GIS In EA 2003

Workshop on Time change analysis

J. M. Stroh

Objectives: Learn one method to analyze change in vegetation cover over time. The workshop assumes a good working knowledge of Idrisi, understanding of the differences in Landsat Thematic Mapper bands (in Jensen), and the analysis of vegetation using the Normalized Difference Vegetation Index (in Jensen).

We will use the Cooper Point area of Thurston County for this workshop. The necessary files are in the folder called Cooper. Make a new folder on C:\temp called change-xxx where xxx are your initials. Cooper has 3 sub folders Copoper85, Cooper90, Cooper95; one raster called CP-mask, 2 vector coverages, Thurpoly and Tumrd83ft, and 1 palette, thresh.smp. Copy Cooper to C:\temp\change-xxx. In Idrisi make C:\temp\change-xxx your working folder and Cooper a resource folder.

Ms. Veena Tabbutt of Thurston County Regional Planning kindly provided the original thematic mapper images for three different years 1985, 1990, and 1995. They came in ERDAS Imagine format. The conversion to Arc Grid files, then to Idrisi raster files was fairly straight forward. The MS Word document "Instructions for Ben" contains the procedure. Thanks to Ben Beug (CAL staff) for doing the 1990 and 1995 years. You can read the "Instructions for Ben" document if you want to know the methodology of going from Erdas imagery to Arc Grid then to Idrisi.

We have the following metadata:

Path, Row, Date of Acquisition, Description

46 27/28 05/31/1995 Landsat 5 TM data. Zero percent cloud cover.

46 27/28 09/22/1990 Landsat 5 TM data. Zero percent cloud cover.

46 27/28 08/23/1985 Landsat 5 TM data. Zero percent cloud cover.

All data in Washington State Plane zone 2 (south) NAD83, feet.

The information below is provided for TM band review. SOURCE: Land Cover Change Report Thurston County regional Planning 2002 originally from LILLESAND AND KEIFER, 1994.

Band 1. 0.45-0.52 Blue Designed for water body penetration, making it useful for coastal water mapping. Also useful for soil/vegetation discrimination, forest type mapping, and cultural feature

identification.

Band 2. 0.52-0.60 Green Designed to measure green reflectance peak of

vegetation for vegetation discrimination and vigor assessment. Also useful for cultural feature identification.

Band 3. 0.63-0.69 Red Designed to sense in a chlorophyll absorption region

aiding in plant species differentiation. Also useful for cultural feature identification.

Band 4. 0.76-0.90 Near infrared Useful for determining vegetation types, vigor and biomass content, for delineating water bodies, and for

soil moisture discrimination.

Band 5. 1.55-1.75 Mid-infrared Indicative of vegetation moisture content and soil moisture. Also useful for differentiation of snow from

clouds.

Band 6. 10.4-12.5 Thermal infrared Useful in vegetation stress analysis, soil moisture discrimination, and thermal mapping applications.

Band 7. 2.08-2.35 Mid-infrared Useful for discrimination of mineral and rock types. Also sensitive to vegetation moisture content.

This workshop follows closely a tutorial in Change and Time Series Analysis, Eastman and others, 1995. The document is available in the CAL. View as many of the 21 files in the Cooper folder as you want to get a feeling for these data. Running HISTO will show the different frequency of radiation intensity for the different bands.

Make a false color composite (24 bit) for the 1985 and 1995 years. Clearly the vegetation (in red) dominates the landscape, but also observe the more urbanized areas. Add the vector road layer Tumrds and you will see highway 101, Campus Parkway, Cooper Point Road and many other roads. Notice the excellent registration of the satellite imagery (TM roads are in the right x-y coordinates).

Now you will attempt to determine differences in land use as indicated by changes in vegetation between 1985 an 1995. First notice a potential problem created by the different acquisition months. We can do nothing about this, except to expect some bias.

Use Convert to convert the red band 3 and near IR band 4 from integer to byte for all 3 years. Give the byte images new names. Make an NDVI image for 1985 using the byte images and the module NDVI in the Image Processing - Transformation - Vegindex menu. Call it ndvi85. Make an NDVI image for 1995. Call it ndvi95. Notice these have both negative and positive real numbers. To make comparison easier we are going to convert the range of values to byte format with a range of 0-255. Do this with STRETCH using the linear option and all other values at default. Call the stretched ndvi85 image ndvi85st and the ndvi95 image ndvi95st. View with the gray256 palette. For you skeptics, does this transformation change the overall pattern of these data (stretch often does)? Make a histogram of ndvi85, and one of ndvi85st and compare on screen. Is the overall pattern the same? If so we can proceed. Note that the details might be different because the "bin" size is different in the 2 histograms.

Our analysis will use differences between the 1995 and 1985 image. Use OVERLAY to subtract ndvi85st from ndvi95st (ndvi95st-ndvi85st) and call the result 95-85. Run a histogram on 95-85. In an ideal world the result would look like a normal distribution with a mean of 0. What is the actual mean? Note that if vegetation changes to + values of NDVI these pixels have more or more vigorous vegetation while - values indicate the opposite. For example a change from forest to cleared land would have a very negative value on the difference image.

Ignore the difference of the mean from 0. Copy the mean and the standard deviation numbers down. Assume that a substantial, really meaningful change is indicated by a change of the NDVI difference of 2 or more standard deviations, ether + or -. RECLASS the difference image 95-85 so all negative values less than 2 standard deviations from the mean have a value of 1, all negative and positive values within 2 standard deviations of the mean have a value of 0, and all positive values more than 2 standard deviations from the mean have a value of 2. Call it 2stddev. The reclassed image should have only 1,0 and 2 for values. If it has anything else you reclassed it incorrectly. View 2stddev with the thresh palette provided (or one of the qualitative palettes). What we have done is create a threshold map of change. Does this make sense?

Run area on 2stddev in tabular format. What threshold category, more vegetation change (2) or less vegetation change (1) has the greater area? Did you expect this, explain?

Because the boundary of the map is larger than Cooper Point I have provided a "mask" Boolean image called cp-mask that can be used to mask out all but the Cooper Point area. Note the lower limit is entirely arbitrary, drawn for convenience between southernmost Budd Inlet and Mud Bay. Use OVERLAY to multiply cp-mask times 2stddev. Run area in tabular mode on just the Cooper Point part of the map. Again, what threshold category, more vegetation change (2) or less vegetation change (1) has the greater area? Did you expect this, explain?

Now go back and do all the threshold analysis, including area, but with a threshold of 1 standard deviation, not 2. Notice some really unreasonable values crop up where they should not. Where does this occur? But, after masking the Cooper Point area, what do you think gives the best representation of change, 1 or 2 standard deviation for the threshold, or is there any best threshold?

As a first approximation this technique gives reasonable results for change. Critique the technique for this particular workshop.