Efficient Implementation for Composite CNN-Based HSI Classification Algorithm with Huawei Ascend Framework

In recent years, deep learning has been merged as a powerful tool to effectively address nonlinear feature-extraction problems and is widely used in the field of hyperspectral image (HSI) classification. With the increased use of Unmanned Aerial Vehicles (UAVs) based HSI remote sensing, faster implementation of the HSI classification algorithm for UAVs can be deployed as aerial edge servers to facilitate edge computing service based remote sensing. In this paper, by fully using complex pixel- and superpixel-level interdependencies, we propose a composite neighbor inspired heterogeneous network which combines pixel-level CNN and superpixel-level based graph convolutional network. After that, we adopt Huawei's Ascend AI processor and Compute Architecture for Neural Networks (CANN) operator library and neural network information flow optimization, to develop an efficient implementation solution for the proposed HSI classification algorithm. By using the capabilities in tensor acceleration, runtime management, and task scheduling of the Huawei Ascend framework, four HSI classification algorithms are implemented and compared. Experiments on three real-world HSIs demonstrate the proposed method outperforms the compared mini-batch deep learning algorithms and have obtained the state-of-the-art performance. In addition, our studies show that the Huawei Ascend framework can greatly alleviate the memory occupation and insufficient accuracy of large-scale HSIs data during training, while Huawei's Ascend series chips have the potential ability for UAVs-based HSI remote sensing.