Multiple-level defoliation assessment with hyperspectral data: integration of continuum-removed absorptions and red edges

Abstract

Hyperspectral data were collected from 40 canopies of saltcedar (Tamarix ramosissima): 10 healthy canopies and 30 canopies defoliated by an introduced biological control agent, the saltcedar leaf beetle (Diorhabda carinata). These data assessed multiple-level defoliations in response to the process of biological control. Two important characteristics of the hyperspectral data – red edges and continuum-removed absorptions – were used to discriminate four defoliation categories of saltcedar (healthy plants, newly defoliated plants, completely defoliated plants and refoliating plants) at the canopy level. The red edge positions were located at ranges of 711–716 nm, 706–712 nm, 694–698 nm and 715–719 nm for the four defoliation stages described above, respectively. These red edge positions alone could not clearly judge the four defoliation categories associated with feeding by the beetles. Only the completely defoliated canopies had distinct red edge positions that could be differentiated from the other three types of canopies. While using a classification tree to integrate the red edge positions and their derivatives with the central band depths of these five continuum-removed absorptions, it was found that only two band depths of the continuum-removed absorptions were selected, which were the red absorption between 570 and 716 nm and the water absorption between 936 and 990 nm in the near-infrared region (NIR). This implied that the continuum-removed absorptions outperformed the red edges for identifying the defoliation categories. The resulting overall accuracy was 87.5%. The producer accuracy was 100%, 70%, 100% and 80% for the healthy plants, newly defoliated, completely defoliated plants and refoliating canopies, respectively. The corresponding user accuracy was 90.91%, 77.78%, 100% and 80%. Therefore, we concluded that single spectral data based variable failed to separate the four stages but a combination of the two continuum-removed absorptions located in the blue absorption and the first water absorption in the NIR improved the identification of defoliated canopies associated with the dynamic defoliation process of the biological control agent. This study developed the defoliation detection techniques of commonly used binary levels (i.e. defoliation and non-defoliation) to multiple vegetation defoliation levels. We anticipate applying these assessment techniques to wide-area collections of hyperspectral data covering the two spectral regions as described above to further evaluate the effectiveness of these biological control beetles and their impact on saltcedar management in the Western United States.