Graph Neural Network for Feature Extraction and Classification of Hyperspectral Remote Sensing Images

Remote Sensing (RS) is a non-contact earth observation technology that obtains ground target information through the reflection or radiation of electromagnetic waves on ground objects (SambhuNath in An introduction to remote sensing. Koros, 2014; Brink et al. in Introduction to remote sensing for conservation practitioners, 2018). Military remote sensing technology can use remote sensing carriers, e.g., satellites and unmanned aerial vehicles (UAVs), to carry out high-resolution, real-time, multi-angle and multi-frequency remote sensing observation and acquisition of enemy military targets. Then, the military intelligence of enemy and battlefield conditions can be monitored, which provides intelligence support and tactical guidance for military command departments. Therefore, this technology holds a pivotal position in modern warfare, and is of great significance in improving combat efficiency and reducing war costs. In the Russia–Ukraine conflict, both sides have realized the importance of high-precision remote sensing images in obtaining the military deployment trends of the other side, and they compete to use remote sensing technology to timely understand each other’s dynamics and develop corresponding strategies accordingly. At the same time, satellite remote sensing images have become the best tool for netizens and journalists to understand and analyze the conflict situation in Ukraine, enabling remote sensing technology to gain widespread attention in public opinion.

Hyperspectral images (HSIs) provide detailed spectral information through hundreds of (narrow) spectral channels, which can be used to accurately classify diverse materials of interest (Rasti et al. in IEEE Geosci Remote Sens 8(4):60–88, 2020; Zhong et al. in IEEE Trans Neural Netw Learn Syst 12:1–13, 2019). However, the increased dimensionality of such data provides a challenge to conventional techniques, and hyperspectral classification has great research value.

Hyperspectral imagery collected from satellite or airborne comprises hundreds of contiguous bands and contains abundant spectral-spatial information. Due to the advantages of HSI, land-cover categories can be distinguished at the pixel level.

Hyperspectral images (HSIs) collect rich spatial-spectral information in hundreds of spectral bands, which are captured by hyperspectral remote sensors, (Rasti et al., IEEE Geosci Remote Sens Mag 8(4):60–88, 2020; Ghamisi et al., IEEE Geosci Remote Sens Mag 5(4):37–78, 2017; Peng et al., IEEE Trans Geosci Remote Sens 57(2):1183–1194, 2018; Lu et al., IEEE Trans Geosci Remote Sens 56(4):2183–2195, 2018), which are more effective in distinguishing different land-covers compared with other multispectral (Zhao et al., IEEE Trans Neural Netw Learn Syst 30(11):3212–3232, 2019) or RGB (red, green, and blue) image (Hong et al., IEEE Geosci Remote Sens Mag 11:4051, 2021). Therefore, HSIs are widely employed in various applications, ranging from military reconnaissance, marine monitoring to disaster prevention and control (Hong et al., IEEE Geosci Remote Sens 9:16820, 2020; Ding et al., IEEE Geosci Remote Sens Lett 19:1–5, 2021). HSI classification (category each pixel into certain label) is a crucial technique for these applications. However, the complex noise effects and spectral variability (Hong et al., IEEE Trans Image Process 28(4):1923–1938, 2019), labeled training samples deficiency (Wan et al., IEEE Trans Geosci Remote Sens 59(1):597–612, 2021), and high spectral mixing between materials (Zhong et al., IEEE Trans Neural Netw Learn Syst 14:1–13, 2019) bring difficulties in extracting discriminative information from HSI for classification.

Hyperspectral images have been widely used in many remote sensing applications due to their rich spectral and spatial information, and up to now, a wide range of applications have been benefited from the development of this direction, including urban development, land monitoring, scene interpretation, and resource exploration 0. Among these applications, HSI classification is a common technique that facilitates the study of the chemical properties of scene materials remotely.

Hyperspectral image (HSI) is a branch of optical remote sensing data, which has advantages of high spectral resolution and detailed spatial structure (Zhang et al. in IEEE Trans. Cybern. 48:16–28, 2016). Among the applications in HSI processing, hyperspectral image classification (HSIC) is a fundamental yet challenging problem (Hong et al. in IEEE Trans. Geosci. Remote Sens. 58:3791–3808, 2020; Jiang et al. in IEEE Trans. Geosci. Remote Sens. 57:851–865, 2019), which aims to assign a specific label to each pixel in the image. HSIC has been widely applied to many scenarios, such as military target detection, vegetation monitoring, and disaster prevention and control.

Recently both graph convolutional neural networks (GCNs) and Transformers have shown promising progress in hyperspectral image (HSI) classification. Transformer-based methods have a great ability to model non-local interactions among spectral and spatial information, whereas GCNs tend to do well in exploiting neighborhood vertex interactions based on their unique aggregation mechanism.

Hyperspectral imagery (HSI) acquired by remote sensing system are composed of hundreds of contiguous and narrow spectral bands with abundant spatial-spectral information in the electromagnetic spectrum (Liu et al. in IEEE Trans Neural Netw Learn Syst 34:8989–9003, 2022; Gao et al. in IEEE Trans Geosci Remote Sens 60:1–15, 2022; Ding et al. in IEEE Geosci Remote Sens Lett 19:1–5, 2022). Due to its unique advantages, HSI has attracted lots of attention and has been widely applied in various fields, including military reconnaissance, urban mapping, biochemical detection, forest fire detection, and target recognition (Ding et al. in IEEE J Select Top Appl Earth Observ Remote Sens 14:4561–4572, 2021; Ding et al. in IEEE Trans Geosci Remote Sens 60:1–12, 2022; Li et al. in IEEE Trans Neural Netw Learn Syst 34:8057–8070, 2022), etc.