SVM- and MRF-Based Method for Accurate Classification of Hyperspectral Images

Abstract

Abstract-The high number of spectral bands acquired by hyperspectral sensors increases the capability to distinguish physical materials and objects, presenting new challenges to image analysis and classification. This letter presents a novel method for accurate spectral-spatial classification of hyperspectral images. The proposed technique consists of two steps. In the first step, a probabilistic support vector machine pixelwise classification of the hyperspectral image is applied. In the second step, spatial contextual information is used for refining the classification results obtained in the first step. This is achieved by means of a Markov random field regularization. Experimental results are presented for three hyperspectral airborne images and compared with those obtained by recently proposed advanced spectral-spatial classification techniques. The proposed method improves classification accuracies when compared to other classification approaches. Index Terms-Classification, hyperspectral images, Markov random field (MRF), support vector machine (SVM).

Spatial‐spectral classification of hyperspectral images: a deep learning framework with Markov Random fields based modelling

“…State-of-the-art methods for HSI classification will be applied in the experiments as comparison, which include: SVM [8], EMP [21], SVM-composite kernel (SVM-CK) [38], LORSL-MLL [6], sparse representation-based classification (SRC) [10], SVM-MRF [25], CNN [31], 2DCNN [32] and 3DCNN [33], [39]. The SVM only considers the spectral information, implemented by the Gaussian kernel.…”

Gaze Direction Estimation Method Using Support Vector Machines (SVMs)

“…They provide a flexible tool to include spatial context into image-analysis schemes in terms of minimization of suitable energy functions. While earlier algorithms for optimizing MRF energy, such as iterated conditional modes (ICM) and simulated annealing (Solberg, et al, 1996), (Tarabalka, et al, 2010b) were time consuming, more advanced methods, such as graph cuts ), (Li, et al, 2012) provided powerful alternatives from both theoretical and computational viewpoints, resulting in a growing use of the MRF-based segmentation techniques.…”

“…~ denotes a pair of spatially adjacent pixels, and , ( , ) is an interaction term for these pixels. Most works in remote sensing image classification use a Potts model (Tarabalka, et al, 2010b), (Li, et al, 2012) to compute this spatial term, which favors spatially adjacent pixels to belong to the same class (or spatial region): (16)(17)(18)(19)(20)(21)(22)(23) where (•) is the Kronecker function ( ( , ) = 1 for = and ( , ) = 0 otherwise) and is a positive constant parameter that controls the importance of spatial smoothing. This model tends to deteriorate classification results at the edges between land-cover classes and near small-scale details.…”

“…This model tends to deteriorate classification results at the edges between land-cover classes and near small-scale details. In order to preserve the border in the output thematic map, "edge" functions have been proposed and integrated in the spatial energy term, such as (Tarabalka, et al, 2010b): (16)(17)(18)(19)(20)(21)(22)(23)(24) where | is the gradient value of the pixel in the direction of , and is a parameter controlling the fuzzy edge threshold.…”

“…For the graph-cut method, probabilistic SVM classification was applied to derive the data energy term, and a Potts model was used to express the spatial term. We fixed the parameter = 1.5 as suggested in (Tarabalka, et al, 2010b). The MRFbased method using the -expansion optimization yields the highest overall classification accuracy.…”