LiDAR metrics selection and neural network estimation performance

Neural Networks (NN) hold the potential for improving a variety of tasks in remote sensing and image processing. They represent a different approach to problems, as they do not rely on statistical relationships. Instead, neural networks adaptively estimate continuous functions from data without specifying mathematically how outputs depend on inputs. This paper evaluates the effect of metrics selection in NN volume estimation performance on LiDAR metrics. The LiDAR data were acquired from a six-year-old Eucalyptus grandis plantation, in April 2009, under maximum leaf area index. The dependent variable volume was collected from 23 rectangular plots, with 400 m2 each , in July, 2009. The neural networks were designed with 10 neurons on the hidden layer. Input and hidden nodes compute logistic function and the output nodes linear one. The first NN received all the metrics extracted from LiDAR data set: sixty six metrics from all returns and other sixty six from first returns. For the second NN the data set was pruned, removing count and weak metrics. To the third NN were presented metrics selected by correlation rank. The RMSE analysis indicated NN adjusted based on the top 20 metrics as the best fitted neural network, followed by the NN adjusted based on pruned data and by NN adjusted to all metrics as input. Prune metrics increases the NN estimation capacity and the tolerance to unstable metrics as well. However, very intensive prunes, e.g. like performed in top 20 metrics selection, can result in overfitting, deteriorating the NN estimation capacity.