The Power of Neural Networks in GIS

Artificial Neural Networks (ANN) have the power to revolutionize the way we use Geographical Information Systems (GIS) and create geospatial data. For the past 15 years AfricaScope has used ANN in the modelling and mapping of nationally representative household surveys. What this has enabled, is for us to take household surveys in South Africa and Africa to model and map Living Standard Measures (LSM), individual or household income or Gross Domestic Product (GDP) data to the smallest spatial unit of analysis. Depending on the ANN being used it is able to indicate how accurate the estimates are from the model.

The video above gives a clear understanding of how ANN work. It clearly shows how it can be used to solve many geospatial problems and create new geospatial datasets. The power of ANN is that the data you feed into it does not necessarily have to be complete, may have different data types and the modelling can be done on very large datasets and be done very quickly. It does not have to follow statistical norms in terms of normality and linearity as an example. It uses the detection of patterns, edges and objects in creating a final output. So its application could be in identifying optimum sites for government services or retail outlets or in the detection of objects on an image.

Although ANN are powerful, data preparation needs a lot of thought. For example, the use of the ANN in speech recognition requires the conversion of a voice recording into a digital form. Once in a digital form it is entered into an ANN so that learning or training can take place. Sections of the voice recording are taken and used to indicate what words they relate to. This learning is done using large datasets so that the as many combinations of words forms part of the learning so that when applying the ANN to a voice recording, it can accurately identify words when doing speech recognition. The video below illustrates how this is done.

Google's Duplex shows the power of ANN. Imagine the application of this type of ANN in the geospatial industry. There is no doubt that these types of algorithms will revolutionize especially the image processing industry. AfricaScope in collaboration with its international partners is looking at harnessing these capabilities of ANN in detecting dwellings of different types on images of different spectral and ground resolutions and then integrating household survey datasets to impute the household size so that population estimates can be rapidly generated at a local level.