Mortalitätsschätzungen in ungleichaltrigen Fichtenwäldern mit Hilfe Neuronaler Netze

Abstract

Within forest growth modeling it is understood that individual tree mortality can be captured realistically by relating the average rate of mortality to a few reliable and measurable size or site characteristics using a LOGIT model. In this paper we describe the application of neuronal networks adhering to the unsupervised learning paradigm to predict individual tree mortality. Using the large and representative Norway spruce data sample from the Austrian National Forest Inventory, we train different types of neural network architectures, namely Multi-Layer Perceptron, Cascade Correlation, and Learning Vector Quantization. For training, we use the following learning rules: Error Backpropagation, Resilient Propagation, and Scaled Conjugate Gradient. With an independent data set we evaluate the neural network types to predict individual tree mortality.