Balancing hyperspectral dimensionality reduction and information preservation for machine learning-based lithological classification using EnMAP hyperspectral imagery
Abstract
This study addresses the applicability of dimensionality reduction (DR) techniques with EnMAP (Environmental Mapping and Analysis Program) hyperspectral imagery for lithological classification in a semiarid area. In this work, various DR techniques such as principal component analysis (PCA), independent component analysis (ICA), minimum noise fraction (MNF), and manifold nonlinear models were applied to EnMAP data. The classification of the ENMAP dataset was then performed using Support Vector Machines (SVM), k-nearest neighbors (KNN), and random forests (RF). The DR methods were evaluated with a reduction rate of more than 50 %. The study demonstrated that most of the used DR techniques show sensitivity between classification accuracy and the number of components until 30 components. Techniques such as PCA transform variants, ICA and Singular value decomposition (SVD) provide high accuracy with a smaller number of components, while others such as locally linear embedding (LLE), local tangent space alignment (LTSA), and Autoencoder (AE) require a larger number of components to achieve a similar level of accuracy but still provide considerable DR results. On the other hand, the Uniform Manifold Approximation and Projection (UMAP) and Isometric mapping (ISOMAP) are not very accurate and, therefore reveal their weakness in preserving relevant information during DR. The SVM model showed higher performance and lower sensitivity to reduction rate changes than KNN and RF. The SVM achieves the ultimate performance in classifying the lithologic units in the study area over both the original bands and the reduced dimensions obtained with the PCA technique (Overall Accuracy (OA) = 96.09 %, Precision (Pr) = 96.17 % and Kappa Accuracy (KA) = 95.55 %). These results emphasize the need to choose an appropriate DR technique to achieve a balance between DR and information preservation. Indeed, the optimal number of components depends on the technique used and the dataset, emphasizing that the proper selection of a technique, together with a possible tuning of hyperparameters, is very important. This work makes a valuable contribution to research in lithological mapping studies and shows that effective dimensionality reduction is a key to improving model accuracy for processing EnMAP data.
