Open AccessEvaluating the Performance of the state-of-the-art HybridSN Deep Learning Algorithm for Airborne Hyperspectral Image Classification
Author(s) -
M. A. A. M. Abidin,
Helmi Zulhaidi Mohd Shafri,
M. M. A. Al-Habshi,
Nur Shafira Nisa Shaharum
Publication year - 2021
Publication title -
iop conference series. earth and environmental science
Language(s) - English
Resource type - Journals
eISSN - 1755-1307
pISSN - 1755-1315
DOI - 10.1088/1755-1315/767/1/012019
Subject(s) - hyperspectral imaging , artificial intelligence , vnir , support vector machine , python (programming language) , computer science , convolutional neural network , pattern recognition (psychology) , random forest , kernel (algebra) , radial basis function , deep learning , perceptron , multilayer perceptron , machine learning , artificial neural network , algorithm , mathematics , combinatorics , operating system
This study aims to evaluate the performance of state-of-the-art HybridSN deep learning algorithm versus standard machine learning (ML) and deep learning (DL) techniques using open-source Python libraries for producing hyperspectral land use and land cover (LULC) classification maps. Japanese Chikusei hyperspectral datasets captured by the airborne platform using Hyperspec-VNIR-C sensor were used in this study. Standard ML methods used in this study were support vector machine linear kernel (SVM-linear), support vector machine radial basis function kernel (SVM-RBF) and random forests (RFs) that were provided in Python’s Scikit-learn library. DL techniques used in this study were multilayer perceptron (MLP), two-dimensional convolutional neural network (2-D CNN) and hybrid spectral convolutional neural network (HybridSN), which integrates the 2-D and 3-D feature learning. These DL models were built based on the sequential model using Keras API. The results show that all the proposed methods obtained overall accuracies (OAs) above 95%. The HybridSN and 2-D CNN models gave the best score with 99.97% OAs for hyperspectral image classification using the Chikusei dataset.