Hyperspectral image segmentation for estimation of biomass at reclaimed heaps

0423994 - ÚVGZ 2014 RIV CZ eng C - Konferenční příspěvek (zahraniční konf.)
Pikl, Miroslav - Zemek, František
Hyperspectral image segmentation for estimation of biomass at reclaimed heaps.
Global Change and Resilience: From Impacts to Responses : Proceedings of the 3rd annual Global Change and Resilience Conference. Brno: Global change research centre, Academy of Sciences of the Czech Republic, v. v. i, 2013 - (Stojanov, R.; Žalud, Z.; Cudlín, P.; Farda, A.; Urban, O.; Trnka, M.), s. 200-203. ISBN 978-80-904351-8-6.
[Global Change and Resilience. Brno (CZ), 22.05.2013-24.05.2013]
Grant CEP: GA MŠMT(CZ) ED1.1.00/02.0073; GA MŠMT OC09001; GA MŠMT(CZ) LM2010007
Institucionální podpora: RVO:67179843
Klíčová slova: hyperspectral * classification * maximal likehood * neural network
Kód oboru RIV: EH - Ekologie - společenstva

This paper presents the preliminary results from a study that aims at estimation of above ground biomass and soil carbon content at reclaimed mining heaps in the Sokolov region. Two image segmentation methods are presented. We applied maximal likelihood (ML) and neural network (NN) classifi ers on airborne hyperspectral data. Th e objective of this part of the study was to prepare a land cover classifi cation of the region. Th e main focus was paid to discrimination of six classes with prevailing forest species cover. Th e classifi cation accuracy of the training sites was 93.75 % for NN and 79.12 % for ML respectively. But ML outperformed NN in overall classifi cation accuracy with 61.54 % compared to 40.9 % of NN. Th e more accurate results of the ML classifi er are probably infl uenced by properties of the training samples. Th e larger size of the training samples derived for ML enabled better representation of class histograms. Th e lower overall NN accuracy could result from high spatial resolution of HS data.
Trvalý link: http://hdl.handle.net/11104/0230119