Recent developments in hyperspectral sensors have made it possible to acquire HyperSpectral Images (HSI) with higher spectral and spatial resolution. Hence, it is now possible to extract detailed information about relatively smaller structures. Despite these advantages, HSI suffers from many challenges also, like higher spatial variability of spectral signatures, the Hughes effect due to higher dimensionality, and a limited number of labeled training samples compared to the dimensions of the spectral space. Superpixels can be a potentially effective tool in tackling these challenges. Superpixel segmentation is a process of segmenting the spatial image into several semantic sub-regions with similar characteristic features. Such grouping by similarity can significantly ease the subsequent processing steps. Because of this, superpixels have been successfully applied to various fields of HSI processing like Classification, Spectral Unmixing, Dimensionality Reduction, Band Selection, Active Learning, Denoising, and Anomaly Detection. This paper focuses on Classification, presenting a detailed survey of superpixel segmentation approaches for the classification of HSI. The Superpixel creation algorithm framework and post-processing frameworks for superpixels in HSI are also analyzed. Also, a brief description of various application areas of superpixels is provided. An experimental analysis of existing superpixel segmentation approaches is also provided in this paper, supported by quantitative results on standard benchmark datasets. The challenges and future research directions for the implementation of superpixel algorithms are also discussed.

A hyperspectral image (HSI) is a collection of several narrow-band images that span a wide spectral range. Each band reflects the same scene, composed of various objects imaged at different wavelengths; the spatial information, however, remains generally consistent across bands. Both types of information, spectral and spatial, can be leveraged to identify and classify objects. Recently, the use of machine learning (ML) in object classification has become increasingly widespread. Regardless of the selected approach, object-specific spectral and spatial information is key to discriminating relevant categories. Whereas spatial information is usually repeated across bands, spectral information tends to be distributed more unevenly and often highly so. This poses the issue of removing redundancy, which is commonly called the band selection ( BS ) problem and refers to identifying an optimal subset of bands for further HSI processing.

The hyperspectral images (HSIs) often suffer from Hughes effect, as it records information of a single scene in several spectral bands. This can be mitigated by reducing the dimension of HSI. A novel framework for hybrid band selection (BS) is proposed in this work. The proposed technique is a multi‐objective approach, which incorporates clustering (spectral) and intra‐band (spatially filtered) de‐correlation measure (Frobenius norm) as maximisation of two cost functions. Heuristic optimisers are very sensitive to their associated hyperparameters. So, in the proposed architecture, Jaya optimisation is used for BS, as it does not possess any algorithm‐specific control parameter. Spatial and spectral features are extracted for both BS and classification (using support vector machine) for evaluating the effectiveness of the proposed/ experimented approaches. The evaluated performance measures are overall accuracy, average accuracy, and kappa (K). The experimental result shows that the proposed BS approach is better or competent with other experimented state‐of‐the‐art methods. The advantages of the proposed framework can be stated as: (i) spectrally distinct and spatially invariant objective formulation; (ii) Jaya optimisation with minimal control parameters; (iii) optimised ranking for more accurate BS; and (iv) performing classification using spatial–spectral features for further band reduction with the desired accuracy.

The problem of classification is shared across various disciplines. Designing even less computationally demanding and more effective classifiers has been a key challenge for researchers for many years. No single classifier can be highly effective for all types of datasets and thus, depending on the data distribution, various classifiers have been proposed in the literature. To our knowledge, feature values have been vastly exploited as the base for discriminating classes, while feature sequence information has been somehow under-exploited so far. In the proposed approach normalised features are sorted and ranked, creating a sequence of finite numbers. The associated rank of the created sequence is used as an additional feature, which in a way defines the sample-specific intra-feature relationship. Three novel dictionary-based approaches such as Sequence Classifier (SC), Sequence-dictionary-based k-Nearest Neighbours Classifier (SDk-NN) and Combined-dictionary-based k-Nearest Neighbours Classifier (CDk-NN) are proposed in this paper. In the case of remotely sensed data, and specifically in Hyper-Spectral Images (HSI), the spectral features (Spectral signatures) represent a strong, object-specific spectral relationship, which is a key point in our proposed approach. In this case, indeed, the proposed classifiers were tested over various (five) HS datasets and found to be effective. Based on the classifiers features, two derived distance measures are proposed and validated for the HS dataset, namely: the Normalised Sequence Distance (NSD) measure and Combined Distance (CD) measure. These measures appear to overperform the conventional Normalised Euclidean Distance (NED) in this context. Also, validation for both binary and multi-class datasets are experimented and their performances are evaluated in terms of accuracy and other standard measures. Experimental results over 21 datasets revealed that the proposed approaches perform comparably, and in some cases even better than other classifiers. Stack-operated, class-specific sparse dictionaries are also introduced in order to reduce the computational complexity, which can be used as an active learning-based approach for optimal training sample selection. Additional tests were performed with variable levels of dictionary sparsity for assessing its impact on accuracy.

Hyperspectral imaging is a key tool in numerous remote sensing applications. Images with tens or even hundreds of spectral bands contain a wealth of information to retrieve multiple geophysical parameters, to perform target detection, or land classification with unbeatable high accuracies. In addition, hyperspectral sensors are becoming smaller, and today they even fit in CubeSats. However, the amount of data they generate is so large that satellite communication systems have severe limitations to download it, especially in SmallSats. It is therefore becoming urgent to develop efficient automated algorithms that can be executed in the limited capabilities of the onboard computers of these satellites, so as to reduce the amount of data to be stored and downloaded, while keeping as much information as possible for a given scene. In this work, a band selection algorithm has been designed to deal with this problem. The proposed algorithm consists of the sequential selection of the spectral bands ranked using the amount of information provided by each band, and also the correlation of these bands with the previously selected ones. The algorithm performance is assessed by means of a suite of classification tests with hyperspectral datasets from different sensors. Results show comparable or even better performance than other existing band selection algorithms, while outperforming in terms of computational complexity, which makes it more suitable for SmallSats with limited computing resources.