Stack Attention-Pruning Aggregates Multiscale Graph Convolution Networks for Hyperspectral Remote Sensing Image Classification

The hyperspectral remote sensing images are classified by traditional neural networks methods can achieve promising performance, but only operate on regular square regions with fixed. This will lead to between neighborhood pixels have limitations in achieve long-distances joint interaction modeling and cross-spacetime information flow for capturing complex spatial-temporal dependencies. Meanwhile ignoring importance detail information and improved utilization of irrelevant information. In the work, we propose a stack attention-pruning multiscale aggregates graph convolution framework (SAP-MAGACN). The framework can automatically learn and selectively attend to the relevant subspace structure by stack attention-pruning module, can effectively disentangle the complex space structure of remote sensing images and capture the rich structural semantics. Meanwhile a refine graph of neighborhood pixels are constructures. Then we adopt the aggregation manner for multiscale graph convolution of pixels nodes in different neighborhood for effective long-range joint interaction modeling. Finally, we leverage dense cross-spacetime edges to completion propagation of multiscale spatial-temporal information, and gradually produce the discriminative embedded features and effectively distinguish the categories of boundary pixels. The experimental results shown the propose SAP-MAGCN outperformance all others state-of-the-art methods on Indian Pines and Salinas public benchmark datasets. Such as the OA, AA and Kappa of our propose SAP-MAGCN frameworks is 96.75%, 95.73% and 97.33%, respectively, on Indian Pines datasets.