Attention-based Deep Feature Learning Network for Scene Classification of Hyperspectral Images

Existing scene classification methods are mainly based on high spatial resolution (HSR) images, which contain limited spectral information, and some complex scenes with similar visual perception (e.g., texture, shape and color), are prone to be misclassified. Hyperspectral images contain detailed spectral information that can realize fine classification of ground objects by mining subtle spectral differences, but existing data is mainly oriented to pixel-wise interpretation. Based on the above facts, we introduce a hyperspectral data set for remote sensing scene classification (HSRS-SC), which contains 1385 images divided into 5 semantic classes. Images in the HSRS-SC data set possess a spatial resolution of 1 meter and a wavelength ranging from visible light to near-infrared, to reflect the rich spatial and spectral features of ground objects. To combine the above two features, an attention-based deep feature learning network (ADFLN) is developed for scene classification of hyperspectral images. In the proposed ADFLN, spatial-spectral hybrid attention mechanism is adopted to highlight key discriminative information by using ResNet18 as backbone. Experiments performed on the HSRS-SC show that hyperspectral images achieve better classification results than RGB images. Furthermore, by introducing hybrid attention mechanism, overall classification accuracy can be further improved.

I. Introduction

Remote sensing images possess tremendous potential in various applications, such as land management, hazard assessment, national defense, and intelligent agriculture, and extracting crucial information from images has become an important topic [1]– [3]. Scene classification aims to assign a semantic label for each queried images, and it can provide scene-level information for downstream tasks [4]– [6]. Scene images always exhibit complex spatial patterns, which leads to within-class diversity and between-class similarity [7]– [9]. To improve classification performance, various methods based on scene representations have been proposed, including hand-crafted feature-based methods and deep learning methods. Although hand-crafted features (e.g., texture, color, structure, and shape) possess the ability to describe simple scenes for classification, it is a huge challenge to effectively extract high-level semantic information for complex scenes because of its limited description ability, which restricts the improvement of performance [10], [11].