Artificial Neuronal Networks
In this book, an easily understandable account of modelling methods with artificial neuronal networks for practical applications in ecology and evolution is provided. Special features include examples of applications using both supervised and unsupervised
- PDF / 27,722,017 Bytes
- 275 Pages / 439.32 x 666.12 pts Page_size
- 23 Downloads / 203 Views
Springer-Verlag Berlin Heidelberg GmbH
Sovan Lek . Jean-Fran 5 -> 1 (#inputs,#hidden nodes, #outputs). The inputs were TM bands 3, 4, 5, elevation, slope and aspect. The number of pixels/points shown are 3555 for the testing data. The RMSE and R2 values for the testing data were 5.6 and 0.69, respectively
ture channels, a 4 ---t 17 ---t 1 neuronal network predicted secondary forest age over a 9 year range with an RMSE of 2.0 years and R2(actual vs. predicted) of 0.38. The corresponding multiple linear regression employing 3 SPOT bands and 4 texture channels had an RMSE of 2.1 years and an R2 of 0.31. Though neither technique could be said to accurately estimate secondary forest age, neuronal networks consistently outperform parametric, linear discriminant and regression procedures using fewer spectral and textural bands. Additional work by Nelson et al. (2000) near the same area in Rondonia using Thematic Mapper multispectral data verifies these findings. For instance, a linear discriminant function using four spectral/textural measures differentiated primary forest, nonforest, and secondary forest with an overall accuracy of 96.6%. Using 3 channels (3 ---t 3 ---t 3), the comparable neuronal net yielded an overall accuracy of 97.2%. The TM spectral and textural data were also used to estimate secondary forest age. The multiple linear predictive regression utilized 6 spectral-texture channels and yielded an RMSE of 1.62 years and an R2 (actual vs. predicted) of 0.35. The comparable neuronal net, using 4 channels, had an RMSE of 1.59 years and an R2 value of 0.37. The differences between the linear and neuronal net results are small, but in this and the Kimes et al. (1999) studies, they are consistent. In general, using fewer bands, utilizing automatic variable selection procedures, and utilizing automatic weighting procedures, neuronal
CHAPTER 2
•
Predicting Ecologically Important Vegetation Variables
37
net results are, in general, comparable to or better than linear discriminant and linear regression results. Ultimately, the scientific community needs to develop physically-based radiative scattering models for the above areas of research. These models need to be accurate and invertible for the desired variables. In research areas where these activities are immature, the neuronal network approach can provide an accurate initial model for predicting vegetation variables. 2.4.2 Neuronal Networks as Baseline Control
A network can be used as a baseline control while developing adequate physicallybased models (Fu 1994). Where adequate field and ground truth data sets exist, a neuronal network can be trained and tested on these data sets. These networks attempt to find the optimum functional relationships that exist between the input variables and the output variables of interest. The networks can be trained in the forward direction on the field data (e.g. vegetation canopy variables are the inputs and radiative scattering is the output). Improvements to the physically-based model are indicated if it cannot su
