Hyperspectral Classification Based on Spectral Indices Learned Through Soft Attention Units

Classification of hyperspectral scenes has been dominated in recent years by Convolutional Neural Networks (CNNs). Spectral and spatial convolutions have proven to be very effective in learning discriminative representations. The power of CNNs however is limited by their small receptive fields. Training deeper CNNs with wider receptive fields is a very difficult task with regard to the small amount of available training samples. Furthermore, CNNs seem to fail in capturing very local spectral features such as absorption peaks. In the present paper, we introduce a new paradigm inspired by physics-based hyperspectral indices and attention mechanisms. Our model learns spectral indices by focusing on specific spectral bands through soft attention units. It achieves high or better overall accuracy and kappa score than state-of-the-art CNNs on Pavia University, Kennedy Space Center and our own real-world dataset while dramatically reducing the number of parameters needed and increasing interpretability.