Distinguishing tree plantation species and secondary forests using Landsat and Hymap

Matthew E. Fagan, Columbia University, mef2153@columbia.edu (Presenter)
J. Pablo Arroyo-Mora, McGill University, juan.arroyomora@mail.mcgill.ca
Steven E. Sesnie, US Fish and WIldlife Service, steven.sesnie@nau.edu
Ruth DeFries, Columbia University, rd2402@columbia.edu

Despite their potential importance for restoring connectivity and carbon storage to deforested agricultural landscapes, tree-dominated agricultural habitats have proven difficult to systematically map in the tropics. Timber plantations and other tree-crops (e.g., oil palm, rubber, shade coffee and cacao) are often hard to distinguish from secondary forests using multispectral sensors. We examine the ability of moderate-resolution hyperspectral imagery (Hymap) to improve classification accuracy over moderate-resolution multispectral imagery (Landsat) in the dynamic agricultural region of northern Costa Rica. We hypothesize that 1) hyperspectral imagery will have higher classification accuracy than multispectral Landsat imagery for all land covers, 2) hyperspectral imagery will accurately distinguish tree plantations to species, and 3) degrading hyperspectral imagery from 15 m to 30 m spatial resolution will not affect its classification accuracy. Using the Random Forests classifier and training data from a 2005 field campaign and aerial photography, we classified Landsat ETM+ imagery and aerial Hymap imagery to fourteen land cover classes. Spectral separability of cover classes was high using Hymap imagery (all >1.99, Jeffries-Matusita), with the exception of primary and secondary forests (1.65). The Landsat imagery had lower spectral separability (0.5-1.54), but lumping the five species of tree plantations increased accuracy markedly (all >1.72). In a preliminary analysis on one image, hyperspectral imagery had higher classification accuracy for most land cover classes and distinguished some tree plantation species but not others. Degrading hyperspectral imagery decreased its accuracy. We conclude that hyperspectral imagery improves discrimination of agricultural tree cover compared to multispectral imagery, but some tree plantation species are spectrally indistinguishable from secondary forest with current technology for passive remote sensing.

Presentation Type:  Poster

Session:  Global Change Impact & Vulnerability   (Tue 11:30 AM)

Associated Project(s): 

• DeFries, Ruth: Growing Up Fragmented: Using Hyperspectral Imagery to Improve Estimates of Ecosystem Services from Tropical Reforestation ...details

Poster Location ID: 176