A Comparison of Land Use and Land Cover

A Comparison of Land Use and Land Cover Change Detection Methods Daniel L. Civco, James D. Hurd, Emily H. Wilson, Mingjun Song, Zhenkui Zhang Center for Land use Education And Research Department of Natural Resources Management & Engineering The University of Connecticut U-4087, Room 308, 1376 Storrs Road Storrs, CT 06269 -4087

Outline • • Background Objectives Study Area and Data Methods – – Post Classification Analysis Cross-correlation Analysis Neural Networks Segmentation & Object-oriented Classification • Results • Conclusions • Recommendations

Northeast Applications of Useable Technology In Land planning for Urban Sprawl A NASA Regional Earth Science Applications Center (RESAC)

Our RESAC Mission To make the power of remote sensing technology available, accessible and useable to local land use decision makers as they plan their communities. To educate the general public on the value and utility of geospatial technologies, particularly RS information.

NAUTILUS Research Better land cover mapping and change detection Urban growth models and metrics Forest fragmentation models and metrics Improved impervious cover estimates

Background • Need for effective methods for deriving information on – Land use change – Forest fragmentation – Urban growth – Loss of agricultural lands – Increase in impervious surface area

Background Farmland conversion in the Stony Brook Millstone Watershed: 1995 -1999 Genesis 3 D rendering of land use derived from Landsat data

Objective Compare the results of different land use and land cover change detection approaches • traditional post-classification • • cross-tabulation cross-correlation analysis neural networks knowledge-based expert systems image segmentation & objectoriented classification

Research & Education Watersheds Presumpscot Su. As. Co Salmon Stonybrook A range of land covers and issues

Stony Brook Millstone Watershed, NJ • Has a strong Watershed Association in existence • Between New York City and Philadelphia • Increased development pressures • Loss of agriculture land to urban sprawl USGS MRLC 265 Sq. Miles

Study Area and Data SITE 1 SITE 2 Stony Brook Millstone Watershed

Study Area and Data March 27, 1989 Site 1 May 4, 2000

Study Area and Data September 3, 1989 Site 1 September 23, 1999

Study Area and Data March 27, 1989 Site 2 May 4, 2000

Study Area and Data September 3, 1989 Site 2 September 23, 1999

Study Area and Data September 3, 1989 September 23, 1999

Methods – Post Classification Analysis – Cross-correlation Analysis – Neural Networks – Segmentation & Object-oriented Classification

Post Classification Analysis 1989 Classification Iteration 1 • Step 1 – Unsupervised classification • Identify known clusters • Extract unknown clusters Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren Unknown

Post Classification Analysis 1989 Classification Iteration 2 • Step 2 – Unsupervised classification • Identify known clusters • Extract unknown clusters Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren Unknown

Post Classification Analysis 1989 Classification Iteration 3 • Step 3 – Unsupervised classification • Identify known clusters • Extract unknown clusters Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren Unknown

Post Classification Analysis 1989 Classification Iteration 4 • Step 4 – Unsupervised classification • Identify clusters Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren Unknown

Post Classification Analysis • Step 5 – Combine iterations into single land cover image Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren

Post Classification Analysis • Step 6 – Smooth image using majority filters Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren

Post Classification Analysis Perform similar procedure on 2000 date 1989 Classification Site 1 2000 Classification

Post Classification Analysis 1989 Classification Site 2 2000 Classification

Post Classification Analysis 1989 Classification 2000 Classification Cross Tabulate Site 2 Change

Methods – Post Classification Analysis – Cross-correlation Analysis – Neural Networks – Segmentation & Object-oriented Classification

Cross-correlation Analysis CCA calculates the sum of the distance of each pixel in each band from the norm • • Z is the distance measure Observed is the pixel value for each band Expected is the mean value of all extracted pixels for each band Std. Dev. Is the standard deviation of all extracted pixels for each band

Cross-correlation Analysis 1989 Deciduous Category • Step 1 – Use 1989 classification as base land cover – Extract vegetated class areas to be analyzed from 1999/2000 ETM image (turf & grass, agriculture & barren, deciduous, and coniferous) September 23, 1999

Cross-correlation Analysis 1989 Deciduous Category • Step 2 – Perform CCA on 1999/2000 imagery – Identify thresholds separating unchanged pixels and changed pixels Probable unchanged rangechanged Z-values from 1 to 5, 794

Cross-correlation Analysis • Step 3 – Create a mask from changed pixels for all categories analyzed (turf & grass, agriculture & barren, deciduous, and coniferous) Turf & Grass Agriculture & Barren Deciduous Coniferous

Cross-correlation Analysis • Step 4 – Extract changed pixels from 1999/2000 image data – Perform unspervised classification to identify new categories Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren

Cross-correlation Analysis • Step 4 – Merge new classes with historic classification to produce updated land cover Dense Urban Turf & Grass Residential Agriculture Deciduous Coniferous Water Wetland Barren

Cross-correlation Analysis 1989 Classification Site 1 2000 Classification

Cross-correlation Analysis 1989 Classification Site 2 2000 Classification

Methods – Post Classification Analysis – Cross-correlation Analysis – Neural Networks – Segmentation & Object-oriented Classification

Neural Networks Nautilus Image Processing System

Neural Networks • Step 1 – Select training features based on points

Neural Networks • Step 2 – Extract digital numbers for: • Each pixel – By class • Each Band

Neural Networks Records 259 Features 7 Classes 9 Band 1 Class actually represented by one-of-n encoding. (i. e. , 0 0 0 1 0 0 0 Band 2 Band 3 Band 4 Band 5 Band 7 Band 6 Class 69 28 22 151 103 129 30 4 65 26 21 167 107 129 31 4 …. …. 77 35 42 66 77 135 41 9 79 34 46 69 82 135 43 9 …. …. 61 20 15 83 53 124 13 5 60 20 15 83 50 124 12 5 …. …. Example of Site 2 Training Data for September 23, 1999

Neural Networks Step 3: Create data set of all possible from T 1 to T 2 changes (constrained by permitted changes)

Neural Networks From To Urban Residntl Turf&Grass Agric Decid Conif Water Wetland Barren Urban Residential Turf&Grass Agriculture Deciduous Coniferous Water Wetland Barren Permitted Changes

Neural Networks • Step 4 – Create neural network classifier • Neural. SIM® – Backpropagation – Export C-code – Compile into NIPS

Neural Networks • Step 5 – Perform full neural networkbased change detection within NIPS

Methods – Post Classification Analysis – Cross-correlation Analysis – Neural Networks – Segmentation & Object-oriented Classification

Segmentation and Objectoriented Classification • Step 1: Preprocessing – Prepare Image Data (2 dates and 2 seasons = 4 images) – Use indices to extract obvious classes – Create a data layer using the knowledge engineer – Add to image data for input into e. Cognition

Segmentation and Objectoriented Classification • Step 1: Preprocessing – Prepare Image Data (2 dates and 2 seasons = 4 images) – Use indices to extract obvious classes – Create a data layer using the knowledge engineer – Add to image data for input into e. Cognition Date 1 Spring NDVI Date 1 Spring NDMI Date 2 Spring NDVI Date 1 Summer NDMI Date 2 Spring NDMI Date 2 Summer NDVI Date 2 Summer NDMI

Segmentation and Objectoriented Classification • Step 1: Preprocessing – Prepare Image Data (2 dates and 2 seasons = 4 images) – Use indices to extract obvious classes – Create a data layer using the knowledge engineer – Add to image data for input into e. Cognition

Segmentation and Objectoriented Classification • Step 1: Preprocessing – Prepare Image Data (2 dates and 2 seasons = 4 images) – Use indices to extract obvious classes – Create a data layer using the knowledge engineer – Add to image data for input into e. Cognition Site 1 Site 2

Segmentation and Objectoriented Classification • Step 2 – Create e. Cognition Projects for date 1 and date 2 – Segment images using both seasons of imagery (excluding layer 7) • Input Data: 7 layers • • 1. Red (spring) 2. NIR (spring) 3. MIR (spring) 4. Red (summer) 5. NIR (summer) 6. MIR (summer) 7. Classified layer

Segmentation and Objectoriented Classification • Step 2 – Create e. Cognition Projects for date 1 and date 2 – Segment images using both seasons of imagery (excluding layer 7) • Segment four levels Level 1 2 3 Scale Color Shape 3 1. 0 0. 0 5 0. 8 0. 2 10 0. 7 0. 3 4 20 0. 5

Segmentation and Objectoriented Classification • Step 2 – Create e. Cognition Projects for date 1 and date 2 – Segment images using both seasons of imagery (excluding layer 7) Pixel 4 Level 1 2 3

Segmentation and Objectoriented Classification • Step 3 – Begin creating class hierarchy – Training segment selection and standard nearest neighbor – Nearest Neighbor Classification of each level

Segmentation and Objectoriented Classification • Step 3 – Begin creating class hierarchy – Training segment selection and standard nearest neighbor – Nearest Neighbor Classification of each level

Segmentation and Objectoriented Classification • Step 3 – Begin creating class hierarchy – Training segment selection and standard nearest neighbor – Nearest Neighbor Classification of each level

Segmentation and Objectoriented Classification • Step 3 – Begin creating class hierarchy – Training segment selection and standard nearest neighbor – Nearest Neighbor Classification of each level

Segmentation and Objectoriented Classification • Step 4 – Adding knowledge to each e. Cognition project – Refinement and final classification with class-related features Utilize other the layer spatial levels attributes 7 mask Water in The classified Level 1 is layer based must Turf and be solely classon 3 Grass the must existence border of water residential in or Level other 3 turf and grass Water in Level 3 is based on the summer red band (layer 4) and the standard nearest neighbor samples Turf and Grass must have residential in level 3

Segmentation and Objectoriented Classification • Step 4 – Adding knowledge to each e. Cognition project – Refinement and final classification with class-related features Site 2, 1, Date 2 1

Segmentation and Objectoriented Classification • Step 5 – Use the knowledgebased classifier in ERDAS Imagine to do a postclassification change detection – Final change classifications for Site 1 and Site 2

Segmentation and Objectoriented Classification • Step 5 – Use the knowledgebased classifier in ERDAS Imagine to do a postclassification change detection – Final change classifications for Site 1 and Site 2 1

Results – Post Classification Analysis – Cross-correlation Analysis – Neural Networks – Segmentation & Object-oriented Classification

Post Classification Analysis Site 1 Urban Agriculture Site 2 Forest Water Agr to Urban Barren Forest to Urban Barren to Urban

Cross-correlation Analysis Site 1 Urban Agriculture Site 2 Forest Water Agr to Urban Barren Forest to Urban Barren to Urban

Neural Networks Site 1 Urban Agriculture Site 2 Forest Water Agr to Urban Barren Forest to Urban Barren to Urban

Segmentation and Object-oriented Classification Site 1 Urban Agriculture Site 2 Forest Water Agr to Urban Barren Forest to Urban Barren to Urban

September 3, 1989 Post-classification Change Detection September 23, 1999 Agriculture To Urban Forest To Urban Barren To Urban

September 3, 1989 Cross-Correlation Change Detection September 23, 1999 Agriculture To Urban Forest To Urban Barren To Urban

September 3, 1989 Neural Network Change Detection September 23, 1999 Agriculture To Urban Forest To Urban Barren To Urban

September 3, 1989 Object-oriented Change Detection September 23, 1999 Agriculture To Urban Forest To Urban Barren To Urban

Conclusions • The results of this research reveal that there is merit to each of the several land use change detection methods studied, but that there appears to be no single best way in which to perform change analysis • The most significant conclusion of this study is that much research remains to be done to improve upon the results of land use and land cover change detection

Recommendations • These investigators firmly believe that an approach based on image-segmentation and rule-based classification is potentially such an improved methodology, and accordingly intend on pursuing the avenues of neural network and object-oriented classification change detection, perhaps in an integrated approach.

Acknowledgement National Aeronautics and Space Administration Grant NAG 13 -99001/NRA-98 -OES-08 RESACNAUTILUS, Better Land Use Planning for the Urbanizing Northeast: Creating a Network of Value. Added Geospatial Information, Tools, and Education for Land Use Decision Makers. Northeast Applications of Useable Technology In Land planning for Urban Sprawl

This presentation is available at resac. uconn. edu

A Comparison of Land Use and Land Cover Change Detection Methods Daniel L. Civco, James D. Hurd, Emily H. Wilson, Mingjun Song, Zhenkui Zhang Center for Land use Education And Research Department of Natural Resources Management & Engineering The University of Connecticut U-4087, Room 308, 1376 Storrs Road Storrs, CT 06269 -4087