However,it is costly and labor intensive to generate annotations for HSI due to the wide variety of sensors used.The current available HSI benchmark datasets only include far fewer labeled pixels in the whole image.Although some researchers have put forward some solutions in the past decades,these methods do not completely exclude the areas in the test set during the data set partition,resulting in the problem of leakage of test information,which makes the current classification accuracy rate too optimistic.In view of the existing data leakage problems,this paper proposes the following two solutions:(1)We design a novel spectral–spatial 3-D fully convolutional network(SS3FCN)to jointly explore the spectral–spatial information and the semantic information.SS3 FCN takes small patches of original HSI as inputs and produces the corresponding sized outputs,which enhances the utilization rate of the scarce labeled images and boosts the classification accuracy.In addition,to avoid the potential information leakage and make a fair comparison,we introduce a new principle to generate classification benchmarks.Experimental results on three popular benchmark datasets,including Salinas Valley(SV),Pavia University(PU)and Indian Pines(IP),demonstrate that the SS3 FCN outperforms state-of-the-art methods.(2)A novel data partitioning scheme and a triple-attention parallel network(TAP-Net)are designed to enhance the performance of HSI classification without information leakage.The dataset partitioning strategy is simple yet effective to avoid overfitting,and allows fair comparison of various algorithms.In contrast to classical encoder–decoder models,the proposed TAP-Net utilizes parallel subnetworks with the same spatial resolution and repeatedly reuses high-level feature maps of preceding subnetworks to refine the segmentation map.In addition,a channel-spectral-spatial-attention module is proposed to optimize the information transmission between different subnetworks.Experiments were conducted on three benchmark hyperspectral datasets,and the results demonstrate that the proposed method outperforms stateof-the-art methods.