Land Cover Classification Using Local Softened Affine Hull

Training samples are usually very scarce in land cover classification, which challenges many supervised classifiers. To deal with this problem, this work presents a new learning approach, called local softened affine hull (LSAH). One of the most attractive characters of this classifier is its ability to expand the training set through exploiting “virtual” prototypes. During classification, this method utilizes some local prototypes around the query sample to construct SAH manifolds for each class, which are then employed to determine the label ownership of the query by the nearest neighbor rule. Because each SAH represents infinite “virtual” samples that approximate the “underlying” or “possible” variants of the real prototypes, the representational capacity of the training set is greatly enlarged; LSAH thus avoids the sample scarce problem. Meanwhile, as the number of the local prototypes is usually very small relative to the total training samples of each class, LSAH can run efficiently in classification tasks. LSAH's performance is demonstrated on some land cover classification tasks using three remote sensing images, in comparison with several popular algorithms, including maximum likelihood classifier, K-nearest neighbor, backpropagation neural network, and support vector machine. Experimental results show that the accuracy of LSAH is significantly higher than those of most of others, with the classification speed being quite comparable.