Spectral Stability of Human-Modified Landscapes

Christopher Small, Columbia University, cs184@columbia.edu (Presenter)
Cristina Milesi, NASA Ames, cristina.milesi@gmail.com

Two of the principal challenges of mapping land cover with decameter-resolution broadband sensors are spectral mixing and spectral non-uniqueness. Spectral mixing occurs when reflected radiance from distinct land cover components is spatially averaged within the Instantaneous Field Of View (IFOV) of a single pixel. The result is a loss of spatial detail that complicates identification of the individual land cover components blurred into a single pixel spectrum. Spectral non-uniqueness occurs when the radiance field of distinct land cover components are spectrally averaged over the wavelengths of a single spectral band. The result is a loss of spectral detail that can make distinct land cover components spectrally indistinguishable. We attempt to address both of these challenges simultaneously by combining standardized spectral mixture models with temporal moments of spectral endmember fractions. At optical wavelengths, spectral properties of land cover can be represented using standardized spectral endmember fractions to represent combinations of the most spectrally and functionally distinct components of land cover; rock, soil and impervious substrates, vegetation, water and shadow. The spectral similarity of many soils and impervious substrates that makes thematic classifications error prone can be resolved by using multi-season composites of spectral endmembers. Temporal variability of reflectance can distinguish spectrally stable impervious substrates from temporally variable soil reflectance resulting from seasonal changes in moisture content (thus albedo) and fractional vegetation cover. The spectral similarity of shadow and water can also be resolved by differences in their temporal variability resulting from seasonal changes in solar illumination and shadow fraction. We use standardized spectral mixture models to represent land cover as linear mixtures of soil and impervious Substrate, Vegetation and spectrally Dark surfaces in Landsat imagery acquired in different seasons over the course of a year. The temporal mean and standard deviation of the resulting SVD components provide supplementary information on the temporal stability of each spectral endmember, thereby distinguishing land cover types that may be spectrally indistinguishable in single date imagery. We illustrate the spectral and temporal characteristics of a wide variety of land cover types using spectral stability spaces defined by seasonal mean and variability of each spectral endmember. The resulting spectral stability spaces clearly discriminate soils and impervious surfaces that are generally spectrally indistinguishable in individual scenes.

Presentation Type:  Poster

Session:  General Contributions   (Tue 4:35 PM)

Associated Project(s): 

• Milesi, Cristina: Mapping of Urban Expansion Using Multi-Decadal Landsat and Nightlights Data Over North America ...details

Poster Location ID: 119