Evaluating the effects of shadow detection on quickbird image classification and spectroradiometric restoration

Jindong Wu, Marvin E. Bauer

Research output: Contribution to journalArticlepeer-review

20 Scopus citations

Abstract

Shadows in high resolution imagery create significant problems for urban land cover classification and environmental application. We first investigated whether shadows were intrinsically different and hypothetically possible to separate from each other with ground spectral measurements. Both pixel-based and object-oriented methods were used to evaluate the effects of shadow detection on QuickBird image classification and spectroradiometric restoration. In each method, shadows were detected and separated either with or without histogram thresholding, and subsequently corrected with a k-nearest neighbor algorithm and a linear correlation correction. The results showed that shadows had distinct spectroradiometric characteristics, thus, could be detected with an optimal brightness threshold and further differentiated with a scene-based near infrared ratio. The pixel-based methods generally recognized more shadow areas and with statistically higher accuracy than the object-oriented methods. The effects of the prior shadow thresholding were not statistically significant. The accuracy of the final land cover classification, after accounting for the shadow detection and separation, was significantly higher for the pixel-based methods than for the object-oriented methods, although both achieved similar accuracy for the non-shadow classes. Both radiometric restoration algorithms significantly reduced shadow areas in the original satellite images.

Original languageEnglish (US)
Pages (from-to)4450-4469
Number of pages20
JournalRemote Sensing
Volume5
Issue number9
DOIs
StatePublished - 2013

Keywords

  • High resolution
  • Land use and land cover
  • Object-oriented
  • Pixel-based
  • Shadows

Fingerprint

Dive into the research topics of 'Evaluating the effects of shadow detection on quickbird image classification and spectroradiometric restoration'. Together they form a unique fingerprint.

Cite this