Estimating forest canopy attributes via airborne, high-resolution, multispectral imagery in midwest forest types

An investigation of the utility of high spatial resolution (sub-meter), 16-bit, multispectral, airborne digital imagery for forest land cover mapping in the heterogeneous and structurally complex forested landscapes of northern Michigan is presented. Imagery frame registration and georeferencing issues are presented and a novel approach for bi-directional reflectance distribution function (BRDF) effects correction and between-frame brightness normalization is introduced. Maximum likelihood classification of five cover type classes is performed over various geographic aggregates of 34 plots established in the study area that were designed according to the Forest Inventory and Analysis protocol. Classification accuracy estimates show that although band registration and BRDF corrections and brightness normalization provide an approximately 5% improvement over the raw imagery data, overall classification accuracy remains relatively low, barely exceeding 50%. Computed kappa coefficients reveal no statistical differences among classification trials. Classification results appear to be independent of geographic aggregations of sampling plots.; Estimation of forest stand canopy parameter parameters (stem density, canopy closure, and mean crown diameter) is based on quantifying the spatial autocorrelation among pixel digital numbers (DN) using variogram analysis and slope break analysis, an alternative non-parametric approach. Parameter estimation and cover type classification proceed from the identification of tree apexes. Parameter accuracy assessment is evaluated via value comparison with a spatially precise set of field observations. In general, slope-break-based parameter estimates are superior to those obtained using variograms. Estimated root mean square errors at the plot level for the former average 6.5% for stem density, 3.5% for canopy closure and 2.5% for mean crown diameter, which are less than or equal to error rates obtained via traditional forest stand cruising by experienced personnel. The employed methodology entails parsimonious parameterization and is supportive of automation. Overall cover type classification accuracy increases from approximately 70% when using original imagery DNs to over 85% when band registration problems are corrected and variable brightness regimes among imagery frames are normalized. Limiting cover type classification to pixels identified as tree apexes is found to improve traditional classification approaches that use all pixels by 35%.; Image-texture analysis based on intensity co-occurrence provides a quantitative evaluation of second order image texture features that carry discriminatory potential for forest cover type classification purposes. Procedure development and evaluation is based on two independent data sets. Classification accuracies exceeding 60% can potentially be achieved by using only image texture information. In its current level of development, procedure applicability may be limited because of substantial computational cost, absence of computer software for automation, and the complexity of methodologies integral to the feature selection process.