<![CDATA[Land cover information is one of the basic variables in determining phenomena on the earth's surface as well as a measuring variable for the development of an area. Land cover with hyperspectral imagery can provide accurate information on the earth's surface. Limited labeled samples and imbalances data are common in land cover hyperspectral image classification. However, the data that available publicly is very limited makes it difficult to evaluate the performance of existing algorithms. Various methods were developed with several public datasets using as few training samples as possible to produce a method that is robust against sample limitations and imbalances data, but these studies have complicated parameters and the accuracy results are not optimal. In this study, several combination of Multiple Spectral Resolution (MSR), 3D Spectral Dilated Convolution (SDC) and Hybrid Convolutional Neural Network (HybridSN) were carried out to increase accuracy in hyperspectral image classification with very small training samples and simpler parameters and computations. This method was tested on two public datasets Indian Pines (IP) and Salinas (SA) with 5% and 1% training data respectively (only 500 training samples) and showed high accuracy results in land cover hyperspectral image classification. The best results were obtained with a combination of SDC and HybridSN methods which were optimized with Adam with 96.58% % Overal Accuracy (OA), 87.83% Average Accuracy (AA) and 96.09% kappa on the IP dataset, and 99.08% Average Accuracy (AA), 99.00% kappa on the SA dataset]]>
Copyrights © 2021
