Akademik Veri Yönetim Sistemi
Turkish Journal of Earth Sciences, cilt.35, sa.3, ss.217-235, 2026 (SCI-Expanded, Scopus, TRDizin)
Hyperspectral image (HSI) classification is of critical importance in many fields including agriculture, geology, environmental monitoring, and urban planning. In recent years, many researchers have utilized deep neural networks (DNNs), known for their high performance in the classification of HSIs. When 2-D/3-D convolutional neural networks are used in HSI classification, filters are applied using input patches typically larger than 11 × 11. This allows spectral and spatial features to be evaluated together. However, this combination creates several problems. Because HSIs have low spatial resolution, they often do not contain strong texture details. Furthermore, features with little relevance to classification make the feature vectors spread out in the input space. More importantly, when large input patches and high sampling rates are used, the training set may implicitly include the test data. To overcome these issues, a new 1-D DNN framework is proposed in this study instead of 2-D DNNs. A novel deep learning model focusing on the identification of features is applied. Features related to the images are extracted with fully connected neural networks trained with Walsh vectors and the classification process is carried out with a minimum distance network. With spectral data alone, the proposed 1-D DNN model achieves average accuracy of 97% on the Indian Pines, Salinas, Pavia Centre, Pavia University, and Botswana datasets. It is observed that, compared to 2-D DNNs, the proposed 1-D DNN achieves high accuracy while avoiding overlap problems and unnecessary complexity, making it a simpler and more reliable choice for HSI classification.