Abstract
The aim of this study was to evaluate three different strategies to improve classification accuracy in a highly fragmented semiarid area using, (i) different classification algorithms with parameter optimization in some cases; (ii) different feature sets including spectral, textural and terrain features; and (iii) different seasonal combinations of images. A three-way ANOVA was used to discern which of these approaches and their interactions significantly increases accuracy. Tukey–Kramer contrast using a heteroscedasticity-consistent estimation of the kappa covariances matrix was used to check for significant differences in accuracy. The experiment was carried out with Landsat TM, ETM and OLI images corresponding to the period 2000–2015. A combination of four images using random forest and the three feature sets was the best way to improve accuracy. Maximum likelihood, random forest and support vector machines do not significantly increase accuracy when textural information was added, but do so when terrain features were taken into account. On the other hand, sequential maximum a posteriori increased accuracy when textural features were used, but reduced accuracy substantially when terrain features were included. Random forest using the three feature subsets and sequential maximum a posteriori with spectral and textural features had the largest kappa values, around 0.9.
Highlights
Several factors hinder the classification of remote sensing imagery in Mediterranean landscapes: the high heterogeneity, due to the presence of small patches dedicated to several different land uses and covers [1]; urban sprawl [2]; the high reflectivity of very dry soils and limestone areas that mask the presence of vegetation [3]; and when rainfed and irrigated areas are mixed, they may be difficult to distinguish by using a single scene image for the analysis [4].Several strategies have been developed to overcome these difficulties
Rodríguez-Galiano [13] and Rodriguez-Galiano et al [14] used several different methods to classify land use in a semiarid environment in southern Spain. They reached the same conclusions as the previous authors, stressing that Support Vector Machine (SVM) and Random Forest (RF) are more robust in the presence of noise in the data
Belgiu and Dragut [15], in a recent review of previous research, concluded that the RF classifier outperforms decision trees, the Binary Hierarchical Classifier (BHC), Linear Discriminant Analysis (LDA) and artificial neural network classifiers in terms of classification accuracy; RF and SVM classifiers are reliable, the accuracy of RF being slightly higher for high dimensional input data such as hyperspectral imagery; the SVM classification is more sensitive to the selected features, and it is more complicated to use as several parameters have to be set
Summary
Several factors hinder the classification of remote sensing imagery in Mediterranean landscapes: the high heterogeneity, due to the presence of small patches dedicated to several different land uses and covers [1]; urban sprawl [2]; the high reflectivity of very dry soils and limestone areas that mask the presence of vegetation [3]; and when rainfed and irrigated areas are mixed, they may be difficult to distinguish by using a single scene image for the analysis [4].Several strategies have been developed to overcome these difficulties. Belgiu and Dragut [15], in a recent review of previous research, concluded that the RF classifier outperforms decision trees, the Binary Hierarchical Classifier (BHC), Linear Discriminant Analysis (LDA) and artificial neural network classifiers in terms of classification accuracy; RF and SVM classifiers are reliable, the accuracy of RF being slightly higher for high dimensional input data such as hyperspectral imagery; the SVM classification is more sensitive to the selected features, and it is more complicated to use as several parameters have to be set These authors highlight the lower sensitivity of RF, compared with other algorithms, to erroneous training data or overfitting and point to the robustness due to the ensemble of classifiers and the randomness in the selection of a feature subset when the nodes of the trees are split. The reason for this, according to the authors, is that SMAP reaches higher accuracy values because it can retrieve information from both intra-class and spatial variability
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
