Modeling training site vegetation coverage probability with a random optimization procedure: An artificial neural network approach

摘要

The main objective of this study is to examine the feasibility of applying feed-forward neural networks to estimate training site vegetation coverage probability based on past disturbance pattern and vegetation coverage history. The rationale behind this study is the excellent approximation and generalization ability of feed-forward neural networks. The data used to train the networks were collected from Fort Sill, Oklahoma, using the U. S. Army's Land Condition-Trend Analysis (LCTA) standard data collection methodology. The basic unit in this study is a transect point. Spatial independence between transect point's vegetation cover, as well as disturbance, was assumed. Two types of vegetation covers were modeled in this study: ground cover and canopy cover. For both types of vegetation cover, the input vector of a transect point consisted of seven variables, namely, the disturbance in years 1989, 1990 and 1991, the covers in years 1989 and 1991, transect plot's plant community type, and the vegetation's life form. The target output was whether the transect point was covered in year 1991. The actual output from a neural network was regarded as the estimated conditional probability of a transect point having vegetation cover in 1991. Due to the imbalance of representation in the training data, widely used gradient methods @e., back-propagation and its variants) failed to produce any useful results. An adaptive and directional random optimization procedure was subsequently experimented as an alternative training algorithm. The algorithm can be regarded as a hybrid between gradient-based and random search optimization methods. It has a self-adjusting variance term, a directional component and can conduct backward searches. In this study, fixed structure, single hidden layer feed-forward networks with one or two hidden units were employed. The performance of the trained networks were compared to that of logistic regression models based on x{sup}2 goodness-of-fit statistic. All the network trainings were conducted on a Connection Machine CM-2 computer. The results suggested that the neural networks were indeed better than the corresponding logistic models. Though the training data might look pathological to many, data sets with similar characteristics can be found in individual tree mortality modeling and in predicting the likelihood of bankruptcies. The random search nature of the employed algorithm and its speed make the algorithm particular suitable for solving this class of problems. The unique features of the employed algorithm and a version of its parallel implementation are also discussed.