Developing a Method to Mask Trees in Commercial Multispectral Imagery

Abstract:

The US Army has an increasing focus on using automated remote sensing techniques with commercial multispectral imagery (MSI) to map urban and peri-urban agricultural and vegetative features; however, similar spectral profiles between trees (i.e., forest canopy) and other vegetation result in confusion between these cover classes. Established vegetation indices, like the Normalized Difference Vegetation Index (NDVI), are typically not effective in reliably differentiating between trees and other vegetation. Previous research in tree mapping has included integration of hyperspectral imagery (HSI) and LiDAR for tree detection and species identification, as well as the use of MSI to distinguish tree crowns from non-vegetated features. This project developed a straightforward method to model and also mask out trees from eight-band WorldView-2 (1.85 meter x 1.85 meter resolution at nadir) satellite imagery at the Beltsville Agricultural Research Center in Beltsville, MD spanning 2012 - 2015. The study site included tree cover, a range of agricultural and vegetative cover types, and urban features. The modeling method exploits the product of the red and red edge bands and defines accurate thresholds between trees and other land covers. Results show this method outperforms established vegetation indices including the NDVI, Soil Adjusted Vegetation Index, Normalized Difference Water Index, Simple Ratio, and Normalized Difference Red Edge Index in correctly masking trees while preserving the other information in the imagery. This method is useful when HSI and LiDAR collection are not possible or when using archived MSI.