A 3D-Inception CNN for Hyperspectral Image Classification

Abstract

The classification of hyperspectral image (HSI) has attracted significant attention from the research community of remote sensing. HSI analysis suffers from overfitting due to the limited number of labelled trainingsamples. As a result, in order to enhance the performance of the HSI classification task, a better efficient neural networkarchitecture should be developed. To tackle this issue, this letter presents a new 3D-Inception CNN (3D-ICNN) modelfor dynamically extracting features by stacking inception modules in the network that can learn more representativefeatures with fewer training samples by adopting variable spatial size convolutional filters and dynamic CNNarchitecture. The experimental results exhibit that the presented model can modify the network design adaptively andachieve higher classification performance. To establish the efficiency and robustness of the presented model, theexperiments are conducted on the publicly available benchmark data sets and also on the new data sets. The proposed3D-Inception CNN model obtained accuracies of 86.25% on Ahmedabad-1(AH1) dataset, 80.30% on Ahmedabad-2(AH2) dataset, 99.95% on the Pavia University (PU) dataset, 99.86% on the Salinas (SA) dataset, and 99.89% on theIndian Pines (IP) dataset