%0 Conference Proceedings
%4 sid.inpe.br/mtc-m18/2012/05.16.17.29
%2 sid.inpe.br/mtc-m18/2012/05.16.17.29.11
%@isbn 978-85-17-00059-1
%T Object-based cloud and cloud shadow detection in Landsat images for tropical forest monitoring
%D 2012
%A Zortea, Maciel,
%A Salberg, Arnt-Børre,
%A Trier, Øivind Due,
%@electronicmailaddress maciel.zortea@nr.no
%@electronicmailaddress arnt-borre.salberg
%@electronicmailaddress oivind.due.trierg@nr.no
%E Feitosa, Raul Queiroz,
%E Costa, Gilson Alexandre Ostwald Pedro da,
%E Almeida, Cláudia Maria de,
%E Fonseca, Leila Maria Garcia,
%E Kux, Hermann Johann Heinrich,
%B International Conference on Geographic Object-Based Image Analysis, 4 (GEOBIA).
%C Rio de Janeiro
%8 May 7-9, 2012
%I Instituto Nacional de Pesquisas Espaciais (INPE)
%J São José dos Campos
%P 326-331
%S Proceedings
%1 Instituto Nacional de Pesquisas Espaciais (INPE)
%K Cloud detection, shadows, classification, segmentation, Landsat.
%X Clouds and cloud shadows often obscure parts of images acquired by optical space-borne sensors. The clouds and cloud shadows need to be detected and labeled as missing data. This enables subsequent methods to make their own decisions about how the missing data should be handled. Here we propose an automatic method to detect daytime cloud and cloud shadows in the context of tropical forest monitoring. In particular, we focus on Landsat 5 TM and Landsat 7 ETM+ images. In addition to the original bands, we investigate the use of additional spectral-derived features, based on pixel-wise differences, ratios, and maximum values derived for all combinations of pairs of top-of-the atmosphere reflectance bands. The subset of features retained for classification, and the boundaries of the classes in the feature space, were identified by optimizing the accuracy of the proposed method using samples collected from spatially disjoint scenes, acquired in different time periods, in an attempt to increase the generalization capability of the proposed approach when applied to unseen scenes. When a new image is to be classified, the idea is to first segment it locally using the Statistical Region Merging algorithm (Nock and Nielsen, 2004). Cloud and cloud shadow masks are then obtained by classifying the averaged pixel values, inside each segment, instead of individual pixels. Finally a simple cloud shape matching algorithm is used to reduce false detection of cloud shadow areas. We found that the proposed object-based technique reduces the spatial noise of the final classified map when compared to traditional single pixel classification. The accuracy of the proposed method appears to be comparable to two alternative algorithms selected for benchmark purposes.
%9 Forest Analysis
%@language en
%3 094.pdf