A neural-evidence pooling approach to predict urban sprawl using multi-temporal remote sensing data

This paper introduces a dynamic neural network based technique to predict the urban growth in Sriperumbudur taluk, India. The novelty of this technique is integration of neural network and theory of evidence to predict future growth which does not require any definition of input parameters, spatial rules or large dataset that involves expert knowledge and time consuming manual work for preparation. First, Land Use Land Cover (LULC) images were obtained by classifying LANDSAT TM satellite image of 30m spatial resolution captured in the year 1991, 2000 and 2009. Multi-Layer-Perceptron (MLP) neural network is used as a classifier. Second, five spatial metrics namely Shannon's entropy, aggregation index, density, nuclearity, and proximity are derived around each non-urban pixel in 2009 by considering 7x7 window of spatial unit in the LULC images of the three years. Next, these metrics are fed to Focused Time Delay Neural Network (FTDNN) and the metrics for year 2018 are predicted. Finally, evidence pooling is applied to the predicted metrics to identify whether the non-urban pixels in 2009 changes to urban pixel in the year. The predicted results shows that 12.5% of nonurban LULC mainly agriculture and wasteland in 2009 will change to urban land use in 2018. In order to test the performance of the proposed technique, urban land use in 2009 is predicted using LULC images of 1991 and 2000 and compared with actual pixels transformed to urban land use in 2009 from 2000. The predicted result is found to be 81.28% accurate with spatial autocorrelation of 0.90. The proposed method is also compared with conventional model SLEUTH and found to be 2.56% more accurate with 0.19 increases in spatial correlation. The results potentially help the city planners and developers to have prior visualization of future urban sprawl in the study area for effective city planning.