Title: Classification of hyperspectral images by utilising a lightweight cascaded deep convolutional network

Authors: Sandhya Shinde; Hemant Patidar

Addresses: D Y Patil International University, DYPIEMR, Pune, India ' Department of Electronics and Communication Engineering, Oriental University, Indore, India

Abstract: In recent years, deep learning frameworks have been increasingly used for hyperspectral image classification (HIC) challenges, with outstanding results. Existing network models, on the other hand, are more sophisticated and require longer computing. Traditional HIC algorithms frequently ignore the relationship between local spatial factors. This paper introduces a novel lightweight cascaded deep convolutional neural network (LC-DCNN) that represents the spatial and spectral properties of hyperspectral pictures. The proposed LC-DCNN's performance is tested for several spectral band reduction approaches used to reduce the computational cost of HIC, such as principal component analysis (PCA) and linear discriminant analysis (LDA). The suggested algorithm's efficacy is validated on the Indian Pines and Salinas datasets using accuracy, recall, precision, and F1-score.On the datasheet for IPs, the LC-DCNN+LDA and LC-DCNN+PCA provides overall accuracy of 99.00% and 98.01%, respectively. In the SD, however LC-DCNN+LDA and LC-DCNN+PCA both provide overall accuracy of 99.6% and 98.62%, respectively. The proposed approach provides superior results (99.6% accuracy) compared with traditional state of arts employed previously for the HIC.

Keywords: hyperspectral image classification; HIC; deep learning; convolutional neural network; principal component analysis; PCA.

DOI: 10.1504/IJCSE.2024.139698

International Journal of Computational Science and Engineering, 2024 Vol.27 No.4, pp.434 - 442

Received: 23 Jan 2023
Accepted: 27 Jul 2023
Published online: 05 Jul 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article