Geleneksel görüntü işleme tekniklerinin direkt olarak uzaktan algılamaya uygulanması, sadece multispektral datalar için geçerli olabilir. Öznitelik vektörü boyutu 100-200 civarında olan hiperspektral dataların analizi için gelişmiş algoritmalara ihtiyaç vardır. Bununla birlikte, uzaktan algılamada, genellikle sınırlı sayıda eğitim örneğinin olması, özellikle öznitelik vektörünün boyutunun büyük olduğu hiperspektral datalarda, parametrik sınıflayıcıların kullanımını kısıtlar. Bu çalışmanın amacı, istatistiksel dağılıma bağlı olmayan, sadece eldeki eğitim örneklerine dayanan bir algoritma geliştirerek yukarıda özetlenen uzaktan algılama için genel sınıflandırma problemlerinin üstesinden gelmektir. Önerilen Sınır Özniteliklerinin Belirlenmesi ve Adaptasyonu (SÖBA) algoritması, karar yüzeylerine yakın sınır öznitelik vektörlerini kullanır ve bu sınır öznitelik vektörleri, maksimum marjin prensibini sağlayacak şekilde adapte edilerek, öznitelik uzayında doğru bölütlemenin yapılmasını sağlar. SÖBA algoritması iki bölümden oluşur. İlk aşamada sınır öznitelik vektörlerinin başlangıç değerleri uygun eğitim kümesi elemanlarından, yönetilebilir sayıda atanır. Daha sonra uygulanan adaptasyon işlemiyle, öğrenme süreci gerçekleştirilerek sınır özniteliklerinin, sonuç değerlerine ulaşması hedeflenir. Sınıflandırma sonuç sınır öznitelik vektörlerine olan en yakın 1 komşuluk (1-EK) kuralı uyarınca yapılır. Ek olarak, SÖBA algoritmasının sınır öznitelik vektörlerinin başlangıç değerlerine ve eğitim kümesi elemanlarının eğitimde kullanılma sırasına bağlı olarak her çalışmasında kabul edilebilir derecede farklı sınır karar yüzeyleri oluşturması, konsensüs yapılarda kullanılması için elverişli bir özelliktir. Böylece birçok defa çalıştırılan SÖBA kararlarının uygun kurallarla birleştirilmesiyle tek bir sınıflayıcının aldığı karardan çok daha doğru kararlar elde edilebilir.

Various types of sensors collect very large amounts of data from the earth surface. The characteristics of the data are related to sensor type with its own imaging geometry. Consequently, sensor types affect processing techniques used in remote sensing. In general, image processing techniques used in remote sensing are usually valid for multispectral data which is relatively in a low dimensional feature space. Therefore, advanced algorithms are needed for hyperspectral data which have at least 100-200 features (attributes/bands). Additionally, the training process is very important and affects the generalization capability of a classifier in supervised learning. Enough number of training samples is required to make proper classification. In remote sensing, collecting training samples is difficult and costly. Consequently, a limited number of training samples is often available in practice. Conventional statistical classifiers assume that the data have a specific distribution. For real world data, these kinds of assumptions may not be valid. Additionally, proper parameter estimation is difficult, especially for hyperspectral data. Normally, when the number of bands used in the classification process increases, precise detailed class determination is expected. For high dimensional feature space, when a new feature is added to the data, classification error decreases, but at the same time, the bias of the classification error increases. If the increment of the bias of the classification error is more than the reduction in classification error, then the use of the additional feature decreases the performance of the decision algorithm. This phenomenon is called the Hughes effect, and it may be much more harmful with hyperspectral data than with multispectral data. Our motivation in this study is to overcome some of these general classification problems by developing a classification algorithm which is directly based on the available training data rather than on the underlying statistical data distribution. Our proposed algorithm, Border Feature Detection and Adaptation (BFDA), uses border feature vectors near the decision boundaries which are adapted to make a precise partitioning in the feature space by using maximum margin principle. The BFDA algorithm well suited for classification of remote sensing images is developed with a new approach to choosing and adapting border feature vectors with the training data. This approach is especially effective when the information source has a limited amount of data samples, and the distribution of the data is not necessarily Gaussian. Training samples closer to class borders are more prone to generate misclassification, and therefore are significant feature vectors to be used to reduce classification errors. The proposed classification algorithm searches for such error-causing training samples in a special way, and adapts them to generate border feature vectors to be used as labeled feature vectors for classification. The BFDA algorithm can be considered in two parts. The first part of the algorithm consists of defining initial border feature vectors using class centers and misclassified training vectors. With this approach, a manageable number of border feature vectors is achieved. The second part of the algorithm is adaptation of border feature vectors by using a technique which has some similarity with the learning vector quantization (LVQ) algorithm. In this adaptation process, the border feature vectors are adaptively modified to support proper distances between them and the class centers, and to increase the margins between neighboring border features with different class labels. The class centers are also adapted during this process. Subsequent classification is based on labeled border feature vectors and class centers. With this approach, a proper number of feature vectors for each class is generated by the algorithm. In supervised learning, the training process should be unbiased to reach more accurate results in testing. In the BFDA, accuracy is related to the initialization of the border feature vectors and the input ordering of the training samples. These dependencies make the classifier a biased decision maker. Consensus strategy can be applied with cross validation to reduce these dependencies. In this study, major performance analysis and comparisons were made by using the AVIRIS data. Using the BFDA, we obtained satisfactory results with both multispectral and hyperspectal data sets. The BFDA is also a robust algorithm with the Hughes effect. Additionally, rare class members are more accurately classified by the BFDA as compared to conventional statistical methods.