Hyperspectral remote sensing image classification and feature mining

In the field of remote sensing, spectral imaging techniques are often classified into three categories based on spectral resolution: multispectral, hyperspectral, and ultraspectral. Generally, with the same spatial resolution, the higher the spectral resolution, the more image bands there are, the larger the data volume, and the richer the information obtained about the object being measured.

Comparison table of spectral technical characteristics

Hyperspectral image classification distinguishes the types of multiple targets by assigning corresponding labels to groups with similar characteristics; samples with the same label are considered to belong to the same class. The accuracy of classification directly or indirectly affects its application performance.

Compared to ordinary images, hyperspectral images have two characteristics:

1) High spectral resolution, nearly continuous spectral curves, multiple bands, and large data volume make fine classification possible, but at the same time, they also have problems such as serious data redundancy and strong inter-spectral correlation.

2) Limited spatial resolution and the presence of a large number of mixed pixels may lead to the curse of dimensionality. Due to their unique structure and data characteristics, hyperspectral remote sensing images are difficult to directly apply using traditional image processing methods, posing a severe challenge to the field of signal and information processing.

Hyperspectral (including multispectral) image classification and other processing techniques were conventional methods before 1990. From 1990 to 2015, the main methods were based on features and general learning. Since 2015, the focus has shifted to deep learning-based methods. Since 2022, there has been a trend towards learning-based and data and knowledge fusion or joint-driven methods.

Multispectral and hyperspectral image classification methods can be broadly classified into four categories.

Hyperspectral remote sensing satellite

Typical technical specifications of representative hyperspectral satellites and their spectral imagers both domestically and internationally:

Feature mining for hyperspectral image classification

The characteristics of hyperspectral image data mainly refer to spectral features, spatial features, temporal features, and spatial-spectral, spatial-temporal, and combined spatial-temporal-spectral features. In hyperspectral image classification, common spectral features include spectral absorption locations, reflectance, symmetry, the location and number of peaks and valleys in the spectral curve, and slope, as well as the gradient and higher-order statistics of the spectral curve. Spatial features of hyperspectral images mainly include amplitude, color, texture, shape, and spatial relationships. Temporal features of hyperspectral images mainly include stability over time, (time) derivative, and rate of change.

Hyperspectral features and feature mining methods

Hyperspectral image feature mining mainly includes feature selection, feature extraction, and feature fusion. The methods for spectral feature mining of hyperspectral remote sensing data and their classification are shown in the table below:

Methods for Spectral Feature Mining and Classification of Hyperspectral Remote Sensing Images

Feature selection generally refers to directly selecting m (m

Feature extraction refers to transforming the original n-dimensional vector space (the original data is hyperspectral cubic data) and then finding its feature subspace. Feature extraction involves finding a mapping relationship that transforms the original three-dimensional dataset into a lower-dimensional m-dimensional feature subspace, generating m new features through this transformation. Feature extraction typically searches for feature parameters reflecting various targets from all bands of data; obviously, the computational cost is much higher than feature selection, making the process inherently slow.

Feature combination or feature fusion utilizes the advantages of multiple features, such as combined spectral-spatial-temporal features, or combinations of direct energy features with transformation features (such as derivatives). The performance of manual feature mining is closely related to factors such as domain expert knowledge, physicochemical properties, models, and experience.

参考：He Mingyi，Li Qiang，Guo Ying，Yan Hongmei.2025. Overview and prospects of intelligent classification of hyperspectral images. Journal of Image and Graphics，30(6):2207-2238