Discriminant analysis-based attention network for hyperspectral target detection

Abstract

Hyperspectral target detection is one of the main applications in remote sensing field, which has several challenges such as imbalance between target and background and how accurately separate targets from background. To deal with these difficulties, the discriminant analysis-based attention network (DAAN) is proposed in this work. To solve the insufficient number of targets, two autoencoder based data augmentation approaches, which are pixel-based and patch-based are suggested. While the augmented target pixels are used for feature space transformation for maximizing the between-class scatters and minimizing the within-class scatters, the augmented target patches are used as input of the network. To increase the separability among targets and background, a discriminant analysis method is introduced. An attention feature map is generated from the discriminant analysis for weighting the feature maps of the proposed network to highlight targets with respect to the background. DAAN uses the low-level features attended by the weight matrix in addition to the high-level features hierarchically extracted by several convolutional kernels. The experiments show high detection performance of DAAN in various hyperspectral images where there is no need to set the free parameters for each dataset.