Optimized Lithological Mapping from Multispectral and Hyperspectral Remote Sensing Images Using Fused Multi-Classifiers

Mahendra Pal,Thorkild Rasmussen,Alok Porwal

doi:10.3390/rs12010177

Mahendra Pal, Thorkild Rasmussen + Show 1 more

Open Access

https://doi.org/10.3390/rs12010177

Copy DOI

Abstract

Most available studies in lithological mapping using spaceborne multispectral and hyperspectral remote sensing images employ different classification and spectral matching algorithms for performing this task; however, our experiment reveals that no single algorithm renders satisfactory results. Therefore, a new approach based on an ensemble of classifiers is presented for lithological mapping using remote sensing images in this paper, which returns enhanced accuracy. The proposed method uses a weighted pooling approach for lithological mapping at each pixel level using the agreement of the class accuracy, overall accuracy and kappa coefficient from the multi-classifiers of an image. The technique is implemented in four steps; (1) classification images are generated using a variety of classifiers; (2) accuracy assessments are performed for each class, overall classification and estimation of kappa coefficient for every classifier; (3) an overall within-class accuracy index is estimated by weighting class accuracy, overall accuracy and kappa coefficient for each class and every classifier; (4) finally each pixel is assigned to a class for which it has the highest overall within-class accuracy index amongst all classes in all classifiers. To demonstrate the strength of the developed approach, four supervised classifiers (minimum distance (MD), spectral angle mapper (SAM), spectral information divergence (SID), support vector machine (SVM)) are used on one hyperspectral image (Hyperion) and two multispectral images (ASTER, Landsat 8-OLI) for mapping lithological units of the Udaipur area, Rajasthan, western India. The method is found significantly effective in increasing the accuracy in lithological mapping.

Highlights

Mapping lithological units of an area using satellite-borne remote sensing data is one of the most challenging geological applications of remote sensing technology, given the complexities involving sub-pixel level microscopic-scale non-linear mixing of minerals and the presence of surface soil, regolith and vegetation
It is clear that all the litho-classes and overall lithological accuracy obtained by the proposed method applied to the Hyperion image are better compared to ASTER and Landsat 8 images, while additional improvement achieved by combined images (Hyperion, ASTER, Landsat 8), which likewise confirmed that the proposed method is effective in enhancement of classification accuracy in lithological mapping by multiple images
The complexity and restricted data quality in hyperspectral and multispectral remote sensing images, and limited robustness in single classification algorithms lead to the development of ensemble-based techniques for geological mapping using multi-classifiers

Summary

Introduction

Mapping lithological units of an area using satellite-borne remote sensing data is one of the most challenging geological applications of remote sensing technology, given the complexities involving sub-pixel level microscopic-scale non-linear mixing of minerals and the presence of surface soil, regolith and vegetation. Expert systems, Support Vector Machine (SVM), deep learning and neural networks are typical examples of non-parametric classifiers which are widely preferred for classification remote sensing images from complex terrain [2,3,4,5,6,7]. The performance of the neural network-based classifiers can be adversely affected due high dimensionality and unavailability of enough and satisfactory training and test samples of remote sensing images from complex terrains [6,9]. In recent years, bagging, boosting, or a hybrid of both techniques are being increasingly used to enhance the classification performance of non-parametric as well as parametric classifiers [6] These methods have been utilized in the framework of decision trees [10] and SVM [2,4,7,11,12,13] to enhance classifications accuracy

Methods

Results

Discussion

Conclusion