Construction and application of characteristic bands of typical land cover based on spectrum-photometric method

• Published in: 2010 18th International Conference on Geoinformatics pp. 1 - 6
• Main Authors: Wenfang Lin, Quanfang Wang, Shuping Zha, Jiayong Li
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-06-2010
• Subjects: Accuracy; Construction of characteristic band; Data mining; Indexes; Object-oriented Classification Approach; Remote sensing; Reservoirs; Rivers; Spectral Analysis; Spectrum-photometric Method; SPOT5; Vegetation mapping
• ISBN: 1424473012; 9781424473014
• ISSN: 2161-024X
• DOI: 10.1109/GEOINFORMATICS.2010.5567712

With high spatial resolution, SPOT5 satellite imagery contains rich texture information, such as geometry and color texture which embodies the details of object surfaces by changing the hue or brightness. However, traditional statistical classification technology based on the spectrum of pixels couldn't make full use of the texture information of SPOT5 satellite imagery because SPOT5-XS imagery only includes four multispectral bands, i.e. green(G) with wavelengths ranging from 0.50 to 0.59 microns, red(R) 0.61 to 0.68, near infrared (NIR) 0.78 to 0.89 and shortwave infrared (SWIR) 0.15 to 1.75. If only these four bands are applied to identify land cover by using remote sensing data, it will often bring some problems, such as the divisibility among various land cover types reduced and the automated information extraction more difficult, which mainly results in the lack of band information and complexity of spectrum. By using SPOT5 imagery covering Taibai Mountain in Shaanxi Province, China, we analyzed spectrum characteristic, relations among bands and texture feature of typical land cover types including evergreen coniferous forest, deciduous broad-leaved forest, uncovered rock riverbed, dry land, residential area, etc. Based on these results, many characteristic bands were built through Band Operation method and merged into the original SPOT5-XS imagery, which includes Normalized Difference Water Index (NDWI), Normalized Difference Vegetation Index (NDVI), Normalized Difference soil Index (NDSI), Difference Vegetation Index (DVI), Adjusted Normalized Difference Vegetation Index (ANDVI), Sum of Vegetation Index (SVI) and two texture bands. Object-oriented classification approach was then applied to the new generated imagery based on decision trees and Error Matrix for accuracy assessment. Experimental results show that the Overall Classification Accuracy (OCA) is 92.22% and Overall Kappa Statistics (Kappa) 89.43%. The classification accuracy of evergreen coniferous forest, deciduous broad-leaved forest and dry land is 95.20%, 94.04% and 93.10%, respectively. These results indicated that the object-oriented method based on the new constructed spectrum features could take advantage of spatial information of high resolution imagery, improve the divisibility among various land cover types and make the information extraction more easily and the high classification accuracy obtained. In particularly, with the texture feature bands, the objects with similar spectrum could be well identified.