Classifying forest pixels using neural networks
The forestry content of the Landsat Thematic Mapper has been investigated in some detail in the literature in the hope that the use of remotely sensed data could reduce the cost and increase the frequency of forest inventories. It has become apparent however that conventional classification methods are inadequate for operational purposes and that new methods are required. In this paper, we investigated the use of neutral networks to approximate raw biophysical data: height; basal area density; and stem density. A multilayer perceptron with a simple architecture trained using a backpropagation algorithm was used to approximate height, basal area and stem density of test pixels. The architecture of the network and the scaling of the data were varied to determined the effect on the results. The effect of two different methods for dividing the data into training and testing were also investigated. The results indicate that scaling and architecture had little influence on the results whereas the division of the data had a definite influence. Height was well approximated followed by basal area and stem density.
Bibliographic Reference: Paper presented: Remote Sensing Society Conference 1996, Durham (GB), September 12-14, 1996
Availability: Available from (1) as Paper EN 40045 ORA
Record Number: 199611057 / Last updated on: 1996-09-30
Original language: en
Available languages: en