Computational Model for Urban Growth Using Socioeconomic Latent Parameters
- 1. Computational Model for Urban Growth Using Socioeconomic Latent Parameters Piyush Yadav (Lero- Ireland ) Shamsuddin Ladha(Tata Research -India) Shailesh Deshpande(Tata Research- India) Edward Curry(Lero- Ireland)
- 2. 16/09/2018 © Lero 2015 2 • Urban Growth • Land Use Land Cover Change • Urban Growth Models and Limitation • Proposed Model • Study Area • Datasets and Pre-processing • Results • Conclusion Outline
- 3. 16/09/2018 © Lero 2015 3 World Population is growing Increased Economic Activities Increased Urban Growth Rate Urban Growth An Aerial View of a part of study area in 2006 and 2014
- 4. 16/09/2018 © Lero 2015 4 A KEY ASPECT OF URBAN GROWTH IS AFFECT ON LAND USE LAND COVER CHANGE LAND COVER INDICATES THE PHYSICAL LAND TYPE SUCH AS FOREST OR OPEN WATER LAND USE DOCUMENTS HOW PEOPLE ARE USING THE LAND SUCH AS AGRICULTURE Land Use Land Cover Change(LULCC)
- 5. 16/09/2018 © Lero 2015 5 Factors Affecting Land Use Land Cover • Predominantly, change over space but remain relatively static with respect to time. • Digital Elevation Model (DEM) Spatial Factors • Change over both time and space. • Proximity to the primary roads Spatio- temporal Factors • Change over time but spatially static for a given study area. • National Gross Domestic Product (GDP) Temporal Factors Direct Factors Indirect Factors Land Use Land Cover Change
- 6. 16/09/2018 © Lero 2015 6 Urban Growth Models Thus the lattice based spatio-temporal models, e.g. Cellular Automata (CA) and Logistic Regression (LR), are effectively used to model the spatial geographic processes. Whereas, the MC model predicts the quantum of growth along with the rate of transfer between different LULC classes (without specifying the exact growth locations). LULC images of two distinct time instances are taken and the probabilities are computed using the frequency of change from one LULC class to another and generate transition probability matrix. Urban Growth models are used for prediction of land use land cover (LULC) changes. LULC modeling is extremely difficult due to complex interactions between multi-scale factors. Schematic of an integrated Markov Chain model
- 7. 16/09/2018 © Lero 2015 7 • Markov model is built only on the cross tabulation of transition frequencies from the existing land cover images, future transition probabilities have to be estimated from the existing ones. • Some elements of the transition probability matrix for the future time period are derived based on a simple power law. • 𝒂 𝟏𝟏 is the transition probability of LC Class 1 persisting after time period 𝒕, then for time period 𝟐𝒕 the persistence rate remains unchanged and hence the new probability value is 𝒂 𝟏𝟏 𝟐 . • In world wide process of increasingly rapid urbanization the assumption of persistent rate change is unrealistic and strongly constrained. Persistent Rate of Change • It is well known that processes such as LULCC and socio-economic changes are linked and interactive (et al. Lefebvre 1991) • MC model does not allow incorporating the effect of temporal variables such as macro-economic factors, socio-economic factors, etc. Temporal Factors Limitation to Present Models
- 8. 16/09/2018 © Lero 2015 8 Our Contribution Hidden Markov Model Introduction of Hidden Markov Model (HMM) Temporal Factors Incorporate temporal factors in LULC change modelling using HMM. Model the underlying temporal factors as Gaussian distributions, conditioned on the hidden states, to learn land cover type transition probabilities Integrate Integrate our model with other spatio- temporal models such as Logistic Regression (LR) to yield richer integrated models than the corresponding MC based integrated models. An urban growth model with multi-scale direct and indirect factors impacting LULC changes
- 9. 16/09/2018 © Lero 2015 9 Our Model A Hidden Markov Model with hidden states (V, I, S) and sample emissions (GDP and Liquidity) Proposed urban growth model: HMM integrated with Logistic Regression model
- 10. 16/09/2018 © Lero 2015 10 Study Area: Pune • Tier-A city situated in the state of Maharashtra, India. • Located 560 m above the sea level. • Famous for Information Technology and Automobile industries and various research institutes. • Considered 45 sq. km of the city area which have gone under rapid urbanisation.
- 11. 16/09/2018 © Lero 2015 11 Temporal Growth Factors Gross Domestic Product National. Amount of goods and services produced within the border of a country in a specific time interval. Interest Rate Cycle National. Revised bimonthly. A tight monetary policy affects the overall investment policy which leads to slowdown and vice versa. Consumer Price Index National. Low inflation creates developmental investment environment. Gross Fixed Capital Formation National. Amount that government spends in the capital formation(such as infrastructure building, land improvements) of the country. Greater the GFCF investment higher is the rate of urbanization . Urban Population Growth Rate National. In order to accommodate a higher influx of people, cities are expanding along their outskirts, leading to the growth in urban agglomerate. Electricity Consumption Regional. Typically, regions with higher electricity demand grow faster than those with lesser demand. Road Length Added Regional. Better connectivity of a region helps in better transportation and thus provides impetus to growth by allowing setup of new industrial complexes and other infrastructure services.
- 12. 16/09/2018 © Lero 2015 12 Temporal Growth Factors Data GDP growth rate (%) Absolute average CPI Inflation (%) Gross fixed capital formation (%GDP) Urban population growth rate (%) Bimonthly interest (repo) rate (%) Per capita electricity consumption in kilowatt-hours
- 13. 16/09/2018 © Lero 2015 13 Land Use Land Cover (LULC) Data LULC data is required for HMM hidden states and LR models as an input. Time period Yearly, 2001 to 2014 (between March to April) Latitude 18.38847838°N - 18.79279909°N Longitude 73.64552005°E - 74.07494971°E Bands 1 to 7 Resolution 30m Pixels 1500 𝑥 1500 Landsat 7 Landsat-7 Specifications
- 14. 16/09/2018 © Lero 2015 14 LULC Data Pre-processing Scan Line Correction (SLC) Atmospheric Correction Solar Correction • In 2003 Landsat-7 SLC in ETM+ instrument has developed a fault thus creating some black lines in the captured images. • Image Smoothening using windowing. • Electromagnetic radiation captured by the satellite sensors is affected because of the atmospheric interference such as scattering, dispersion, etc. • Subtract the digital number (DN) of water pixels in band 4 (infrared band) as it has very low water leaving radiance. • DN values were then converted to spectral radiance. 𝑳 = 𝑳 𝒎𝒊𝒏 + 𝑳 𝒎𝒂𝒙 𝟐𝟓𝟒 − 𝑳 𝒎𝒊𝒏 𝟐𝟓𝟓 𝒙 𝑫𝑵 • For clear Landsat images, solar correction of the images was done by converting spectral radiance to exoatmospheric reflectance. 𝝆 𝒑 = 𝝅 ⋅ 𝑳 𝝀⋅ 𝒅 𝟐 𝑬𝑺𝑼𝑵 𝝀 ⋅ 𝒄𝒐𝒔𝜽 𝒔
- 15. 16/09/2018 © Lero 2015 15 • Classified into seven broad LULC classes on the basis of the nature of the landscape. • Forest Canopy, Agriculture Area, Residential Area, Industrial Area, Common Open Area, Burnt Grass, Bright Soil, and Water Body. Classes • For classification a labeled set of pixels for each class of interest was collected (500 to 3000 samples per class). The feature vector for each pixel consisted of all seven band values. • Support Vector Machines • Manual Correction (Concrete and Quarry) SVM Classification • Vegetation, Impervious Surface, and Soil VIS Classes LULC Data Classification
- 16. 16/09/2018 © Lero 2015 16 A Quick Recap LULC Data
- 17. 16/09/2018 © Lero 2015 17 Spatio-Temporal Factors Digital Elevation Model (DEM) and Slope Proximity to primary roads: Mask CARTOSAT 1 Water bodies were masked out from the LULC image 3 D View DEM Image Primary Road Layers
- 18. 16/09/2018 © Lero 2015 18 Results HMM Experiments Computed MC transition probabilities for 2001-2002, Learned HMM transition probabilities for 2014, Computed MC transition probabilities for 2014 • Used Gaussian HMM library in Scikit Learn • We designed a HMM with the three hidden states (V, I, and S) and temporal factors • HMM was initialized with MC transition probabilities for the year 2001 to 2002 • A stable model was obtained empirically after 50000 iterations with a threshold of less than 0.01
- 19. 16/09/2018 © Lero 2015 19 Results Land Change Modelling Experiments • Terrset’s Land Change Modeler. • Transition sub-models were defined for four LC change types, i.e., V to S, V to I, S to V, and S to I. • Slope gradient and primary roads layer were used as the primary driver variables . 𝒔𝒖𝒊𝒕𝒂𝒃𝒊𝒍𝒊𝒕𝒚 = 𝟏 𝒔𝒍𝒐𝒑𝒆 𝒈𝒓𝒂𝒅𝒊𝒆𝒏𝒕 𝟎.𝟏 • Suitability map. Greater the value higher the suitability and vice-versa. • Suitability for urbanization is high in areas such as roads, low lying river basin, and around the urbanized areas where the slope gradient is less. • Towards, the south end the suitability drops significantly, as the area has hills and valleys. • Four of the sub models were built using Logistic Regression.
- 20. 16/09/2018 © Lero 2015 20 Results Soil to Impervious Soil to Vegetation Vegetation to Impervious Vegetation to Soil Heat maps depicting transition probabilities from one state to another
- 21. 16/09/2018 © Lero 2015 21 • The two models were then used to predict changes for the year 2014. Results Actual land cover image of 2014 obtained from classification Predicted land cover image of 2014 (HMM-LR) Predicted land cover image of 2014 (MC-LR) • Visually it is evident that the HMM based predicted image is significantly better, in terms of similarity with the actual classified LC image than the MC based predicted image .
- 22. 16/09/2018 © Lero 2015 22 HMM-LR MC-LR V I S V I S Precision 0.48 0.49 0.60 0.54 0.38 0.34 Recall 0.48 0.52 0.59 0.54 0.32 0.39 Results • Blob Analysis of urban and non urban regions. Blobs denote concentrated urban regions. • Green blobs are true positives, blue blobs are false negatives, and red blobs are the false positives. • HMM-LR false positives are smaller in size and less dense than those of the MC-LR. The HMM output is well balanced and resembles the actual output better. • 11% increment in precision of the persistence of Impervious Surface (I) is observed. • Precision of Soil (S) class type has jumped up by 26%. • Drop in the precision of Vegetation (V) class type by a marginal 6% . This is because vegetation cover is an outcome of relatively easy process as compared to S and I . Blob Analysis of urban areas. Left to right: (i) Actual, (ii) MC-LR, (iii) HMM-LR Precision and Recall for integrated models
- 23. 16/09/2018 © Lero 2015 23 Conclusion • Markov Chain (MC) models are limited in their urban prediction capabilities due to the assumption of constant rate of persistence of land cover class types and inability to model the temporal factors. • We have proposed a new temporal model using Hidden Markov Model. • We have demonstrated the usefulness of our model over MC by predicting urban growth for an upcoming city of India (Pune). • We believe that this inquiry into HMM based models provides yet another tool that will equip the urban modelers, planners and decision makers to better design sustainable urban environments. • 11% and 26% increment of precision in Impervious Surface and Soil Class respectively.
- 25. 16/09/2018 © Lero 2015 25 Back UP SLIDES
- 26. 16/09/2018 © Lero 2015 26
- 27. 16/09/2018 © Lero 2015 27
- 28. 16/09/2018 © Lero 2015 28