Artificial Neural Networks for the Generation of Isohyets by Considering Land Configuration

Abstract

Feedforward multilayer neural networks have been used for the generation of isohyets of the mean annual rainfall in Cyprus (i.e. contours of mean annual rainfall). Archived data of precipitation recorded over a period of sixteen years at 51 meteorological stations have been used for training a suitable artificial neural network. Various neural network architectures and learning rates were tested, aiming at establishing a network which can yield the best possible estimation of precipitation at any arbitrary location on the island. Such a neural network was subsequently used for drawing isohyets. A multiple hidden layer architecture was chosen. This kind of architecture has been adopted for solving problems with similar requirements.

Initially, five stations were excluded from the training data set for validation purposes. Precipitation data for the remaining 46 stations were used for training and testing the network. The parameters used for the training of the network were a) the X and Y coordinates of each station measured from a random reference point, b) the station elevation and c) the mean annual precipitation for the respective station. The correlation coefficient obtained for the training data set was 0.921. The validation of the network was performed by using unknown data for the five stations which were not included in the training phase. The correlation coefficient for the unknown cases was 0.916. The prediction error of the mean annual rainfall was confined to less than 8.8% which is considered quite adequate. A sensitivity analysis was then carried out to investigate the effect of the station elevation on the estimated precipitation. This showed that there are stations for which the elevation is very important whereas for some others is insignificant.

In order to broaden the data base, the five stations used for the validation of the technique were embedded into the training data set and a new training of the network was performed. The architecture of the network, the momentum, the learning rate and the initial weight values were the same as in the validation phase. The correlation coefficient value for the training was increased to 0.962 which was expected due to the increase in the number of data used. It is believed that the accuracy of prediction increases too.

Subsequently, a grid with a grid distance of 10 km was drawn over a detailed topographic map of Cyprus and the X and Y coordinates and the elevation of each grid-point were recorded. This information was then supplied to the network which produced an estimate of the mean annual precipitation at each grid-point. The X and Y coordinates and the mean annual precipitation at both the original meteorological stations and the grid-points were then used as input to a specialised contour drawing software in order to draw the isohyets. It should be noted that the recorded data inherently include "elevation" information, which is part of the site specificity, whereas the estimated precipitation at the grid-points include the elevation information because the network was trained by considering "elevation" as one of the input parameters.

The effect of orography upon rainfall is widely known. Therefore, it is believed that the proposed method for implicitly involving the station elevation in isohyet drawing is more realistic than the traditional methods which make use of only the X and Y station coordinates.