Object-based hyperspectral image classification using a new latent block model based on hidden Markov random fields

Efficient high-dimensional analyses of hyperspectral datasets and their utilization within classification algorithms is a popular topic in the field of data analytics. A powerful tool for summarizing a large array of datasets is the latent block model (LBM), which finds homogeneous blocks within the data using the finite mixture model (FMM). In this study, for the first time, LBM was modified by replacing the hidden Markov random field (HMRF) instead of using FMM to consider the spatial relationship between pixels and, thus, to develop a new object-based classification algorithm. The proposed clustering algorithm was named LBMHMRF and was used along with the support vector machine (SVM) algorithm to classify land cover/land use (LCLU) categories using two hyperspectral datasets. Unlike LBM, HMRF, and MultiHMRF, the LBMHMRF algorithm allows for the use of more spectral information without estimating a large number of parameters and produces a model with high computation costs saving feature. Additionally, the segmentation results are produced in a shorter period of time compared to the above-mentioned algorithms. It was observed that the proposed object-based classification algorithm (i.e., LBMHMRF + SVM) had the highest potential in terms of visual and statistical accuracies as well as computation time compared to the pixel-based SVM, object-based HMRF + SVM, and MultiHMRF + SVM. The average overall classification accuracies considering the different datasets and cases investigated in this study were 93.1%, 94.6%, 95.7%, and 96.4% for SVM, HMRF + SVM, MultiHMRF + SVM, and LBMHMRF + SVM, respectively.