Virtual Support Vector Machines with self-learning strategy for classification of multispectral remote sensing imagery

We follow the idea of learning invariant decision functions for remote sensing image classification with Support Vector Machines (SVM). To do so, we generate artificially transformed samples (i.e., virtual samples) from available prior knowledge. Labeled samples closest to the separating hyperplane with maximum margin (i.e., the Support Vectors) are identified by learning an initial SVM model. The Support Vectors are used for generating virtual samples by perturbing the features to which the model should be invariant. Subsequently, the model is releamed using the Support Vectors and the virtual samples to eventually alter the hyperplane with maximum margin and enhance generalization capabilities of decision functions. In contrast to existing approaches, we establish a self-learning procedure to ultimately prune non-informative virtual samples from a possibly arbitrary invariance generation process to allow for robust and sparse model solutions. The self-learning strategy jointly considers a similarity and margin sampling constraint. In addition, we innovatively explore the invariance generation process in the context of an object-based image analysis framework. Image elements (i.e., pixels) are aggregated to image objects (as represented by segments/superpixels) with a segmentation algorithm. From an initial singular segmentation level, invariances are encoded by varying hyperparameters of the segmentation algorithm in terms of scale and shape. Experimental results are obtained from two very high spatial resolution multispectral data sets acquired over the city of Cologne, Germany, and the Hagadera Refugee Camp, Kenya. Comparative model accuracy evaluations underline the favorable performance properties of the proposed methods especially in settings with very few labeled samples.