eCite Digital Repository

A robust multi-kernel change detection framework for detecting leaf beetle defoliation using Landsat 7 ETM+ data


Anees, A and Aryal, J and O'Reilly, MM and Gale, TJ and Wardlaw, T, A robust multi-kernel change detection framework for detecting leaf beetle defoliation using Landsat 7 ETM+ data, ISPRS Journal of Photogrammetry and Remote Sensing, 122 pp. 167-178. ISSN 0924-2716 (2016) [Refereed Article]

Copyright Statement

Copyright 2016 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

DOI: doi:10.1016/j.isprsjprs.2016.10.011


A robust non-parametric framework, based on multiple Radial Basic Function (RBF) kernels, is proposed in this study, for detecting land/forest cover changes using Landsat 7 ETM+ images. One of the widely used frameworks is to find change vectors (difference image) and use a supervised classifier to differentiate between change and no-change. The Bayesian Classifiers e.g. Maximum Likelihood Classifier (MLC), Naive Bayes (NB), are widely used probabilistic classifiers which assume parametric models, e.g. Gaussian function, for the class conditional distributions. However, their performance can be limited if the data set deviates from the assumed model. The proposed framework exploits the useful properties of Least Squares Probabilistic Classifier (LSPC) formulation i.e. non-parametric and probabilistic nature, to model class posterior probabilities of the difference image using a linear combination of a large number of Gaussian kernels. To this end, a simple technique, based on 10-fold cross-validation is also proposed for tuning model parameters automatically instead of selecting a (possibly) suboptimal combination from pre-specified lists of values. The proposed framework has been tested and compared with Support Vector Machine (SVM) and NB for detection of defoliation, caused by leaf beetles (Paropsisterna spp.) in Eucalyptus nitens and Eucalyptus globulus plantations of two test areas, in Tasmania, Australia, using raw bands and band combination indices of Landsat 7 ETM+. It was observed that due to multi-kernel non-parametric formulation and probabilistic nature, the LSPC outperforms parametric NB with Gaussian assumption in change detection framework, with Overall Accuracy (OA) ranging from 93.6% (κ = 0.87) to 97.4% (κ = 0.94) against 85.3% (κ = 69) to 93.4% (κ = 0.85), and is more robust to changing data distributions. Its performance was comparable to SVM, with added advantages of being probabilistic and capable of handling multi-class problems naturally with its original formulation.

Item Details

Item Type:Refereed Article
Keywords:change detection, Landsat 7 ETM+, least squares probabilistic classifier, naive Bayes classifier, supervised classification
Research Division:Engineering
Research Group:Geomatic engineering
Research Field:Photogrammetry and remote sensing
Objective Division:Environmental Management
Objective Group:Terrestrial systems and management
Objective Field:Assessment and management of terrestrial ecosystems
UTAS Author:Anees, A (Mr Asim Anees)
UTAS Author:Aryal, J (Dr Jagannath Aryal)
UTAS Author:O'Reilly, MM (Associate Professor Malgorzata O'Reilly)
UTAS Author:Gale, TJ (Dr Timothy Gale)
UTAS Author:Wardlaw, T (Dr Timothy Wardlaw)
ID Code:112641
Year Published:2016
Web of Science® Times Cited:14
Deposited By:Geography and Spatial Science
Deposited On:2016-11-21
Last Modified:2017-10-25

Repository Staff Only: item control page