Regression_Presentation2
Download as PPT, PDF0 likes151 views
The document describes a regression model developed to predict percent built-up land cover in Pucallpa, Peru using normalized difference vegetation index (NDVI) values derived from Landsat imagery. Landsat and Google Earth imagery were analyzed to determine percent built-up land cover around sample points. NDVI and percent built-up land cover were then used to develop a regression model. The model was able to predict percent built-up land cover with an R-squared value of 0.776, providing planners and managers a low-cost tool for rapid urban area assessment.
Regression_Presentation2
- 1. A REGRESSION MODEL FOR PREDICTING PERCENT BUILT-UP LAND COVER FROM REMOTELY SENSED IMAGERY OF PUCALLPA, PERU Presented by: Drake H. Sprague M.A. Candidate Advisor: Dr. Maria Garcia-Quijano Department of Geosciences Florida Atlantic University Boca Raton, Florida
- 2. Cities in LDCs absorb annually 20 – 30 million new residents due to rural-to- urban migration of the poorest citizens (Smith, 2001) UN (2000) estimated 74% of Latin American population in urban areas; this is expected to increase to 81% by 2020 (UN, 2006) Numerous related long term impacts, including loss of most fertile agricultural lands (Imhoff et al., 1997) Loss of life and damage to property due to disasters is greater in LDC urban area than in those of developed countries (Montoya, 2003) Planners and emergency managers in LDCs urgently need timely intelligence about urban areas, however, high costs and complex analysis methods may prevent them from acquiring the information they require Introduction Worldwide, urban areas are growing rapidly
- 3. National census counts in most LDCs are infrequent due to high costs and data becomes quickly outdated (Lo, 2006) Remote Sensing methods have been used for obtaining population using high- resolution (submeter) imagery which may be too costly for use but infrequently in LDCs (Gluch et al., 2006) Moderate-resolution (10 to 30 m) imagery is, by comparison, less expensive and its uses for deriving rapid population estimates should be considered Introduction How can agencies in LDCs account for rapid urban population growth?
- 4. Comparative costs / km² Introduction How much does this imagery cost?
- 5. Provide planners and emergency managers in LDCs with an inexpensive tool that they can easily implement for rapid urban area assessments Purpose
- 6. Develop a method to quickly assess built-up land cover for use as a proxy for population density estimation using available resources: Moderate-resolution satellite imagery Free high-resolution imagery from Google Earth Local expert knowledge Using a regression model, predict the percentage of built-up land cover in Pucallpa, Peru, as a function of a single variable common to the Amazon region: Green vegetation Key criteria: Low cost, simple and statistically robust Objective
- 7. Null hypothesis: No relationship exists between the intensity of built-up land cover and the concentration of surrounding green vegetation in estimated from remotely sensed data using the Normalized Difference Vegetation Index (NDVI) in the city of Pucallpa, Peru. Research Questions
- 8. Literature Review Urban studies using Moderate Resolution Satellite Imagery in LDCs 1977, an allometric growth model was used to estimate the population in 13 Chinese cities using color composites from Band 5 (red) and Band 7 (infrared) from Landsat Multispectral Scanner (MSS) imagery with 79-meter spatial resolution. Best results obtained for cities of between 500,000 and 2.5 million (Lo et al., 1977). An urban planning study was done in 2000 using 20-meter SPOT (Systeme Pour l’Observation de la Terre) XS (Multispectral) imagery to analyze the growth of Ouagadougou, Burkina Faso between 1986 and 1997. The Spatial Reclassification Kernel (SPARK) algorithm was applied to distinguish between socio-economic regions within the city. Results found that the imagery could be used to accurately estimate urban growth, but was too coarse in resolution to be used with the SPARK algorithm (de Jong et al., 2000).
- 9. Literature Review Urban studies using Moderate Resolution Satellite Imagery in LDCs 30 meter Landsat ETM+ imagery was used in 2006 as a basis for monitoring the evolution of urban land cover changes in Manaus, Brazil, at the sub-pixel level using multiple endmember spectral mixture analysis (MESMA). Results found the vegetation and impervious surface features corresponded well with reference data, but soil features did not, due to limitations in the reference data (Powel et al., 2006). • The time investment, cost and complexity of the methodology in this study would be impractical for most agencies in LDCs, especially if a rapid assessment of urban areas is all that is required.
- 10. Located in Peru’s low-altitude jungle region, 155 meters above sea-level Between 74° 31’ and 74° 39’ W and 8° 18’ and 8° 26’ S Map: Gobierno Regional de Ucayali, 2006 Study Site – Pucallpa, Peru
- 11. Landsat ETM+ Sept. 2000 Image Preprocessing: • Image Subset • Stack • Register Landsat Sept. 2002 Image Preprocessing: • Image Subset • Stack • Register Intuitive Map (5 urban classes) In Situ – Pucallpa (GPS) Readjust Class Scheme (110 points) NDVI 3 x 3 Filter Derive GE %BU Coverages Co-registered Map BU% & NDVI Intersect Training Set 75% Validation Set 25% Analyze Outliers Analyze Outliers RUN MODEL Map of BU Urban Intensity Google Earth 2004 Expert Validation Stratified Random Sampling ISODATA 200 Clusters Process Flow Chart A Regression Model for Predicting Percent Built-up Land Cover Using Remotely Sensed Imagery of Pucallpa, Peru
- 12. Landsat 5 - Thematic Mapper (TM). Seven spectral bands over a ground swath of 185 × 175 km 30 x 30 m spatial resolution Landsat 7 - Enhanced Thematic Mapper Plus (ETM+) Includes the above, plus an additional Panchromatic 8th band with 15 x 15 m spatial resolution – especially useful for updating maps and monitoring urban growth (Cheng, 2000). LANDSAT IMAGING SYSTEMS Data Sources
- 13. ETM+ acquired September 7, 2000 Provided by: Centro Internacional de Agricultura Tropical (CIAT) ETM+ (or TM) acquired September 1, 2002 Provided by: Gobierno Regional de Ucayali Landsat Worldwide Reference System (WRS) Pucallpa is located within Path 006 / Row 066 Dry season acquisition date; imagery less affected by atmospheric noise than in the wet season. LANDSAT IMAGERY OF PUCALLPA
- 14. 2004 Google Earth – Digital Globe’s QuickBird Satellite Imaging System DATA SOURCES
- 15. 2005 Air photo Mosaic – Fuerza Aérea del Perú (FAP) DATA SOURCES N
- 16. Landsat Scenes of Study Area DATA SOURCES September 1, 2000 185 × 170 km swath (Natural Color) Red – Band 3, Green – Band 2, Blue - Band 1 20 × 16 km (320 km²) (False Color) Red – Band 4, Green – Band 3, Blue - Band 2
- 17. Project to the Universal Transverse Mercator (UTM), Zone 18 South on the World Geodetic System (WGS) 1984 Horizontal Datum No atmospheric correction was applied to either image Landsat image. Most remote sensing studies involving imagery of a single date forego this procedure as it is considered unnecessary (Song et al., 2001) Landsat Imagery - Preprocessing
- 18. The Iterative Self-Organizing Data Analysis Technique (ISODATA) clustering algorithm was used to stratify the Landsat image into a set of 200 clusters: ISODATA changes the number of clusters by merging, splitting, and deleting as it passes (or iterates) through the RS data. With each iteration, the algorithm evaluates the statistics of the clusters. It will merge two clusters if the distance between their mean points is less than a predefined minimum distance. It will split a single cluster if its standard deviation is greater than a predefined maximum value. Or, if a cluster has fewer than the minimum specified number of pixels, it will be deleted. This is repeated to cluster sets, until either no significant change in the cluster statistics exists, or it has reached the maximum number of iterations (Lillesand et al., 2004) After masking pixels falling outside of the urban areas, the remaining clusters were manually collapsed into five nominal urban / built-up intensity levels. Urban / Built-up Intensity Map
- 19. 200 spectral clusters Reduced to 5 urban classes
- 20. Stratified random sampling - sampling procedure for testing an area that has been subdivided into land-cover strata through an image classification scheme. It assigns a minimum number of sample points to each land-cover category so that the size of each category is the same regardless of its areal extent in proportion to total size of the study area (Jensen, 2005). 36 sample points assigned per strata (180 sample points) based on the estimated time and cost required in referencing each point. Some points might fall in areas inaccessible for in situ referencing, so that up to six points per category could be left unsampled without compromising the statistical validity of the model. Sampling Design
- 21. Pucallpa field guide from 2004 Google Earth imagery Other field tools: 1:65000 planning map provided by the Gobierno Regional de Ucayali Handheld GPS for in situ referencing of the sample points In Situ Referencing
- 23. IN SITU REFERENCING Moto-taxi: traditional mode of transportation used for in situ referencing
- 24. IN SITU REFERENCING - PUCALLPA Central Pucallpa with paved roads and concrete structures
- 25. IN SITU REFERENCING - PUCALLPA Areas prone to flooding
- 26. IN SITU REFERENCING - PUCALLPA Central, heavily built region of Pucallpa
- 27. IN SITU REFERENCING - PUCALLPA Northern fringe region with hastily built wooden structures
- 28. IN SITU REFERENCING - PUCALLPA Informal settlement with hastily built structures in the southern fringe region
- 29. IN SITU REFERENCING - PUCALLPA Predominantly residential area with concrete structures
- 30. IN SITU REFERENCING - PUCALLPA Southern fringe region with open fields and scattered wooden structures
- 31. IN SITU REFERENCING - PUCALLPA Central, heavily built areas
- 32. IN SITU REFERENCING - PUCALLPA Recently-settled area in western fringe region
- 33. IN SITU REFERENCING - PUCALLPA Scattered development in western fringe
- 34. IN SITU REFERENCING - PUCALLPA Eastern extent of old-city Pucallpa along Ucayali River
- 35. IN SITU REFERENCING - PUCALLPA Recent settlement in northwestern region
- 36. IN SITU REFERENCING - PUCALLPA Large informal settlement along the extreme southern fringe
- 37. IN SITU REFERENCING - PUCALLPA Aerial view of Pucallpa heading northeast
- 38. IN SITU REFERENCING 158 points referenced in situ. Remaining 22 points not referenced due to inaccessibility, i.e., located in marshland, jungle, deep within private property, etc.
- 39. Meeting with officials from the National Institute of Statistics and Informatics (INEI) to validate the referenced sample points . Due to time constraints, 110 total points were validated using a municipal planning map. Each point was assessed according to its land use and approximate population density per hectare (city block). From this a new schema was derived: 5 population density categories Expert Validation
- 40. First component for building a regression model: amount (%) of built-up land cover (buildings, roads, and bare soil ) at each reference location 30-meter buffer around each sample point: account for possible GPS positional errors and the 30 × 30 m resolution of Landsat imagery Attempt was made to register subsets of the air photo mosaic then digitize polygons representing built-up features This was too time intensive due to image distortion and without a planimetric map Source: Fuerza Aérea del Perú, 2005 Data Processing
- 41. DATA PROCESSING – BUILT-UP LAND COVER Second attempt successful using Google Earth Sample points dropped onto the GE scene of Pucallpa and a screenshot was then acquired at each point Corner coordinates for each screenshot recorder for image-to-image rectification by coordinates to Landsat imagery x, y -adjustment was necessary due to GE’s Simple Cylindrical projection
- 42. DATA PROCESSING – BUILT-UP LAND COVER 30-meter buffer around point BU areas manually digitized
- 43. NORMALIZED DIFFERENCE VEGETATION INDEX (NDVI) Two spectral bands, Red and Near Infrared (NIR), to estimate the presence and health of vegetation within a given area. Exploits the typical spectral behavior of healthy green vegetation, with an absorption feature on the red portion of the electromagnetic spectrum due to photosynthetic pigments, and high reflectance in the NIR region due to the spongy mesophyll (Rouse et al. 1974). Range of values between -1.0 and 1.0. Highly vegetated areas will typically have NDVI values greater than 0.4. NDVI = (NIR - Red) / (NIR + Red) At 30 x 30 m, NDVI data is sensitive to contextural information, such as small areas of water – these might skew the results and cause inaccuracies in the regression model. To minimize these effects, a 3 x 3 low-pass filter was applied to ‘smooth’ and generalize the data. DATA PROCESSING
- 44. 2002 NDVI Darkest areas represent least vegetated regions
- 45. 2002 NDVI After data ‘smoothing’ with 3 x 3 Low-Pass Filter
- 46. BU % = a + b × NDVI + e Intercept a is the point where the line intersects the vertical y-axis. Slope b represents the change in the dependent variable expected from a unit change in the independent variable. Residual term e indicates that there is a difference between the predicted y and observed y in the paired data (Rogerson, 2001). NDVI values were extracted at each of the sample points in ArcMap. Sample data split into two groups: Calibration set 75% / Validation set 25% Regression Analysis
- 47. Of 82 original training points, one was discarded from the model due possibly to exaggerated effects on the NDVI from nearby water, particularly after applying a low pass filter. NDVI at this location was relatively low at 0.0989; however, predicted BU % was very high at over 95%. INITIAL TEST - EXTREME OUTLIER
- 48. Predicted BU % = 84.141 - 229.581 × NDVI The slope in this equation is -229.581 shows that an increase in NDVI value of 0.01 will result in an average 2.30% decrease of built-up land cover. The intercept of this equation, 84.141 indicates that on average, an NDVI value of 0.0 will result in a predicted BU% of 84.14%. When NDVI is 1.0, predicted BU% will be 0.0. The Regression Model
- 49. RESIDUALS – DETECTING OUTLIERS Residuals: the difference between a value predicted by the regression line and the observed value for the dependent variable. Points should be homogenously distributed along the curve (above and below)
- 50. REGRESSION ANALYSIS Regression model re-run using 81 of the 82 training points
- 51. REGRESSION ANALYSIS 0.500000000000.400000000000.300000000000.200000000000.100000000000.00000000000-0.10000000000 NDVI3X3 100.000000000000 80.000000000000 60.000000000000 40.000000000000 20.000000000000 0.000000000000 %BuiltUp 81 8079 78 77 76 75 74 73 72 71 70 69 68 67 6665 64 63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 3837 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 54 3 2 1 R Sq Linear = 0.776 Regression model re-run using 81 of the 82 training points
- 52. OUTLIERS This location was observed and digitized as 95.97% built-up; whereas the model predicted it as 61.42% built-up
- 53. VALIDATION OF REGRESSION MODEL Testing the model using the remaining 28 validation points revealed a mean predictive error of 7.6% and a standard deviation of 27.02
- 54. MINIMUM ERROR Observed built-up land cover was 5.7%; the model predicted 76.34%
- 55. MAXIMUM ERROR Observed built-up land cover was 67.45%; the model predicted 0% built-up
- 56. URBAN BUILT-UP INTENSITY MAP
- 57. URBAN BUILT-UP INTENSITY MAP
- 58. URBAN BUILT-UP INTENSITY MAP
- 59. NDVI MAP
- 60. NDVI MAP
- 61. NDVI MAP
- 62. Evaluation of Google Earth Served as guide for compiling an initial urban density map and for extracting built-up land cover information. Contains a wealth of satellite and aerial imagery availabile at no cost anywhere a connection with the Internet can be established. Much of the imagery is available at high (approximately 1 meter) spatial resolution. Google Earth imagery is georeferenced; it can substitute in some cases for digital planimetric maps. Much of its imagery is at least two years old; imagery of Pucallpa was three years old. Google Earth is still preferable to relying on aerial imagery 10 years or more out of date, or no imagery at all.
- 63. Strong relationship between these variables - thus the null hypothesis of no relation was rejected – this makes possible the use of a regression model to make rapid assessments of built-up land cover in Pucallpa and other places similar to it. Moderate-resolution imagery is considered best suited for urban analysis at a regional rather than local scale (Gluch, 2006). However, when combined with high-resolution imagery, such as provided by Google Earth when available, the potential uses of moderate-resolution imagery are multiplied. Conclusions & Recommendations
- 64. STATISTICAL ROBUSTNESS: Regression model successfully explained 77.4% (R² = .774) of the variability observed in the %BU land cover. COSTS: Cost of Imagery: For this project, imagery cost was $0.00. A planning agency will need to factor at least $600.00 for a Landsat image Cost in Time: In situ referencing of 158 points (4-5 days) Capturing and georectifying 158 GE screenshots (3 days) Expensive Software used: ERDAS Imagine, ESRI ArcMap COMPLEXITY: Basic image processing and GIS procedures used throughout Conclusions & Recommendations ASSESSMENT OF METHODOLOGY
- 65. Conclusions & Recommendations QUESTIONS: How will this model perform in other parts of the Amazon region and in other world regions? What will be the seasonal effects of vegetation on the model? How can this model be adapted to derive more detailed information about the human populations found within built-up regions? What are the tradeoffs of adding additional variables to the model to increase its predictive capabilities?
- 66. Future Work LandScan: a global population database of the United States Department of Energy’s (USDOE) Oak Ridge National Laboratory (ORNL) Global Population Project (land cover, roads, slope, and night time lights ) Peru recently conducted its first national census since 1993 – results should be used for further study into Pucallpa’s population dynamics Other low-cost imagery sources (ASTER, ALI) should be considered as alternatives to imagery produced by the aging Landsat constellation. Free or low-cost GIS systems, such as SPRING of Brazil’s National Institute for Space Research and IDRISI, should be used to further enhance overall practicality of these methods for use by agencies in LDCs
- 67. The End
Editor's Notes
- #50: *