Improving Selection of Spectral Variables for Vegetation Classification of East Dongting Lake, China, Using a Gaofen-1 Image

Abstract

Abstract:There is a large amount of remote sensing data available for land use and land cover (LULC) classification and thus optimizing selection of remote sensing variables is a great challenge. Although many methods such as Jeffreys-Matusita (JM) distance and random forests (RF) have been developed for this purpose, the existing methods ignore correlation and information duplication among remote sensing variables. In this study, a novel approach was proposed to improve the measures of potential class separability for the selection of remote sensing variables by taking into account correlations among the variables. The proposed method was examined with a total of thirteen spectral variables from a Gaofen-1 image, three class separability measures including JM distance, transformed divergence and B-distance and three classifiers including Bayesian discriminant (BD), Mahalanobis distance (MD) and RF for classification of six LULC types at the East Dongting Lake of Hunan, China. The results showed that (1) The proposed approach selected the first three spectral variables and resulted in statistically stable classification accuracies for three improved class separability measures. That is, the classification accuracies using three or more spectral variables statistically did not significantly differ from each other at a significant level of 0.05; (2) The statistically stable classification accuracies obtained by integrating MD and BD classifiers with the improved class separability measures were also statistically not significantly different from those by RF; (3) The numbers of the selected spectral variables using the improved class separability measures to create the statistically stable classification accuracies by MD and BD classifiers were much smaller than those from the original class separability measures and RF; and (4) Three original class separability measures and RF led to similar ranks of importance of the spectral variables, while the ranks achieved by the improved class separability measures were different due to the consideration of correlations among the variables. This indicated that the proposed method more effectively and quickly selected the spectral variables to produce the statistically stable classification accuracies compared with the original class separability measures and RF and thus improved the selection of the spectral variables for the classification.

Estimating the Growing Stem Volume of Chinese Pine and Larch Plantations based on Fused Optical Data Using an Improved Variable Screening Method and Stacking Algorithm

“…The DC-FSCK can examine the linear and nonlinear relationships between feature variables and GSV and consider the self-correlation and combination effects among the feature variables. The SRA and RF can only select the variables with good linear correlation or importance, but discard the rest variables that might have a high saturation level and contain useful information [19]. Song et al [19] proposed a spectral variable selection method by integrating Jeffreys-Matusita distance and correlation among feature variables for image classification of wetlands and found that their method offered greater potential than RF.…”

“…The SRA and RF can only select the variables with good linear correlation or importance, but discard the rest variables that might have a high saturation level and contain useful information [19]. Song et al [19] proposed a spectral variable selection method by integrating Jeffreys-Matusita distance and correlation among feature variables for image classification of wetlands and found that their method offered greater potential than RF. However, their method ignores the combination effects of the feature variables.…”

“…However, this also leads to difficulty in selecting the variables that can significantly increase the GSV estimation accuracy. Principal component analysis (PCA) is often used to reduce the information redundancy, as it linearly combines original variables and transforms them into a set of new variables that are not correlated with each other [13,19]. The new variables are not bio-physically meaningful.…”

“…The new variables are not bio-physically meaningful. In addition, Stepwise Regression Analysis (SRA) [28,30], Random Forest (RF) [13,19,46], the sure independence screening procedure based on Pearson correlation (PC-SIS) [47], the sure independence screening procedure based on distance correlation (DC-SIS) [48], and the fast iterative features selection method for k-nearest neighbor (kNN-FIFS) have also been widely used for variables selection in classification or regression modeling [49]. Unlike PCA, these methods keep the original variables and do not generate new variables when reducing the dimensionality of sample data.…”

“…Ensemble learning algorithms, such as RF, eXtreme Gradient Boosting (XGBoost) and Stacking, often have better performance than traditional regression algorithms such as Multiple Linear Regression (MLR) [53][54][55][56][57][58]. However, the ensemble learning algorithms are generally more sensitive to the characteristics of training samples, including sample sizes and representatives of population features, and thus require more training samples than the traditional parametric algorithms [13,19,45]. In addition, the performance of the algorithms may be site-specific [13,45].…”