Distributional modelling in forestry and remote sensing
The use of distributional models in forestry is investigated, in terms of their capability of modelling distributions of forest mensurational attributes, for modelling and inventory purposes. Emphasis is put on: (i) the univariate and bivariate modelling of tree diameters and heights for stand-level modelling work, and (ii) heuristic methods for use and analysis of distributions which occur in multi-temporal EO imagery, (for the inventory-related tasks of land-use mapping, change detection and growth modelling). In univariate distribution modelling, a new parameterization of the widely-used Johnson’s SB distribution is given, and new Logit-Logistic, generalised Weibull and the Burr system (XII, III, IV) models are introduced into forest modelling. The Logit-Logistic distribution is found to be the best among those compared. The use of regression-based methods of parameter estimation is also investigated. In the domain of bivariate distribution modelling of tree diameters and heights the Plackett method (a particular form of copula) is used to construct Plackett-based bivariate Beta, SB and Logit-Logistic distributions, (the latter two are new), which are compared with each other and the SBB distribution. Other copula functions, including the normal copula, are further employed (for the first time in forest modelling) to construct bivariate distributional models. With the normal copula, the superiority of the Logit-Logistic in the univariate domain is extended into the bivariate domain. To use multi-temporal EO imagery, two pre-processing procedures are necessary: image to image co-registration, and radiometric correction. A spectral correlation-based pixel-matching method is developed to “refine” manually selected control points to achieve very accurate image co-registration. A robust non-parametric method of spectral-distribution standardization is used for relative radiometric correction between images. Finally, possibilities for further research are discussed.