Robust forest cover indices for multispectral images

Authors: BECKER, S.J - National Geospatial-Intelligence Agency; Daughtry, Craig; Russ, Andrew - Andy

Submitted to: Photogrammetric Engineering and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/1/2018
Publication Date: 8/6/2018
Citation: Becker, S., Daughtry, C.S., Russ, A.L. 2018. Robust forest cover indices for multispectral images. Photogrammetric Engineering and Remote Sensing. 84:505-512. https://doi.org/10.1201/9781420055139.
DOI: https://doi.org/10.1201/9781420055139

Interpretive Summary: Trees occur in forest and non-forest land cover classes and provide significant ecosystem services, including watershed protection, wildlife habitat, and recreational uses. Trees in non-forest land cover classes are also important for such diverse applications as planning routes for utilities, roads, and trails, monitoring land cover changes, modeling environmental quality, and enhancing quality of life. Remotely sensed images are often used to create thematic maps of land cover at a range of spatial scales, but accurately identifying trees in mixed land-use scenes is often challenging. The study site in central Maryland included coniferous and deciduous trees associated with agricultural fields and pastures, residential and commercial buildings, roads, parking lots, and wetlands. Using multispectral images acquired by the WordView-2 satellite, we developed two forest cover indices and protocols that accurately and reliably classified trees in the study site throughout the year. These forest cover indices were robust and could be used to rapidly create and update land cover maps. Additional research is required to evaluate these forest cover indices for other ecosystems as well as other multispectral sensors.

Technical Abstract: Trees occur in many land cover classes and provide significant ecosystem services. Remotely sensed multispectral images are often used to create thematic maps of land cover, but accurately identifying trees in mixed land-use scenes is challenging. We developed two forest cover indices and protocols that reliably identified trees in WorldView-2 multispectral images. The forest cover indices exploited the product of either the reflectance in red (630-690 nm) and red edge (705-745 nm) bands or the product of reflectance in red and near infrared (770-895 nm) bands. For two classes (trees vs. all other), overall classification accuracy was >85% for the four images that were acquired in each season of the year. Additional research is required to evaluate these indices for other scenes and other sensors.