Many studies have used multiple linear regression to develop a model linking landscape variables to perceived scenic beauty. This paper uses a single data set to generate both a regression based model and a model using artificial neural nets. The results of the two models are compared in terms of their predictive capability, their residuals and maps of scenic beauty produced by extending each model across the study area. The results show considerable similarities in the preferred predictor variables.

Many researchers have sought to identify reliable predictors of perceived scenic beauty using regression analysis. With preference, scenic beauty, visual quality or some other synonym or near-synonym as the dependent variable, independent variables have variously been either very physical measures of two dimensional image components (Shafer & Bush, 1977), measures of the distribution of terrain and land uses within the view (Bishop & Hulse, 1994), estimates of more abstract or composite physical qualities (Kaplan et al, 1989) or intermediate measures of human response (Kaplan, 1979). Some studies have taken several sets of measures and applied them separately or together to compare their predictive powers (Gobster & Chenoweth, 1989; Steinitz, 1990).

The vast majority of these studies have used multiple linear regression to derive their predictive model. Such models include more than one independent variable to explain the behaviour of the dependent variable (scenic beauty in this case). The coefficients (bp) of each variable (Xip) in the linear additive model

Yi = b0 + b1Xi1 + b2Xi2 + ......... + bpXip + ei

are estimated by the method of least squares.

Although some high R-squared values showing a good fit between the model and the scenic beauty estimates have been achieved few models have transferred well from one data set to another. This arises in part from the likelihood that, in different environments, judgments of scenic beauty are made on either subtly or significantly different grounds. It surely also arises from the fact that the linear model is an inherently poor surrogate for the complex factor linkages which take place in the human brain in making even a snap judgment of scenic beauty. There are two very obvious problems with the linear multiple regression model. The first is that it would be simply a fluke if our perceptions followed a linear model: how many of us would agree that twice as much water, smoothness or mystery in a view renders it twice as beautiful? The second problem is that there are inevitable interdependencies between the factors which cannot be properly included in a linear regression model: a water body in an already attractive setting may not contribute as much to perceived beauty as a similar water body in an otherwise desolate environment.

The ideal result in terms of mathematical congruence with popular theoretical models of human visual preference would be a model which linked basic physical variables with second order visual preference predictors such as mystery, complexity, legibility and coherence [Kaplan, 1979]. Second order variables of this type would then become predictors for either visual preference itself or an additional intermediate predictive layer. Many authors have used multiple linear regression models to propose both direct and staged linkages between physical variables and visual quality. These have generally been treated sceptically by other researchers because while high R-squared values are observed in individual studies the results in other locations have found similar predictors but no convergence of coefficients to stable or widely accepted values.

There has recently emerged a technology which may offer a viable alternative to the use of linear models for scenic beauty estimation. This technology is a computational procedure originally developed in an attempt to mimic the complex neural interactions of the human brain. It treats the space between stimuli (inputs) and response (outputs) as a network of nodes whose individual responses are triggered by the inputs and other nodes of the net (Figure 1). Because of this claim to mimic the neuron level stimulus/response characteristics of the brain and to network these neurones together this technology is called the neural net [Rumelhart & McClelland, 1987]. In mathematical terms a neural network model is defined (Müller & Reinhardt, 1990, p12) as a directed graph with the following properties:

1. A state variable ni is associated with each node i

2. A real-valued weight wik is associated with each link (ik) between two nodes i and k.

3. A real-valued bias Ji is associated with each node i.

4. A transfer function fi[nk,wik,Ji,(k_i)] is defined, for each node i, which determines the state of the node as a function of its bias, of the weights of the incoming links, and of the states of the nodes connected to it by these links.

The transfer function is usually taken to have the form f(Skwiknk - Ji), where f(x) is either a discontinuous step function or its smoothly increasing generalisation known as a sigmoid function. In one very common implementation the model is 'trained' by providing multiple sets of known input and output values. The computer initially assigns random weights to the links and then using a technique known as back propagation, iteratively adjusts the weights in order to minimise the mismatch between computer outputs and known outputs.

Figure 1. The typical structure of an artificial neural net. The number of input, hidden and output nodes are set by the user. There may also be more than one hidden layer.

The artificial neural network is designed to work best on massively parallel computing devices where each processor takes the place of a single neuron/node. The connectionist pathways of the parallel computer environment can however be programmed on a conventional single processor machine. The process merely takes a little longer.

The potential of neural nets has been explored recently in the landscape planning [Gimblett et al, in press] and architectural design [Coyne & Yokozawa, 1992] contexts. After a comparison of neural net and more conventional modelling approaches in spatial interaction modelling (using journey to work data) Openshaw (1993) concluded that there were a number of possible benefits to be gained by using neural nets. These included: better performing models, greater representational flexibility, handling of explicitly noisy data and the chance to exploit future NN developments in other fields.

Artificial neural nets have not yet however been used in landscape preference studies. In addition to the potential benefit identified by Openshaw there is the potential, based on their ability to deal with non-linearity and factor interdependence, that they may provide for more stable results than traditional regression approaches.

This paper take a single data set of potential predictor variables and scenic beauty judgements and seeks to model the stimulus/response linkage using both multiple linear regression and neural networks. The procedures, results and possible interpretations that can be drawn from each method are compared. It was not an objective of this work to determine which provides a better model. A single experiment is insufficient for that. The objective was rather to show that the intuitively appealing neural net approach is applicable to scenic beauty modelling and to determine whether a NN model is, in general terms, consistent with prior theoretical and empirically generated predictors of visual values.

The data set used differs from that typically used in scenic beauty studies in two ways. Firstly the scenic beauty judgments are based on 360 degree video panoramas rather than single static images. Secondly the potential predictor variables have been computed from mapped land form and land cover data for the area using a geographic information system. This contrasts with the more normal practice of estimating parameters from the images themselves. The rationale for both these choices is described further in Bishop & Hulse (1994) but, in short, relates to a desire to be able to predict and map visual quality across wide areas directly from mapped (and preferably digital) data.

All the sample points for this study fall within an area roughly 20km by 10 km in western Victoria, Australia. The data set was originally developed in conjunction with a study of the visual impact of electricity transmission lines. Consequently, transmission structures are part of the scenery at several of the sample locations. This is consider valid and a reasonable component of the overall modelling process given the ubiquity of transmission lines in our rural landscapes.

A digital coverage of the area encompassing elevation, vegetation type, vegetation height, stream location and transmission tower locations was prepared and converted to a 50 m (0.25 ha) grid for later manipulation. The area was zoned according to distance from the transmission lines and then grid cells chosen at random within each zone. A total of 25 locations were thus identified. The centre of each chosen cell was then mapped onto large scale (1:10300) aerial photographs.

A video camera was taken into the field and by using the aerial photographs the sample points were located on the ground. In most cases this was not difficult because accurate scaling could be done from 'landmarks' such as the intersection of boundary fences, distance from forest edges or individual trees. At each location the camera was mounted on a tripod. A 360 degree clockwise panorama was recorded commencing due north, taking 60 seconds to complete and keeping the camera horizontal throughout the sweep.

The 25 panoramas where shown to an audience of 48 university students who were asked to record their visual preference for each panorama on a scale of 1 to 9. They were first shown a preview of 5, 15 second excerpts from the panoramas and asked to used these as guidance in determining the end points in their use of the 9 point scale. Each respondents mean and standard deviation was used to convert their raw responses to Z-scores and the Z-scores were averaged across the response group to produce the scenic beauty estimate for each location (Brown & Daniel, 1990).

Factor analysis of the raw responses of the individual respondents showed one factor explaining 33.1% of the variation in scores, a second factor 10.7% and 12 factors with eigenvalues greater than 1.0. This is very similar to the agreement levels found in a previous study based on video panoramas (Bishop & Hulse, 1994) but is low compared to many slide based studies.

As the objective involves prediction of scenic beauty based on available mapped data, values for independent variables were computed directly from the digital data set using the ARC/INFO GRID module. The sample locations were stored as precise co-ordinates (albeit in the centre of the sampled grid cell) and all other data was encoded on the 0.25 ha grid. The elevation values were the result of gridding of 10 m contours from a 1:25,000 source map. Vegetation distribution was digitised from 1:10300 aerial photographs. Vegetation height was estimated from the same photographs in conjunction with site visits. Four height ranges were used in the initial estimation and these were converted to three heights (7m, 15m and 25m in the data base). Streams were digitised as all blue line on the 1:25,000 maps. Tower allocations were based on their true locations supplied by the electricity authority.

A number of ARC/INFO macros ('amls') were written to automate the data extraction process (Bishop & Robey, 1994). These used the GRID module's capacity to undertake viewability, proximity and overlay analysis. First vegetation heights were added to terrain heights to generate a visual height field. The viewshed of each sample location was then calculated. All subsequent analysis was based on these derived viewsheds. Thus parameters such as total seen area, minimum distance to water, number of towers, range of visible elevations, extent of foreground eucalypt forest and maximum foreground slope were calculated. Table 1 lists all the derived variables used in subsequent analysis. No attempt was made in this study to weight the variables according to any actual or potential "dominant" view. It is arguable that within the 360o panoramas there would be some view directions which would attract a viewer's attention more than others and would therefore have more influence on their rating of the location. This prospect is discussed further in the conclusions.

Table 1. Variables derived from mapped information using the geographic information system.

Variable Name	Measure within viewshed of....
tot_area	Total Visible Area
str_length	Stream Length
Str_in	Stream Length Index
no_towers	Number of Towers
Tow_in	Tower Index
MinStreamDist	Minimum Stream Distance
MinTowerDist	Minimum Tower Distance
Zmax	Maximum Elevation
Zmin	Minimum Elevation
Zrange	Range of Elevation
Smax	Maximum Slope
Smin	Minimum Slope
fZmax	Maximum Elevation (foreground)
fZmin	Minimum Elevation (foreground)
mZmax	Maximum Elevation (midground)
mZmin	Minimum Elevation (midground)
bZmax	Maximum Elevation (background)
bZmin	Minimum Elevation (background)
fSmax	Maximum Slope (foreground)
mSmax	Maximum Slope (midground)
bSmax	Maximum Slope (background)
a_f1	Area of Pine Plantation (foreground)
a_m1	Area of Pine Plantation (midground)
a_b1	Area of Pine Plantation (background)
a_f2	Area of Eucalypt Forest (foreground)
a_m2	Area of Eucalypt Forest (midground)
a_b2	Area of Eucalypt Forest (background)
Euc_bg	Eucalypt Forest proportion of background
Shlt_fm	Area of Shelter Belt (fore & mid ground)
a_f3	Area of Shelter Belt (foreground)
a_m3	Area of Shelter Belt (midground)
a_b3	Area of Shelter Belt (background)
a_f6	Area of Grazing Land (foreground)
a_m6	Area of Grazing Land (midground)
a_b6	Area of Grazing Land (background)

The first round of calculated parameters was compared site by site with the video panoramas. In a small number of cases there were obvious discrepancies between what was calculated as being visible and what was actually visible. Three obvious cases were those in which the sample point had been just within forest close to the forest edge. The data base had located them within the forest and thus with very limited visual range whereas the panoramas showed views beyond the forest on one side of the sample point. The points for computation were move one cell (50 m) in the direction of the forest edge and the parameters recomputed. The results now fitted well with observation.

In addition to very straight-forward measures of transmission tower and stream visibility such as number of towers or distance to the nearest stream it seemed appropriate to include a variable that reflected both the number of towers (or stream sections) visible and their distance. Thus, a tower index was defined as:

where the distance is measured in meters. Thus, a visible tower at 200 m would contribute 5 to the index, while a visible tower at 2 km would only add 0.5. The stream index was similarly defined except that each visible cell containing a river or stream was counted in the index.

First a simple correlation analysis was run between the mean visual quality estimate and each of the potential predictors. The correlation coefficients suggested that high preference scores come with enclosure rather then exposure, neither shelter belts nor grazing land in the fore or mid ground and, more surprisingly, low slopes in the back-ground.

Multiple regression analysis enhances the picture of preference we can build in this environment. Table 2(a) shows that six variables are able to explain 77.8% of the variation in preference with an adjusted R2 of 0.704. This is a good result and compares well with an earlier study of this type (Bishop & Hulse, 1994). These six variables and their coefficients indicate high preference when shelter belts are not in the fore and mid ground, when there are higher slopes in the mid ground, when the tower index is low, when the proportion of native forest in the background is low, when the overall elevation range is high but, again a surprise, when the maximum slope is low.

Table 2. The result of the regression analysis using normalised varaibles: showing (a) that the best six variables explain nearly 80% of the variation in scenic beauty estimate, and (b) that all six variables are highly significant and with similar levels of influence.

Table 2(b) shows the coefficients of the independent variables. Each variable was normalised to vary from 0 to 1 and so the coefficients indicate the strength of the influence of each variable. The six are quite similar in their levels of contribution. All are significant at p< 0.01.

Development of a neural net model based on the same data is more complex than the regression case because the number of options is much greater. A neural net model can have any number of intermediate or hidden layers, and any number of nodes within each of those layers.

It seemed appropriate therefore to choose a simple starting point and then choose variations on that initial condition to determine whether improvements in predictive capability could be made. The starting point was to use the same set of inputs as for the regression model reported in Table 2. A neural net with a single hidden layer of four nodes was specified (Figure 2). The mean preference score was the single output node.

	Tow_in	Zrange	Smax	mSmax	Shlt_fm	Euc_bg
H 1	0.7	-5.3	-14.1	0.2	1.4	3.8
H 2	0.1	-0.5	5.1	-0.5	5.5	-0.0
H 3	-0.5	4.0	3.8	-3.5	5.4	0.1
H 4	-3.8	10.5	-10.8	-0.8	-1.9	-1.0

	H 1	H 2	H 3	H 4
Output	10.7	5.0	-6.9	4.3

Figure 2. The neural net structure used in this work and the weights derived from one training run using the 25 site training set. Note that another run would produce a somewhat different set of weights. Each hidden and output node also has a 'bias' factor associated with it but these are not shown.

The data for the 25 sites formed a set of 25 examples for training the net. The public domain neural net software PlaNet was used (Miyata, 1991). This takes the net specification and the training data, allocates random weights to the network links and determines the fit between the predicted outcomes and the actual or target outcomes. A back propagation process is then used to adjust the weight to create a better fit between output values and target values. It is the nature of the neural net learning process that the fit between outputs and targets continues to diminish gradually through continued cycling of the learning process.

With only 25 cases and more than 25 weightings and biases available for adjustment a perfect result should be achievable with sufficient iterations. However, the reduction in error is rapid in the early cycles but then becomes much slower. A clear indication of the potential fit between a set of inputs and the outputs is achieved by reviewing the error level after about 5000 cycles.

Table 3 Showing the neural net error after 5000 learning cycles with different sets of input variables. The input set with the same variable as identified as the best model in the regression analysis and used in subsequent comparison with the regression model is shown in bold type.

No.of Inputs	Input 1	Input 2	Input 3	Input 4	Input 5	Input 6	Error after 5000 cycles
6	Tow_in	Zrange	Smax	mSmax	Shlt_fm	Euc_bg	0.0003
6		Zrange	Smax	mSmax	Shlt_fm	Euc_bg	0.0012
5	Tow_in		Smax	mSmax	Shlt_fm	Euc_bg	0.0007
5	Tow_in	Zrange		mSmax	Shlt_fm	Euc_bg	0.0043
5	Tow_in	Zrange	Smax		Shlt_fm	Euc_bg	0.0005
5	Tow_in	Zrange	Smax	mSmax		Euc_bg	0.0012
5	Tow_in	Zrange	Smax	mSmax	Shlt_fm		0.0016
6	No_tow	Zrange	Smax	mSmax	Shlt_fm	Euc_bg	0.0003
6	Tow_in	Zrange	Smax	bSmax	Shlt_fm	Euc_bg	0.0004
6	Tow_in	Zrange	Str_in	mSmax	Shlt_fm	Euc_bg	0.0034
6	Tow_in	Zrange	Smax	mSmax	Shlt_fm	Tot_ar	0.0010
2			Smax		Shlt_fm		0.0062

The first training run over 5000 cycles with the set of input variables that gave the best result in the regression analysis produced and error rating of .00796. This number is itself of no easily discernible meaning but it gives a point of comparison for the subsequent attempts to isolate a better model. Table 3 shows some of the variations attempted and the error values after 5000 cycles. Certain conclusion can be drawn from these results:

1. each of the variables used was important to the neural net model, this was clear since the omission of any one produced an increase in the 5000 cycle error level. The largest increases in error followed the omission of the maximum slope. This echoed the regression result as this variable had the large coefficient and the largest partial F in the normalised regression. Fore or mid-ground wind breaks and shelter belts were also high in priority in both modelling procedures.
2. small changes in the selected input variables gave comparable performance in model fit - for example replacement of the 'tower_index' with the 'no_of_towers', or replacement of midground slope maximum with background slope maximum. In each of these cases there was a high level of correlation between the original variable and its substitute (0.805 and 0.766 respectively).
3. other changes in the input variables gave a considerably higher error level than the initial set.
   

On the basis of these results it was clear that variables which are good predictors in a regression analysis are also good predictors within a neural net model. This was in itself a significant result but raised a number of additional questions.

1) What is the predictive capability of the neural net as compared to a regression based on the same variables?

2) How to the residuals in the regression compare to the misfits in the neural net model - do the same ground locations tend to be difficult to model in both cases?

3) If each model is applied across a landscape to produce a broad area mapping of landscape values how will the two maps compare?

These questions are explored in the next section.

Predictive Capability

In order to compare the predictive capability of the two modelling approaches attention was concentrated on the preferred regression model and a neural net developed using the same independent variables.

A direct comparison of predictive ability can be obtain by correlation analysis of the observed mean preference values with (a) the fitted values generated by the regression equation and (b) the outputs from the neural net.

As indicated above the latter fit will depend on how many iterative learning cycles are permitted, but the improvement in fit becomes very slow after a quite small number of cycles provided a reasonable fit is possible. A comparison was therefore undertaken with the result at 200 cycles and at 5000 cycles to compare these results with each other as well as with the regression result.

Correlations found were:

Mean preference V Fitted regression values 0.854

Mean preference V Neural net (200 cycles) 0.854

Mean preference V Neural net (5000 cycles) 0.936

It is easy to determine the relationship between independent and dependent variables in the case of regression analysis by examination of the coefficients. In the case of neural net modelling however the interpretation is less straightforward. It is possible to check firstly whether the predictors are being used in the same sense (i.e. direct or inverse relationship) as in the regression. By using sensitivity analysis one can also get a feeling for the magnitude of their relative contributions.

Examination of the node link weightings as shown in Figure 2 suggests complete consistency with the regression weights, i.e. negative weights on shelterbelts, the tower index and maximum slope but positive weights on background native forest and slope in the mid-ground. In order to test the apparent direction and strength of factor contribution to preference a set of sensitivity runs were conducted. A single site was selected. This was site number 12 and chosen because none of its factors score were at the extreme of the site ranges and its preference score was reasonably close to the mean (0.364). Each factor score at this site was varied through 20 steps of 0.05 each up to 0.5 above and below the actual site score. The normalised site value of elevation range (Zrange), for example, was given sensitivity test values ranging from 0.825 to 0.920. Its actual score was 0.876. After running the true data for all sites through 2000 training cycles, the sensitivity test values were passed through the net. This showed how variation in a single factor score about its actual site value effects the output (preference) score. Thus, in the example above, as Zrange changed from 0.825 to 0.920 the output score changed from 0.284 to 0.387. In other words, in the neural net model generated from the training set the influence of Zrange on visual preference is positive. Table 4 shows the range and direction of change in preference score based on the same range of variation in factor score. Although the relationships are non-linear, it is encouraging that the ranges are comparable in their relative contributions to the coefficients of the preferred regression model.

Table 4. Sensitivity analysis in the neural net showing the direction and degree of change in output generated by a change in normalised input of 0.1 in each of the predictor variables. The regression coefficients on the same variables are shown for comparison.

Variable	Direction of Influence	Change in output	Regression coefficient
Tow_in	Negative	-0.037	-0.384
Zrange	Positive	0.103	0.466
Smax	Negative	-0.156	-0.965
mSmax	Positive	0.050	0.502
Shlt_fm	Negative	-0.102	-0.634
Euc_bg	Negative	-0.067	-0.333

Comparing the residuals

A very high level of coherence (r = 0.95) between regression and neural net residuals was found in the NN analysis after 200 cycles (Fig 3a). Continuing the NN iterations reduced this very high similarity, and after 5000 cycles the correlation was 0.37 with much lower NN residual values (Figure 3b). This clearly indicates that the two procedures are finding similar patterns in the landscape and that the same sample points create difficulty in generating a predictive model each in case.

Figure 3. The relationship between the residuals of the regression analysis and the 'residuals' of the neural net analysis (a) after 200 training cycles, (b) after 5000 training cycles. The close fit in (a) indicates that the sample sites tend to diverge from each model in the same manner. The very much smaller neural net residuals in (b) are indicative of the on-going fine-tuning possible in the non-linear model.

Mapping visual values

Predictor variables were generated for a further 961 points within the study area. These points, with 100 m separation formed a 30 km by 30 km grid. Both the preferred regression model and the corresponding neural net model were run on each of these points. The two sets of predicted visual values were significantly, although not particularly strongly, correlated (r = .30, p< 0.001) as illustrated in Figure 4. Two predictive maps of visual value for the sub area were then produced to compare the spatial distributions. Because the regression values were skewed to the right (high values) while the neural net values were skewed left (low values) it was difficult to obtain a direct visual comparison of the relative, rather and absolute, comparative mappings. To better appreciate the degree of similarity of the ranking across the maps, each was sliced into 5 equal area visual categories. The two resulting maps (Figure 5) thus represent two quite different approaches to modelling visual values. Each has been used to generate a scenic beauty mapping of the conventional very low, low, medium, high and very high style of classification. The maps are both sufficiently similar and sufficiently different to encourage further comparison and evaluation of these techniques.

Figure 4. The output values generated by application of each model to the 961 regularly spaced points across the study area.

The maps have the major areas of high and low rated scenic beauty in similar locations. The smaller areas however are frequently quite different: the area in the north-east corner being a good example.

(a) (b)

Figure 5. The predicted scenic beauty maps for a 3 km by 3 km portion of the study area: (a) based on the regression model, (b) based on the neural net model. Each cell is 1 ha surrounding the 961 grid points. Each set of model outputs was sliced into 5 equal area categories. The major features of the two maps are similar but there are also areas of substantial difference. Light areas are mapped as the highest level of scenic beauty.

NNs do not reproduce the complex functions of the brain. They are sometime regarded as simply another approach to non-linear regression. They do however offer an alternative approach to the modelling of complex phenomena, including scenic beauty. This paper shows that the use of an alternative approach may help to reinforce some conclusions drawn from regression models while, through differences in the results, providing additional insights. There is no clear evidence here that, in this location, one model is any better as an operational tool than the other. Despite the intuitive appeal of the non-linear, interdependent neural net model there is still considerable ground to be explored before it could be established that NNs overcome the historical or theoretical deficiencies of regression modelling. Indeed, the similarity of the results suggest that perhaps the two approaches are similarly deficient. The question of whether it is the similarities or the differences which will eventually prove more telling remains unresolved.

Some promising directions exist which may help bring resolution. These involve: enhancement of the NN technique; refinement of the data set; and finally application to a wide range of existing data sets with appropriate ground truth.

Enhancement in technique

Recent developments in neural net application may make it possible for user to proceed beyond the often 'black-box' characterisation of neural net based models. Coyne [1991] and Coyne and Yokozawa [1992] refer to a variant of neural nets as 'connectionist techniques'. While acknowledging that there is no 'deep' knowledge embedded in connectionist systems they propose the use of such systems with no hidden units and without explicit inputs and outputs. That is, every unit (variable) is potentially linked to each other unit. As with traditional neural net training large data sets can be used to 'train' the connective system. The work reported here could be extended by use of the connectionist technique. Visual quality could initially be treated as simply another variable to be associated with the physical variables in the environment. Once a connectionist system is 'trained' by the examples, typical high visual quality environments may be described by 'clamping' (holding constant) visual quality and using this to 'activate' (bring into the predictive model) closely associated variables through a series of iterative cycles. Visual quality within different landscape types may be described by clamping visual quality together with a variable considered representative of the landscape type.

Refinement of the data set

In describing the data preparation portion of this study it was pointed out the independent variables were not weighted by view direction. To include a consideration of 'dominant view' in the modelling would however require either an independent rating of visual quality, a prior model of dominant view, or a means of recording each respondent's attention patterns. Even with such information there is no clearly appropriate weighting function to be applied. One option which might be applied in a future study would be to divide the viewshed by quadrant or sub-quadrant and generate independent variables for each sector. This would help to identify the dominant view and its significance while also perhaps indicating the factors which contribute to a dominant view within a panorama.

Repeated application/ground truth

More studies of this kind, particular if accompanied by a greater degree of ground truthing to establish the validity of the resulting models appear to be necessary. However it may be that much of the work required for such studies could be by-passed. There exists a large body of work reporting regression analyses which has not produced results which have satisfied landscape managers. The data sets underlying those earlier studies are an immense resource available to support a rapid assessment of neural net modelling.

Arrival at valid and repeatable models does not however necessary provide easy answers to the best location for new development. It is not sufficient to identify the cells which map as low scenic beauty and determine these as the most suitable for disruption. These may be the cells which are contributing to high scenic values at some other location - near or distant. The approach must rather be to introduce the proposed changes into the digital data set, use the model to remap the visual values and by subtraction identify the visual impact of the changes. Such an approach is only possible with visual models based on mapped data and consideration of the total viewshed at each ground location.

This work is part of an on-going investigation of techniques for computer based visual analysis. The work has been funded by the Australian Electrical Supply Industry Research Board, VicRoads and the State Electricity Commission, Victoria. Gerard Hennicke wrote the AMLs and some useful utilities for working with PlaNet. Thanks also to the reviewers of this paper for adding to the scope and strength of the conclusion.

BISHOP, I.D. and HULSE D.W. (1994) Predicting scenic beauty using mapped data and geographic information systems, for Landscape and Urban Planning, 30, 59-70.

BISHOP, I.D. and ROBEY M. (1994) Implementing and environmental impact model within a geographic information system, Proceedings AURISA '94, Sydney pp281-291.

BROWN, T. and DANIEL, T.C. (1990) Scaling of Ratings: concepts and methods, USDA Forest Service Research Paper RM-293, Fort Collins CO.

COYNE, R.D. (1991) Modelling the emergence of design descriptions across schemata, Environment and Planning B, 18,427-458

COYNE, R.D. and YOKOZAWA M. (1992) Computer assistance in designing from precedent, Environment and Planning B, 19,143-171.

GIMBLETT, R.H., BALL G.L. and GUISSE A.W. (1994) Autonomous rule generation and assessment for complex spatial modeling, Landscape and Urban Planning, 30, 13-26.

GOBSTER, P.H. and CHENOWETH R.E. (1989) The dimensions of aesthetic preference: a quantitative analysis, Journal of Environmental Management 29:47-72.

KAPLAN, S. (1979) Perception and Landscape: Conceptions and Misconceptions, From Elsner, G.H. and Smardon, R.C. (eds.) 'Our National Landscape: a conference on applied techniques for analysis and management of the visual resource', USDA Forest Service Report PSW-35.

KAPLAN, R., KAPLAN S. and BROWN T. (1989) Environmental preference: a comparison of four domains of predictors, Environment and Behavior 21:509-530.

MIYATA, Y. (1991) A user's guide to Planet v5.6: a tool for constructing, running and looking into a PDP network, Computer Science Dept., University of Colorado, Boulder.

MÜLLER, B., & REINHARDT, J. (1990). Neural Networks: an introduction. Berlin: Springer-Verlag.

OPENSHAW, S. (1993) Modelling spatial interaction using a neural net, in M.M.Fisher and P.Nijkamp (eds) "Geographic Information Systems, Spatial Modelling and Policy Evaluation", Springer-Verlag, New York.

RUMELHART, D.E. and McCLELLAND J.L. (eds) (1987) Parallel distributed processing: explorations in the microstructure of cognition, Volume 1, Foundations. MIT Press, Cambridge MA

SHAFER, E.L. and BUSH R.O. (1977) How to measure preferences for photographs of natural landscapes, Landscape Planning, 4,237-256.