Neural network model for consequence analysis of developmental proposals

Suju M. George, P. Ram Babu and P. Khanna
National Environmental Engineering Research Institute,
Nagpur India 440 020

Application of the integrated system GIS-RDBMS-Neural networks for regional system identification is reported in the context of sustainable development. A Geographical Information System linked with an external Relational Data Base System maintained spatial and numerical data, and provided inputs to Neural networks. The integrated system is used for consequence analysis of alternate developmental scenarios, and choice of a preferred scenario. The case of developmental planning in National Capital Region of India is presented for illustrating the effectiveness of this modelling approach.

GIS, an external RDBMS, and neural network models have been integrated into a coupled system where the two neural network models interact with RDBMS and GIS to derive input data which are of spatial and nonspatial in nature Neural Network I (NN-I) has been designed to identify the interlinkages between inputs, viz. demographic and population characteristics, activities, resource status and resource use patterns; and output, viz. environment status. Neural Network II (NN-II) identifies the interlinkages between inputs, viz. demographic and population characteristics, activities, resource status and use patterns, amenities, and environment status; and output, viz. urban and rural quality of life levels.

NN-I has been configured for 56 input neurons and 6 output neurons; while NN-II for 35 input neurons and 3 output neurons. The number of input neurons and output neurons correspond to the number of input and output descriptors, respectively.

The data values for the descriptors of population and demographic characteristics; activities, viz. industry, agriculture, mining; resource status and use pattern; environmental status; amenity levels; and quality of life levels have been collated from primary and secondary sources, satellite imagery, and processed data incorporating expert assessments. The data on demographic and population characteristics have been drawn from census data. The activity levels have been collated from primary surveys and secondary sources. The resource status data have been drawn from satellite imageries and subsequent GIS analysis, surveys, and secondary data. The data on environmental status has been the aggregation of media specific parameters, and the expert assessments that resulted in identification of hot spots; as delineated in the project report prepared by National Environmental Engineering Research Institute (Carrying Capacity 1995).

The pre-processed and collated data have been re-processed at GIS and RDBMS for obtaining tehsil level information, arriving at accessibility indices, and resource use intensities. GIS thematic maps, viz. land use, land capability, land erosion, water quantity and quality, and other environmental media hotspots have been used to obtain the spatial data at tehsil level. Spatial analysis capabilities of GIS have been further used to combine information from separate base maps to derive descriptors such as fraction of flood prone area, area effected by environmental impairment and resource status degradation, stress on transportation, and resource use intensities. The spatial data derived from GIS have been stored in an external RDBMS. The nonspatial data have been directly fed into RDBMS. The data processing module at the input stage of the neural networks normalise the data for forming input and target vectors for NN-I and NN-II.

The 37 input vectors, generated as explained above and each corresponding to that of a tehsil; and 22 input vectors encompassing the expert's knowledge, both totalling to 59 input vectors have been used to train the NN-I. Likewise, a set of 56 input vectors which include 11 vectors corresponding to knowledge from experts form the training patterns for NN-II.

The backpropagation algorithm that minimizes the mean squared error has been used as the training algorithm.

The training patterns have been presented in a batch constituting epochs. An epoch is an iteration to adjust the weights to learn partly all the training patterns. The learning rate has been varied from 1.0 to 0.00001 by continuously monitoring the learning process. On-line variation of learning rate has been used to expedite convergence. The training has been continued till the mean squared error reduced to less than 0.00001 for every pattern. The number of iterations required for convergence is less than 20,000 for both NN-I and NN-II in the cases of the network configurations selected for the final prognostics.

The integrated system has been implemented on HP 9000 series workstation. The program modules, data pre-processing modules and neural network models are coded in 'C' programming language.

The trained NN-I and NN-II have been used for the prediction of environmental degradation status and equitable quality of life levels consequent upon developmental interventions. The output of NN-I provides the environmental degradation status levels with respect to air, water, and land media; and problems relating to solid waste and noise;. while the output of NN-II gives equitable quality of life levels in terms of maximum and minimum quality of life levels in urban areas, and average quality of life in rural areas.

The maximum and minimum quality of life levels observed across urban areas of the tehsil, and average quality of life in rural areas have been taken together as indicator of equity of quality of life across the tehsil.

The results of consequence analysis, for select tehsils, constituting environmental status degradation indices and Quality of Life levels are given in Table 1 for present, Business as Usual, and Preferred scenarios for the year 2021 A.D. The normalised values shown in the Table 1 assign 0.7 to the region with highest intensity of hotspots in the present scenario.

Consequence analysis is an integral part of developmental planning based on the premises of carrying capacity. Neural network based modelling approach provides a means of identifying a system of complex and non linear relations amongst resource endowment - socio - economic activities - environmental media status - quality of life levels. The data requirement for training a neural network for regional system identification is less in comparison to integrative causal models, media and resource specific regression models, and optimization models Table 1: Quality of Life Levels in Selected Tehsils of National Capital Region for Present (Pres-Sc),Business as Usual(BaU), and Preferred (Pref-Sc) Scenarios for the Year 2021 A.D.