Improved mapping of forest type using spectral-temporal Landsat features

TitleImproved mapping of forest type using spectral-temporal Landsat features
Publication TypeJournal Article
Year of Publication2018
AuthorsPasquarella, Valerie J., Holden Christopher E., and Woodcock C.E.
JournalRemote Sensing of Environment
Pagination193 - 207
Date PublishedJan-06-2018
Keywordsclassification, Forest type, Landsat, Spectral-temporal, time series

Multi-spectral imagery from the Landsat family of satellites has been used to map forest properties for decades, but accurate forest type characterizations at a 30-m Landsat resolution have remained an ongoing challenge, especially over large areas. We combined existing Landsat time series algorithms to quantify both harmonic and phenological metrics in a new set of spectral-temporal features that can be produced seamlessly across many Landsat scenes. Harmonic metrics characterize mean annual reflectance and seasonal variability, while phenological metrics quantify the timing of seasonal events. We assessed the performance of spectral-temporal features derived from time series of all available observations (1985–2015) relative to more conventional single date and multi-date inputs. Performance was determined based on agreement with a reference dataset for eight New England forest types at both the pixel and polygon scale. We found that spectral-temporal features consistently and significantly (paired t-test, p ≪ 0.01) outperformed all feature sets derived from individual images and multi-date combinations in all measures of agreement considered. Harmonic features, such as annual amplitude and model fit error, aid in distinguishing deciduous hardwoods from conifer species, while phenology features, like the timing of autumn onset and growing season length, were useful in separating hardwood classes. This study represents an important step toward large-scale forest type mapping using spectral-temporal Landsat features by providing a quantitative assessment of the advantages of harmonic and phenology features derived from time series of Landsat data as compared with more conventional single-date and multi-date classification inputs.

Short TitleRemote Sensing of Environment