Analysis of the effect of training sample size on the performance of 2D CNN models

Abstract

Hyperspectral remote sensing plays a significant role in the research of Earth observation owing to rich spectral information. Convolutional Neural Networks have been commonly used in hyperspectral image classification with the rapid development of deep learning algorithms. In this study, the effect of sample size on the performance of 2D CNN models was analyzed using freely available Pavia hyperspectral data for a 9-class classification problem. Thematic maps were produced with different number of samples and the accuracies of the thematic maps were compared. The results were verified for the effectiveness of different number of samples considering accuracy metrics (overall accuracy, F-score and Kappa coefficient). As a result, overall accuracies of 86.42, 91.84, 94.20 and 95.36% were produced for Deep Learning models using 50, 100, 200 and 400 samples, respectively.