You are here

Harvard Forest >

Harvard Forest Symposium Abstract 2017

  • Title: Improved mapping of forest species composition using Landsat time series
  • Primary Author: Valerie Pasquarella (University of Massachusetts - Amherst )
  • Additional Authors: Luca Morreale (Boston University); Jonathan Thompson (Harvard Forest)
  • Abstract:

    Accurate spatial data on forest species composition is integral to understanding a wide variety of ecological processes from nutrient cycling and carbon dynamics to host-pest and predator-prey interactions. Characterizing the composition of Northeastern forests is particularly challenging due to relatively high species diversity that reflects latitudinal, elevational and geological gradients as well as land use legacies. There has been interest in using remote sensing imagery to map forest species composition for decades, however few large-scale forest type products exist and user community feedback suggests fewer still are satisfactory for use in local-scale applications.

    Following the opening of the Landsat archive in 2008, there has been rapid innovation in the processing and analysis of time series of all available Landsat imagery. Spectral-temporal features represent an exciting new suite of products that can be derived from dense Landsat time series. Unlike more conventional features, which typically represent reflectance values at discrete points in time, spectral-temporal features are derived from time series of all available Landsat surface reflectance observations for a given pixel, and therefore provide a more robust characterization of land surface properties. There are currently two main types of spectral-temporal features: (1) harmonic metrics, which characterize mean annual and seasonal variability in reflectance, and (2) phenological metrics, which quantify the timing of seasonal events (Figure 1). Previous work has established that spectral-temporal features consistently and significantly outperform single-date and multi-date inputs for classification of relatively pure forest types in western Massachusetts. While this suggests the potential for using the Landsat temporal domain to improve mapping of forest composition, this pilot study was somewhat limited in scope.

    In this project, we extend previous forest composition mapping efforts to the broader New England landscape in an effort to provide improved spatial datasets for both forest management and modeling. Our goal is to use spectral-temporal features as well as other environmental datasets (e.g. climate, topography) produce the most realistic map of forest composition we can achieve while considering the needs of potential users and reproducibility of the approach. We will be testing and comparing two different methods for generating forest type maps: classification and imputation. These two approaches fundamentally differ in their processing and use of training information, and are expected to generate distinctly different map products as a result. We plan to produce a series of classification and imputation maps to be included in map-to-map comparisons. We also intend to complete an accuracy assessment of individual map products, and characterize spatial uncertainty in results.

    We anticipate this study will provide new insights on differences in imputation and classification approaches for mapping forest attributes, as well as issues of scale among training and predictor datasets. Furthermore, the forest composition products generated by this study will be developed with explicit consideration of the needs of the forest modellng community, as well as other groups requiring more detailed information on forested lands. Therefore, we expect products will be more broadly useful, and will enable researchers at Harvard Forest and elsewhere in New England to improve forest inputs to existing models and potentially answer new science questions.

  • Research Category: Regional Studies
    Ecological Informatics and Modelling
    Conservation and Management

  • Figures:
  • 2017_HFabstract_figure.pdf