Band selection using hyperspectral data from airborne and satellite sensors

Abstract

Hyperspectral data offer refined spectral discrimination of ground targets, but come at a substantial cost.  For some sensors, the number of spatial pixels (swath width) needs to be reduced to acquire a large number of bands.  In addition, coarser spatial resolution is required to achieve enough signal from narrow bands.  This study aimed to investigate whether it was possible to reduce the number of bands and broaden their widths, while achieving the same or higher application accuracy as with hyperstructural data.

Three innovative band selection methods were developed as tools for this investigation.  They were designed primarily for Maximum Likelihood Classification (MLC) applications, but their use with respect to other applications was discussed.  All algorithms aimed to optimise the band location, width and number with respect to the MLC accuracy for the given classification task.  The supervised band selection (SBS) algorithm is based on conventional feature selection techniques, while the unsupervised band selection (UBS) method aims to decorrelate the band set.  The unsupervised clustering-based SBS (CSBS) algorithm uses the SBS, but with classes being defined by clustering.

The three approaches were evaluated on real data sets.  All algorithms gave physically meaningful band sets, which achieved similar or higher MLC accuracies than band sets of current airborne and satellite sensors.  The sub-optimality of the SBS bands was found to be least (7%) for sets with maximum three bands.  The band number criteria were shown to be effective estimates of the intrinsic data dimensionality, although some subjectivity remains.  Only SBS may be used to test whether narrow band data have a significant advantage over broad band data.  UBS depends on dark image data for band expansion and requires each band to be normally distributed, which is only justified if the scene is made up of a single material type.  CSBS has the drawback of producing inconsistent results depending on the initialisation and parameter settings of the clustering routine.

The methods can be applied with programmable sensors in a repeat-pass fashion:  Band selection may be performed on hyperspectral data acquired over a representative part of the scene.  Then, multispectral data may be collected over the same scene with the optimised band set under similar solar and atmospheric conditions.  Both UBS and CSBS may be employed in-flight.  For non-programmable sensors, a more generic band set is sought for a given classification scheme, which needs to be optimised to a large number of scenes.

The author believes that current data acquisition is inefficient in that spectrally redundant data are collected with most imaging spectrometers, often using narrow band data where this is not necessary.  Application of the above band selection methods to real data sets showed that for all three classification tasks, the number of bands to acquire could be reduced dramatically with a maximum loss of 5% in MLC accuracy, and for two out of the three tasks, coarsening the spectral sensor resolution may be justified.  This would allow collecting supplementary data and refining the sensor’s spatial resolution.  Coupled with algorithms to optimise other acquisition parameters, the band selection methods developed in this thesis lead the way towards an intelligent remote sensing expert system for data acquisition.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata