DAIS Process Document for the King County Department of Natural Resources

For LandCover Classification of Multispecrtal Imagery for King County and Southwestern Snohomish County


July 30, 2002
Table of Contents

I.     Introduction.. I-2

II.    Pre-processing.. II-2

Procedures. II-2

Step 1:  Run NIR to Blue Index. II-2

Step 2:  Determine Threshold. II-2

Step 3:  Apply Mask. II-2

III.   Unsupervised Classification.. III-3

Goal. III-3

Procedures. III-3

Step 1:  Isodata Cluster. III-3

Step 2:  Pervious Clusters Identification. III-3

Step 3:  Impervious Category Labels Assignment III-3

IV.   Supervised Refinement.. IV-3

Goal. IV-3

Procedures. IV-3

Step 1:  Recode. IV-3

Step 2:  Refine recode. IV-4

Step 3:  Run hybrid supervised/unsupervised categorization. IV-4

V.    Minimum Mapping Unit (MMU) V-4

Goal. V-4

Procedures. V-4

Step 1:  File Recode. V-4

Step 2:  Clump. V-4

Step 3:  Clump Elimination. V-4

Step 4:  Mask. V-4

VI.   Clipping.. VI-5

Goal. VI-5

Procedures. VI-5

Step 1:  Clip. VI-5

Step 2:  Name to scheme. VI-5

DAIS Process Document for the King County Department of Natural Resources

For LandCover Classification for Multispectral Imagery for King County and Southwestern Snohomish County

I.                   Introduction

The following instructions are the steps taken by Marshall and Associates, Inc. (MARSHALL) for processing the DAIS product for the King County Department of Natural Resources (KCDNR). 

II.                Pre-processing


The goal of this phase is to produce all pixels potentially impervious while keeping the number of pervious pixels to a minimum.


Step 1:  Run NIR to Blue Index

1.1   Run generic index, using blue (band 1) and NIR (band 4) for (4-1)/(4+1).

1.2   Input float point.

1.3   Output floating point single, continuous (do NOT ignore 0.0 calculation).

Step 2:  Determine Threshold

Threshold the index to determine a spectral impervious stratum containing all impervious pixels (roads, sidewalks, roofs, etc.) with as few commission errors as possible (e.g. non-constructed bare surfaces).  The index normalizes for the effects of shadows, reducing the differential spectral response of the sunlit and shadowed portions of the imagery.

2.1   Determine index threshold for masking. 

2.2   Display index image in pseudocolor, “color up” until appropriate cutoff point is reached.

Step 3:  Apply Mask

Apply threshold value from step 2 to mask out all impervious pixels.  The resulting file is the impervious strata.


III.              Unsupervised Classification


The goal of this phase is to spectrally refine the impervious strata by using all four bands to produce distinction between pervious and impervious.


Step 1:  Isodata Cluster

Cluster the impervious strata file with the Isodata algorithm (unsupervised classification).  Clustering options in ERDAS Imagine include ‘Initialize from the statistics,’ using the principle axis to initialize and scaling the range of signatures with ‘Automatic.’  The other parameters, such as number of clusters, number of processing iterations and convergence threshold are determined on the basis of the image feature variability. Note:  There is a trade off in image analyst cost between creating too many clusters to be labeled and not creating enough clusters to identify all image features.

Step 2:  Pervious Clusters Identification

It is possible that one or more of the impervious stratum clusters could be uniquely associated with a pervious surface.  This proposition can be tested by observing whether any of the pixels associated with a cluster from the impervious stratum are found only in pervious category training sets and not in any of the impervious training sets.  Shadow may cause non-impervious features to look like some types of impervious.  If pixels in the impervious stratum occur in shadow, a representative sample should be analyzed to determine if they are true impervious or rather shadowed versions of pervious categories (e.g., forest and open water).  If they are true impervious, proceed to steps 2 and 3.

Step 3:  Impervious Category Labels Assignment

Assign the value of 1 for High Confidence impervious, 50 for Low Confidence Impervious, and 0 for other.  This also tracks error of commission, allowing for further revision either with ancillary data, heads up analysis, or cluster busting.

IV.            Supervised Refinement


The product from these processing steps is the final categorization.  The goal is to spectrally refine low confidence and/or mixed pixels by utilizing supervised signature creation.


Step 1:  Recode

1.1   Systematically identify signature from each mixed cluster confused feature.

1.2   For each missed cluster, prominently color sample areas of mixed clusters to create unique signatures.

Step 2:  Refine recode

2.1   Use the seed properties to create a spectral signature by seeding an Area of Interest (AOI) for the various features within the confused cluster. The seeding properties (AOI growing properties) are set by the spectral Euclidian distance with the darker signatures requiring a smaller distance then the brighter.

2.2   Add to unsupervised signature file.

2.3   Remove corresponding mixed unsupervised sigs from sig file.

Step 3:  Run hybrid supervised/unsupervised categorization

Use a ‘supervised’ clustering algorithm, such as maximum likelihood, to create a revised cluster image.  In the above steps the signatures have all been developed through the collection of spectral signature statistics, the signatures are thus parametric and only parametric rules apply.

V.               Minimum Mapping Unit (MMU)


The goal of this phase is to produce recoded/refined clumps at the desired MMU, 0.0025 acres, regardless of impervious surface classification confidence.


Step 1:  File Recode

Create two files, recoding the cluster busted, labeled files to one file of 0s (background and pervious surfaces) 1s, and 50s and another file of 1s (impervious high and low confidences) and 2s (pervious surfaces, including those excluded from the impervious stratum, background).

Step 2:  Clump

Clump the 1s and 2s file, using the ERDAS clump function, utilizing four orthogonally connected pixels.

Step 3:  Clump Elimination

Run the clump file through the ERDAS Elimination command and eliminate clumps 40 pixels or less (determined by the desired MMU).

Step 4:  Mask

Take the elimination file as a mask to the 0s, 1s, and 50s file, removing from it all cells that are not 1 on the 1 and 2 eliminated file.

This is accomplished by the following Boolean arguments:

4.1   If the product of step1 is >0 and the product of step 4=1, then return the value of step 1.

4.2   If the product of step 1=0 and the product of step 4=1, then return a value of 50.

4.3   If the product of step 4=2, then return a value of 0.

4.4   If the product of step 4=0, then return a value of 0.





VI.            Clipping


The goal of this phase is to create a series of files indexed to tiling scheme.  This allows the data to be viewed and distributed in smaller file size pieces.


Step 1:  Clip

Clip to KCDNR tiles and name as indicated.

Step 2:  Name to scheme

Name to scheme supplied by KCDNR.