Real Time Recurrent Network

Benthic macroinvertebrates were collected at sample sites located in an urbanized stream (the Yangjae stream), a tributary of the Han River in Korea. Figure 7 shows monthly changes in the densities of the abundant groups among the selected taxa collected at the sample sites in the Yangjae stream during the survey period. Input values with greatly different numerical density values were avoided. The data were transformed by natural logarithm in order to emphasize the differences in the low densities. Subsequently, the transformed data were proportionally normalized between 0 and 1 in the range of the maximum and minimum densities for each taxon collected during the survey period. Data collected from April 1996 to March 1997 were used for the training set, while data collected from April 1997 to March 1998 were used as new data for testing the trained network.

In concurrence with the input of biological data, the corresponding sets of environmental data were also provided to the modified RTRN. We trained with environmental data, plus to community data. In this study, the total number of environmental factors was provided as the new external input, but, unlike the community data, neurons accepting environmental factors did not have

"D

10 5

10 5

'<wozorJ

Oligochaeta

- Chironomus

- Chironomidae

- Gastropoda

- Hirudinea

Oligochaeta

- Chironomus

Figure 7 Changes in densities (log-transformed) of selected taxa at the sampling sites in the Yangjae stream, from April 1996 to March 1998.Adapted from Chon T-S, Kwak I-S, ParkY-S, KimT-H, and Kim YS (2001) Patterning and short-term predictions of benthic macroinvertebrate community dynamics by using a recurrent artificial neural network. Ecological Modelling 146: 181-193, with permission from Elsevier a

recurrence feedback. Monthly observations of water velocity and depth, amount of sedimented organic matter, and volume of substrates smaller than 0.5 mm were used as the input. Seven neurons were used for community data for external inputs, additional four neurons for receiving environmental factors, and 13 neurons for hidden nodes. The data for the previous 3 months were given as the input in a sequence with recurrent feedback, while the data for the fourth month were provided as the matching output.

The training data sets were in accord with the matching output. In order to verify the predictability of the trained network, we provided new community data from April 1997 to March 1998. Figure 8 shows the results of the model comparing with field data. Generally, dominant taxa such as Oligochaeta, Chironomus, and Chironomidae showed good matches between the field observations and the predictions from the recurrent networks with community data and with community plus environmental data. Pearson's correlation coefficients between the predicted data and the field data ranged from 0.55 (F = 34, P <0.001) to 0.80 (F = 9.0, P <0.001) when only the community data were used as the input, and ranged from 0.60 (F = 4.0, P<0.001) to 0.94 (F = 32.3, P < 0.001) when both the community and the environmental data were used as the input. The predicted data trained with the community plus environmental data appeared to be closer to field data at the time of data collection (July 1997) when flooding occurred in the Monsoon season.

The trained RTRN was also useful in revealing the environment-community relationships. The sensitivity tests were carried out to show changes in response of different taxa of communities by providing variations to each input value (ranging +50% and —50%) of the environmental variables (Figure 9). For the simplicity of the sensitivity analysis in the network, variation term was given only to the input of the last month. In terms of different training periods and selected taxa, the sensitivity tests effectively showed important environmental variables in determining community changes. For the data of July 1997, when the flooding occurred during this period used for training, all four environmental variables of organic matter, depth, velocity, and substrates (smaller than 5 mm) caused a high variation of communities in a wide range (Figure 9).

The sensitivity tests effectively showed important environmental variables in determining changes in specific taxa. Densities of Chironomidae and Hirudinea, for example, varied greatly in response to different input ranges (Figure 7). Densities of Oligochaeta, in contrast, were characteristically insensitive to input variables.

Worm Farming

Worm Farming

Do You Want To Learn More About Green Living That Can Save You Money? Discover How To Create A Worm Farm From Scratch! Recycling has caught on with a more people as the years go by. Well, now theres another way to recycle that may seem unconventional at first, but it can save you money down the road.

Get My Free Ebook


Post a comment