In the field of remote sensing, the classification of hyperspectral images (HSI) has gained popularity. Convolutional neural networks (CNNs) are a potent visual model that have attracted a lot of attention recently due to their outstanding performance in a variety of visual recognition problems. However, CNNs' performance may be limited if they can't fully utilize the extensive spectral information contained in hyperspectral images. One approach to overcome this limitation is to incorporate attention mechanisms into the CNN architecture. Attention allows the model to focus on the most relevant parts of the input, enabling it to better capture the spectral characteristics of different materials. This chapter focuses on all the studies that are carried out with the convolutional neural network and attention module approach on the hyperspectral image classification in remote sensing between 2012 and 2022. The major objective of this study is to review, identify, evaluate, and analyze the performance of CNN models and attention module in hyperspectral image classification.
Top1. Introduction
Acquisition of Hyperspectral images depends on imaging spectrometers set up in different fields. The imaging spectrum was first used in the 1980s. It is used to obtain images in the ultraviolet, visible, near-infrared, and mid-infrared regions of electromagnetic waves. The imaging spectrometer can take many continuous and very narrow band images so that every pixel in the wavelength range used can obtain a fully reflected or emitted spectrum. Therefore, hyperspectral images feature high spectral resolution, many bands, and abundant information. The processing methods of hyperspectral remote sensing images mainly include image correction, noise reduction, transformation, dimensionality reduction, and classification. Unlike ordinary images, hyperspectral images are rich in spectral information, and this spectral information can reflect the physical structure and chemical composition of the object of interest, which is helpful for image classification (Li et al., 2019). Hyperspectral image classification is the most active part of research in the hyperspectral field. Computer classification of remote sensing images is the specific application of automatic pattern recognition technology in the field of remote sensing (Li et al., 2019). Furthermore, as hyperspectral imaging systems advance, the acquired information will become increasingly detailed. This will result in both enhanced spatial resolution and significant improvement in spectral resolution. With the further development of hyperspectral imaging systems, the amount of information contained in hyperspectral images will further increase and the scope of application of hyperspectral images will also expand.
There are several fundamental issues with hyperspectral image classification technology that need to be resolved. Firstly, the sheer volume of data generated by hyperspectral images can be overwhelming, requiring advanced data storage and processing capabilities. Moreover, the high dimensionality of these datasets introduces the curse of dimensionality, complicating traditional classification approaches. Noise reduction and atmospheric corrections also become paramount due to the sensitivity of these sensors to external conditions. Furthermore, the variability in spectral signatures among similar materials, possibly due to slight changes in composition or external factors, challenges the distinct classification of these materials. To address these concerns, there is a need for more sophisticated algorithms, improved sensor technology, and enhanced preprocessing techniques that can optimize the performance and accuracy of hyperspectral image analysis.
In the early stages of hyperspectral image classification research, spectral information was typically the main focus, and several classification techniques were developed, including support vector machine (SVM), random forest (RF), neural networks, and logistic regression.
Deep learning-based methods extract the most effective spatial features and enable object identification from hyperspectral data. Deep learning techniques, which have shown great success in image processing compared to traditional classification methods based on artificial features and surface learning models, aim to learn hierarchical and distributed features from raw data instead of using artificial features. These features obtained using layered architectures and nonlinear activation functions are more robust than traditional methods (Yang et al., 2018). Convolutional Neural Network (CNN) is one of the most popular deep learning methods, uses shared weights to reduce the number of trainable parameters and efficiently extract local links and local information. It also takes parts of fixed-size images to extract spatial attributes and keep spatial information unchanged (Yue et al., 2015, Liang et al., 2016).
The attention mechanism is a deep learning technique used to emphasize important features in a data representation. Recently, the attention mechanism has been popularly employed in language modeling and computer vision tasks. Its success mainly depends on the reasonable assumption that human vision tends to only focus on selective parts of the whole visual space when and where needed (Ghamisi et al., 2020). In hyperspectral image classification with attention-aided methods, the attention mechanism is used to focus on the most informative parts of the hyperspectral image, thus improving the classification accuracy. This can be done by weighting the contribution of each feature to the final prediction, where features with higher weights are considered more important.