Evaluation of discriminative and generative classification models
for tree species classification
The main objective of this work was the evaluation of four selected classification models of discriminative (AdaBoost, Self Organizing Maps and Support Vector Machines) and generative (Gaussian Maximum Likelihood) nature regarding their suitability for tree species classification including the information of the according development stage (qualification, dimensioning and maturing). For this, a classification setup, which allows a comparative and transparent analysis of each of the classification performances in a bootstrap environment as well as the classification of raster data, was created in the R programming language.
According to the No Free Lunch Theorem (Wolpert & Macready, 1995), there is no classification model to favor in general for any given classification problem. However, in this study the specific problem of classifying the five major tree species European Beech, Pedunculate and Sessile Oak, Norway Spruce, Douglas Fir and Scots Pine divided each into the respective development stages within the study site was investigated, for which it can be stated that all discriminative classification models outperformed the generative model.
Considering all examined factors, such as classification accuracy and precision, parameterization effort or computational time, AdaBoost was found to yield highest accuracies and precision as well as the most comfortable implementation procedure in terms of the classification setup in R and can therefore be named as the most suitable classification model for the given application. The strength of the ensemble based AdaBoost is caused by the simple principle of the weighted learning from misclassifications during the classification process. The inclusion of the AdaBoost in a spatially adaptive classification approach, as it was developed by Stoffels et al. (2012), should be investigated on further improvements in following studies.
The SOM achieved the second best results regarding the classification accuracies and precision, but is characterized by exhaustive implementation efforts, mainly referring to the parameterization optimization. Furthermore, limitations of the implementation in the R environment restricted the applicability of SOM for the given classification problem. As a consequence, the usage of the SOM algorithms included in the R package kohonen is not recommended within a classification framework based on large datasets.
Against the expectations, the SVM performed insufficiently in terms of classification accuracy. It is assumed that the major challenge for the SVM algorithm in this case is the imbalance of the class reference distributions. Besides, it is doubtful, if appropriate parameter values were found, although wide ranges of potential values for the required parameter were assessed during the parameter tuning. These two major issues should be subject to further investigations.
The generative GML achieved lowest classification accuracies for the discrimination of the stand age classes. This is definitely owed to the fact, that the conceptual requirements of the GML were not given. The presence of overlapping and non-normally distributed class probability densities significantly reduces the GMLs´ capability of performing sufficiently. Moreover, it must be assumed, that the reference data base is - to some extent - biased. Therefore, the powerful maximum likelihood principle with nonparametric class model estimations, e.g. a kernel density estimation, might be more adequate for real-world applications, for which the class-wise reference data base is known not to fulfill the requirements of a selected parametric model estimation.
The problem of tree species identification through image classification using multispectral remote sensing data is a complex one involving four key elements: spatial and spectral resolution of the raster data, reference data base, classification model and forest characteristics (Gosh et al., 2014). However, this study focuses the third key element, the investigation of discriminative and generative classification models. The results revealed that the required accuracies for forest management systems and extended research attributed to forest classification highly depend on the operational classification model. Nevertheless, different raster data, reference data or another study area will not ultimately follow the conclusions of this thesis. In general, a classification problem has always to be considered as a complex framework with interdependent factors. Optimized key elements regarding the classification problem, such as the availability of satellite observations, which cover large areas within the optimum phenological stages and show optimized spatial and higher spectral resolution for forest applications, as it is provided by the recently launched Sentinel-2 satellite, will complement the integration of image classification products into operational forest inventory and monitoring systems.