Progressive Non-Local Aggregation Network for Remote Sensing Image Scene Classification

Remote sensing image scene classification (RSISC) as an interesting research direction in the field of remote sensing image interpretation for its broad applications. Recently, the non-local attention mechanism has been demonstrated effectiveness in computer vision and remote sensing image understanding. However, most non-local attention-based methods use a further projection on the same feature map can only bring limited performance gain. In addition, they are usually suffered from expensive computation costs. To solve the above problems, we present a novel progressive non-local aggregation network (PNLANet), which can progressively aggregate long-range feature dependencies from multi-scale feature maps effectively and efficiently. Firstly, a lightweight cross-layer non-local (L-CL-NL) module is designed to compute the long-range dependencies between different convolutional feature maps. Then, this module is recurrently utilized to progressively acquire global contextual scene representation from multi-scale feature maps. The effectiveness of PNLANet is validated on two challenging benchmarks, i.e., AID and NWPU-RESISC45. The results show that it can obtain richer global understanding ability with cheaper computation costs as well as improve classification performance.