1) Artificial neural networks were used to model urban growth by reducing subjectivity and calibration time compared to traditional models.
2) The neural network was trained using spatial data on driving factors like distance to roads, city core, and existing development from GIS to predict urban growth probability.
3) The neural network model was able to accurately simulate past urban growth in Dehradun, India in 2001 and 2005 based on maps and data from 1997, 2001, and 2005, demonstrating the ability to predict urban expansion.
2. Need of Urban growth modelling
•Urban areas are growing at a very fast rate
•Need for Urban growth models to predict areas of future growth.
•So that proper infrastructure facilities can be provided in these areas
3. Gaps in urban growth modelling
• Subjectivity
• Model calibration is done by trail and error
• Complicatedness
• Models are non spatial in nature
•. Grossness.
4. Research Aims
•Reduce subjectivity in urban growth modelling
•Reduce the model calibration time
•Need to couple GIS and RS with urban growth models for
modelling urban growth
5. In order to reduce the subjectivity and calibration time :
Artificial Neural Network ( ANN) were used
• They are able to learn the patterns directly from the database without much
human intervention
• ANN make no assumption regarding the distributional properties of data
• Mixture of data types can be used
• No restrictions on using non numeric data
• They can solve highly non linear problems
6. Urban growth = f ( dist. to city core,
dist. to road,
dist. to nearest built-up,
Percentage of built up in neighbourhod )
Urban growth = ANN ( dist. to city core,
dist. to road,
dist. to nearest built-up,
Percentage of built up in neighbourhod )
9. Urban Growth 1997-2001 Urban Growth 2001-2005
797 ha. Changed from Non Built-up to Built-up 1108ha. Changed from Non Built-up to Built-up
10.
11. Dist. to Roads
Dist. to City Core
Four driving variable grids created in GIS
Percentage of Built - up
Dist. to Built - up
12. Dist. to nearest built-up Dist. To city core Dist. to roads Density of built up in neighbourhod
dist. City core dist. Built-up dist. Road density Built-up
Target value
0.1677317 0.0031023 0.0072796 0.2
1
0.1676008 0.0031023 0.0025737 0.36
1
0.1675815 0.0031023 0.0025737 0.2
0
0.1674457 0.0043873 0.0081388 0.32
1
0.1672564 0.0031023 0.0156554 0.2
1
0.1672126 0.0043873 0.0025737 0.08
0
0.1669403 0.0031023 0.0072796 0.52
1
0.1669403 0.0031023 0.0072796 0.52
1
0.1664677 0.0031023 0.0077212 0.52
1
0.1659447 0.0031023 0.005755 0.2
1
Training Data NN -output Urban growth 1997-01
0.7
0.8 1=Growth
0.5 0 = No growth
ANN 0.55
0.3
0
0.3
0.6
0.7
0.9
13. Multilayer perceptron (MLP) feed-forward Artificial neural network
Network Architecture
2400 Training samples
Target Value=1
0.85
0.9
e=1-.85
e=1-.9
Output Layer
Input Layer Hidden Layer
Supervised back-propagation learning algorithm (BP) has been used for
training the network
Stopping criteria: 1. Fixed number of iterations take place 2. Error drops
below a certain level 3. The network starts over training.
Input Layer = 4 Neurons
1200 Validation samples to prevent the network from overtraining
1200 Testing samples to ( 1= the generalization capability of the neural network
Output Layer =1 Neuron test Growth, 0= No Growth)
14. Network training
Input layer Hidden layer Output layer Remote sensing
GIS database data
f1
f2 Network output Desired output
f3 Change
Error
f4 detection
Training dataset
Optimal
CA simulation weights
Stop simulation f’1 Potential for urban growth (P)
Database for study area f’2
f’3
yes no
Stopping criteria fulfilled f’4
Masking of exclusionary areas
Updation of database
Allocate cell to built-up If P > threshold value
15. Network Architecture
1.What should be the number of hidden layers
2.What is the number of neurons in each hidden layer
The architecture of neural network was decided
1.Heuristically
2.Trial and error. more than 50 ANNs were designed
Heuristically :The number of nodes in a single hidden layer
Kanellopoulos and Wilkinson (1997) 3Ni
Hush (1989) 3Ni 3
Hecht-Nielsen (1987) 2Ni +1
Wang (1994b) 2Ni /3 9
Ripley (1993) (Ni+No)/2
Paola (1994) 2+No *Ni+1/2 No (Ni2+Ni) -3 12
Ni + No
No is number of input nodes, Ni is number of output nodes
16. 1. Trial and error. More than 50 ANNs were designed, using
single and double hidden layer.
4-6--1 4-3-3-1 4-6-3-1 4-9-3-1 4-12-3-1 4-15-3-1 4-18-3-1 4-21-3-1
4-15--1 4-3-6-1 4-6-6-1 4-9-6-1 4-12-6-1 4-15-6-1 4-18-6-1 4-21-6-1
4-18--1 4-3-9-1 4-6-9-1 4-9-9-1 4-12-9-1 4-15-9-1 4-18-9-1 4-21-9-1
4-21--1 4-3-12-1 4-6-12-1 4-9-12-1 4-12-12-1 4-15-12-1 4-18-12-1 4-21-12-1
4-3-15-1 4-6-15-1 4-9-15-1 4-12-15-1 4-15-15-1 4-18-15-1 4-21-15-1
4-3-18-1 4-6-18-1 4-9-18-1 4-12-18-1 4-15-18-1 4-18-18-1 4-21-18-1
4-3-21-1 4-6-21-1 4-9-21-1 4-12-21-1 4-15-21-1 4-18-21-1 4-21-21-1
Network having a single hidden layer with 9 neurons
17. Dehradun
Actual-2001
Simulated-2001
Moran index of 0.29 (Moran Index of 0.29 )
(Percentage accuracy of 74%)
18. Simulated-2005 Actual-2005
Moran index of 0.33 (Moran Index of 0.30 )
(Percentage accuracy of 76%)
20. Conclusions
•ANN helps to reduce the calibration time and
subjectivity in the modelling process
•GIS is used for handling of spatial data , to obtain site
attributes and training data for neural network.
• ANN is used to reveal the relationships between future
urban growth probability and site attributes