VHR Semantic Labeling by Random Forest Classification and Fusion of Spectral and Spatial Features on Google Earth Engine

Semantic labeling is an active field in remote sensing applications. Although handling highly detailed objects in a Very High Resolution (VHR) optical image, and the VHR Digital Surface Model (DSM) is a challenging task, it can improve the accuracy of the semantic labeling methods. In this paper, a semantic labeling method is proposed by fusion of optical and normalized DSM data. The spectral and spatial features are fused into a heterogeneous feature map to train the classifier. The evaluation database classes are impervious surface, building, low vegetation, tree, car, and background. The proposed method is implemented on the Google Earth Engine. The method consists of several levels. First, the Principal Component Analysis (PCA) is applied to the vegetation indices to find the maximum separable color space between the vegetation and non-vegetation areas. The Gray Level Co-occurrence Matrix (GLCM) is computed to provide the texture information as the spatial features. Several Random Forests (RFs) are trained with an automatically selected train dataset. Several spatial operators follow the classification to refine the result. The LeafLess Tree (LLT) feature is used to solve the underestimation problem in the tree detection. The area, and major and minor axes of the connected components are used to refine building and car detection. The evaluation shows a significant improvement in the tree, building, and car accuracy. The overall accuracy and Kappa coefficient are appropriate.