An interesting strategy for deriving tree canopy using high resolution images is illustrated in the following publication:
A strategy was developed for deriving tree canopy density at a spatial resolution of 30 m. This strategy relies on high resolution images for reference data development and uses regression tree and multiple linear regression to model tree canopy density from Landsat 7 ETM+ images. The applicability of this strategy was demonstrated in three areas of the United States, each of the size of the mosaic of two ETM+ scenes. The results were relatively consistent in the three study areas. The 1 m DOQ imagery proved a valuable source for deriving reference tree canopy density data. Tree canopy was separable from non-canopy surfaces using a decision tree classifier. The regression tree was found more robust than multiple linear regression for estimating tree canopy density from ETM+ images. The residual error of model prediction arises not only from the complex nature of mixings between tree canopy and non-canopy surface materials, but also from With the increasing availability and decreasing cost of both high resolution and ETM+ images, the developed strategy likely will be applicable in many regions of the world. For operational applications of this strategy over large areas, however, some related issues need to be further investigated. The first relates to uncertainties in the reference data, arising from classifying high resolution images. Knowledge on how such uncertainties translate to errors in the 30 m reference canopy density data and affect the developed canopy density model and its prediction capability should provide guidelines as to what accuracy levels are acceptable in classifying high resolution images. The second issue is how to select the most relevant variables for modeling tree canopy density. In this study we used 7 ETM+ bands of two acquisition dates, which might not be an optimal set of variables for modeling tree canopy density. Using the most relevant variables for model development may lead to simpler models with better prediction capability. We will further investigate these issues in developing the tree canopy density data layer for the NLCD 2000 project.
To view the full publication, please use the following link:
http://landcover.usgs.gov/pdf/canopy_density.pdf
Attached is another publication entitled "DERIVATION OF TREE CANOPY COVER BY MULTISCALE REMOTE SENSING APPROACH " which may interest you.