| dc.description.abstract |
Purpose: In this study, with 3D CNN, it is aimed to extract both spectral and spatial features of hyperspectral remote sensing images simultaneously. In this way, the structural features of 3D hyperspectral images are fully utilized. Theory and Methods: In this study, a new 3D CNN method for hyperspectral remote sensing image classification is proposed. 3D CNN extracts spectral and spatial features from hyperspectral remote sensing images simultaneously to achieve better classification accuracy. The 3D CNN method chooses the neighborhood block as the input of the network model. PCA is applied on the hyperspectral image cube as a pre-processing step to remove the spectral band redundancy. Then, the hyperspectral cube is divided into small overlapping 3D patches. These patches are processed to create 3D feature maps using the 3D core function on multiple adjacent bands to preserve common spatial spectral information. Finally, the resulting deep classifier model is being trained endto- end. Results&Conclusion: Experimental studies were conducted on indian pines, pavia university, salinas and kennedy space center datasets to evaluate the performance of the proposed 3D CNN method. Overall accuracy, average accuracy and Kappa values for the Indian pines data set were obtained as 99.93%, 99.72% and 99.92%, respectively. It was achieved as 99.99%, 99.96% and 99.99% for Pavia University. For the Salinas data set, 100% results were found in all evaluation metrics. For the Kennedy space center dataset, 99.81%, 99.68% and 99.78% results were obtained. The proposed 4CF-Net method was compared with 7 hyperspectral image classification methods based on deep learning. Experimental studies show that the proposed 4CF-Net method achieves the best overall accuracy, average accuracy, and kappa coefficient in all four data sets. © 2022 Gazi Universitesi Muhendislik-Mimarlik. All rights reserved. |
|