Quantifying forest health trends across time and space: A hyperspectral approach using multi-spectral remote sensing imagery

Background/Question/Methods

Forest health is of increasing concern to land managers due to a convergence of many environmental stressors and increasing pressures from a growing human population. While plot level forest monitoring provides site specific assessments, remote sensing of forest condition is required to quantify the condition of forested ecosystems.  Hyperspectral data contains a wealth of information related to vegetation condition, but its prohibitive cost and a lack of availability make it difficult for landscape scale studies.  Multi-spectral remote sensing instruments have been used for decades to detect forest decline but are typically limited to coarse categories of forest condition (primarily characterized as defoliation classes) for a single species of interest.  In an attempt to merge the information and detail available using hyperspectral techniques with the widespread image availability of multispectral imagery, we tested a hyperspectral approach to forest decline assessment, modified to fit the spectral resolution of Landsat TM5 imagery.  The goal is to determine if it is possible to extract more detailed information on forest health than is typically attempted with multispectral imagery, and if this approach is robust for application across forest type, image locations and acquisition dates.  Calibrated to a range of forest condition including measures of traditional decline symptoms (dieback, transparency, live crown ratio, etc.) as well as early stress symptoms (chlorophyll fluorescence), a Landsat TM5, five-term linear regression (r2 = 0.621, p< 0.0001, RMSE 0.403) was created based on a unique combination of hyperspectral derived vegetation indices. 

Results/Conclusions

When rounded to a class-based system for comparison to more traditional methods, this equation predicted decline across 42 mixed species plots with 63% accuracy for a 10 class rating, and 100% accuracy for a 5 class rating.  This was a significant improvement over the traditional multi-spectral approach using NDVI to predict the same classes (10 class 49% accuracy and 5 class 100% accuracy). Applying this equation to different Landsat scenes, without ground truth calibration input resulted in lower accuracies than the original scene in which the equation was derived (62% accuracy for a 10 class rating and 96% accuracy for a 5 class rating).  Still, this surpasses the accuracy of the traditional multi-spectral approach (10% accuracy for a 10 class rating and 82% accuracy for a 5 class rating), indicating that we can in fact expect more from our multi-spectral analyses.