Hyperspectral Image Classification Based on Logical Analysis of Data

Abstract

Hyperspectral imaging is a relatively new technique for remote sensing. Earth observation technology and applications are migrating from just plane imaging in a few spectral bands towards intensive spectral imaging. The hyperspectral image is composed of a very narrow continuous spectral band with hundreds of bands. This spectral band usually covers all visible light, near-infrared, mid-infrared and thermal infrared areas. A hyperspectral imaging spectrometer mostly adopts the scanning type or push broom type, and can collect the data of hundreds of spectral bands. Unlike a traditional imaging spectrometer, the hyperspectral spectrometer provides more intensive spectral reflectance values for each pixel in the image rather than having only the interval between the bands. The classification of hyperspectral data aims at acquiring spectral information to distinguish between land cover types or material in all pixels in an image. The classification of a hyperspectral remote sensing image is divided into supervised and unsupervised classifications, parametric and non-parametric, crisp and fuzzy classifications. In this paper, we use the Logical Analysis of Data (LAD) approach in order to classify the spectral signature for each spatial pixel in the image. LAD is a supervised classification technique, which is based on combinatorial and Boolean logical analysis and optimization theory. It can classify data into two or more classes by generating patterns from observations to distinguish one class from others. This approach is based on three stages: 1) the transformation of all types of data into a binary form, 2) the generation of patterns that characterize and distinguish each class, and 3) the theory formation that establishes the model for use in future classification of unclassified new observations. In our experiment, a hyperspectral dataset is divided into two parts: one for training and the other for testing. To illustrate the procedure of using LAD for hyperspectral data classification, the software cbmLAD [46] is used to extract knowledge from the training set of data and to generate patterns that characterize and distinguish each class. In this paper these patterns are called “spectral patterns.” The other set of hyperspectral data is used to test the accuracy of the model that was developed based on the generated patterns. Finally, the accuracy of the classification model is evaluated. The results showed that LAD has a classification accuracy that is comparable to other well-known machine learning techniques, such as Convolutional Neural Network (CNN). Moreover, the patterns generated by LAD offer a unique explanatory knowledge and interpretability of the obtained results. For future research, this property will be used in order to consider the hyperspectral data as partially observed phenomena, due to environmental changes or due to changes in spectral reflectance over time.