The production of land cover maps through satellite image classification is a frequent task in remote sensing. Random Forest (RF) and Support Vector Machine (SVM) are the two most well-known and recurrently used methods for this task. In this paper, we evaluate the pros and cons of using an RF-based kernel (RFK) in an SVM compared to using the conventional Radial Basis Function (RBF) kernel and standard RF classifier. A time series of seven multispectral WorldView-2 images acquired over Sukumba (Mali) and a single hyperspectral AVIRIS image acquired over Salinas Valley (CA, USA) are used to illustrate the analyses. For each study area, SVM-RFK, RF, and SVM-RBF were trained and tested under different conditions over ten subsets. The spectral features for Sukumba were extended by obtaining vegetation indices (VIs) and grey-level co-occurrence matrices (GLCMs), the Salinas dataset is used as benchmarking with its original number of features. In Sukumba, the overall accuracies (OAs) based on the spectral features only are of $81.34\%$ , $81.08\%$ and $82.08\%$ for SVM-RFK, RF, and SVM-RBF. Adding VI and GLCM features results in OAs of $82\%$ , $80.82\%$ and $77.96\%$ . In Salinas, OAs are of $94.42\%$ , $95.83\%$ and $94.16\%$ . These results show that SVM-RFK yields slightly higher OAs than RF in high dimensional and noisy experiments, and it provides competitive results in the rest of the experiments. They also show that SVM-RFK generates highly competitive results when compared to SVM-RBF while substantially reducing the time and computational cost associated with parametrizing the kernel. Moreover, SVM-RFK outperforms SVM-RBF in high dimensional and noisy problems. RF was also used to select the most important features for the extended dataset of Sukumba; the SVM-RFK derived from these features improved the OA of the previous SVM-RFK by 2%. Thus, the proposed SVM-RFK classifier is as at least as good as RF and SVM-RBF and can achieve considerable improvements when applied to high dimensional data and when combined with RF-based feature selection methods.

The classification of the ever-increasing collections of remotely sensed images is a key but challenging task. In this letter, we introduce the use of extremely randomized trees known as Extra-Trees (ET) to create a similarity kernel [ET kernel (ETK)] that is subsequently used in a support vector machine (SVM) to create a novel classifier. The performance of this classifier is benchmarked against that of a standard ET, an SVM with both conventional radial basis function (RBF) kernel, and a recently introduced random forest-based kernel (RFK). A time series of Worldview-2 images over smallholder farms is used to illustrate our study. Four sets of features were obtained from these images by extending their original spectral bands with vegetation indices and textures derived from gray-level co-occurrence matrices. This allows testing the performance of the classifiers in low- and high-dimensional problems. Our results for the high-dimensional experiments show that the SVM with tree-based kernels provide better overall accuracies than with the RBF kernel. For problems with lower dimensionality, SVM-ETK slightly outperforms SVM-RFK and SVM-RBF. Moreover, SVM-ETK outperforms ET in most of the experiments. Besides an improved overall accuracy, the main advantage of ETK is its relatively low computational cost compared to the parameterization of the RBF and RFK. Thus, the proposed SVM-ETK classifier is an efficient alternative to common classifiers, especially in problems involving high-dimensional data sets.

Random forest (RF) is a popular ensemble learning method that is widely used for the analysis of remote sensing images. RF also has connections with kernel-based method. Its tree-based structure can generate a Random Forest Kernel (RFK) that provides an alternative to common kernels such as Radial Basis Function (RBF) in kernel-based methods such as Support Vector Machine (SVM). Using RFK in an SVM has been shown to outperform both RF and SVM-RBF (i.e., using an RBF kernel in an SVM) in classification tasks with a high number of features. Here, we explore new designs of RFKs for remote sensing image classification. Different RF structural parameters and characteristics are used to generate various RFKs. In particular, we explore the use of RF's depth, the number of branches between terminal nodes of trees, and the predicted class probabilities for designing and evaluating new RFK. Two depth-based kernel are proposed: an RFK at the optimal depth, and a multi-scale one created by combining RFKs at multiple depths. We evaluate the proposed kernels within an SVM by classifying a time series of Worldview-2 images, and by designing experiments having a various number of features. Benchmarking the new RFKs against RBF shows that the newly proposed kernels outperform RBF kernel for the experiments with a higher number of features. For the experiments with a lower number of features, RFKs and RBF kernel perform at about the same level. Benchmarking against standard RF also shows the general outperformance of the proposed RFKs in SVM.

The authors wish to make the following correction to the paper