Module 7
Caribbean Regional Node

To undertake a supervised classification of a Landsat TM image to show the major marine and terrestrial habitats of the Caicos Bank.
 
 

This lesson relates to material covered in Chapters 9-11 of the Remote Sensing Handbook for Tropical Coastal Management and readers are recommended to consult this for further details of the techniques involved. The lesson introduces you to multispectral classification of imagery using a simple two-dimensional box-classification of the "feature space" of two depth-invariant bottom index images.

Familiarity with Bilko for Windows 2.0 is required to carry out this lesson. In particular, you will need experience of using Formula documents (see Part 5 of the Introduction) to carry out mathematical manipulations of images. Some calculations need to be performed independently; these can either be carried out on a spreadsheet such as Excel or using a calculator.

The image used as the basis for this lesson was acquired by Landsat 5 TM on 22nd November 1990 at 14.55 hours Universal Time (expressed as a decimal time and thus equivalent to 14:33 GMT). The Turks & Caicos are on GMT - 5 hours so the overpass would have been at 09:33 local time. This image has been geometrically corrected, radiometrically and atmospherically corrected (Lesson 3), and finally water column corrected (Lesson 5) to produce two depth-invariant bottom index bands; one from bands 1 and 3 (DI_TM13.DAT) and one from bands 2 and 3 (DI_TM23.DAT). The third depth-invariant band (from bands 1 and 2) will not be used here. The subscenes provided are of the South Caicos area only and are floating point images, i.e. each pixel is stored as a floating point number and occupies four bytes. To allow a mask image to be made to mask out the land areas, you are also provided with the band 5 image of the same area (TMBAND5.GIF).

You are provided with a spreadsheet (HABITATS.XLS) containing field survey data on seven habitat classes:

1. Dense seagrass,
2. Sparse seagrass,
3. Sand,
4. Dense Montastraea reef,
5. Gorgonian plain,
6. Lobophora dominated macroalgal areas,
7. Coral patch reefs.

The first task is to mask out the land areas on the two depth-invariant bottom index images (DI_TM13.DAT and DI_TM23.DAT). We will use the near-infrared Landsat TM band 5 image to make the mask and then multiply the depth-invariant images by it.

The main task in this lesson is to classify major submerged habitats. To allow contrast stretches which will display these best and to remove the distraction of terrestrial habitats which are best classified separately using a combination of infra-red and visible wavebands, we should mask out the land areas. These is easily achieved using a Landsat TM band 5 infra-red image (TMBAND5.GIF) where there will be very little reflectance from water covered areas but considerable reflectance from land areas. This allows water and land areas to be fairly easily separated on the image and a mask of either land or water to be created with a simple Formula document.

A land mask image has all land pixels set to zero and all water pixels set to 1, so when used to multiply another image it leaves sea pixel values unchanged but sets all land pixels to zero.

This requires a formula of the type:

where @1 is the TMBAND5.GIF image. The formula takes each pixel in the @1 image and compares it to the threshold value, then IF the pixel has a value which is less than or equal to (<=) the threshold value it sets the output image pixel to 1. Otherwise (ELSE) the output image pixel is set to 0. Thus the output image has all land pixels set to 0 and all water pixels set to 1.

In this section you will use field survey data on where different habitats are located on the images (in HABITATS.XLS) to derive spectral signatures for major marine habitats and then use these signatures to classify the image. The classification method which will be tried is to create a simple box classifier for each habitat using the two depth-invariant bottom index images. That is, you are seeking to define discrete two dimensional areas in feature space which relate to specific habitats.

The first step is to find out what reflectance values in each depth-invariant band relate to which habitats.

Clearly it is not feasible to separate Lobophora dominated algal areas from coral patch reefs using just these two depth-invariant bands. Thus these two habitats need to be combined for classification.

This improves the classification scheme but two further anomalies need addressing. As is evident from Figure 6.1 the Montastraea reef class swallows the dense seagrass class because of two outliers. For the purposes of this simple box-classification it is perhaps best to risk misclassification of some of the Montastraea reef by restricting the Montastraea class to a box around the five training sites which group together (Figure 6.2). Similarly the one gorgonian plain outlier with a high depth-invariant TM band 2/3 bottom index results in a lot of overlap with the coral patch reef/Lobophora class. Restricting the gorgonian plain class box to the remaining points risks leaving gorgonian plain unclassified but should improve classification of the coral patch reef/Lobophora class. The revised box-classifier boundaries which reflect the classification scheme in Figure 6.2 are listed below in Table 6.4.

Bear in mind that we have used a very small sample of field survey points in constructing our classification and thus may be underestimating the spread of values in feature space. This could lead to a lot of the image being unclassified.

You will now try a classification based on the tight boxes in Figure 6.2 and the very limited number of training sites (field survey stations).

1. The highest pixel value in both blocks is 7. Thus pixels with values above 7 can be considered as being land.
2. On line one you should have @1 * @3 ; On line two you should have @2 * @3 ;
3. Sand and sparse seagrass have clear non-overlapping boxes and are thus clearly separable from all other habitats.
4. Coral patch reefs and Lobophora dominated algal areas occupy very similar areas in feature space for these two depth-invariant bands and are hopelessly confused spectrally. The dense seagrass category is contained within the dense Montastraea reef parallelepiped, although the majority of pixels in each class cluster in different parts of feature space.
5. There is overlap between dense Montastraea reef areas and both dense seagrass and coral patch reef habitats.
6. Because of the outlier with a relatively high depth-invariant bottom index value in TM band2_3, gorgonian plain is likely to be confused with both Lobophora dominated algal areas and coral patch reefs.

7. @1 + @2 + @3 +@4 + @5 + @6 ;
8. The main problem with the CLASSIF1.GIF file is that too much of the submerged area is unclassified. Given the broad classes involved which include most of the habitats likely to be encountered, this suggests that the training sites do not adequately represent the habitat classes in question. Also areas of water on land such as the salinas classify as various marine habitats as they have similar pixel values.
9. Much more of the submerged habitats are classified as the greater number of training sites has better sampled the spread in feature space for each habitat.

Ground-truthing data from 7 training sites for each of 7 habitat classes.
 

The CLASSIF1.FRM formula document which uses the Figure 6.2 boxes as a basis for classification.

# Formula document to classify a Landsat TM image of the shallow sea around South Caicos.

#

# This document uses two depth-invariant bottom index images DITM1.DAT (@1)

# and DITM2.DAT (@2)

#

# Dense seagrass class boundaries

CONST DenSeagMin1 = 4.46 ; CONST DenSeagMax1 = 5.05 ;

CONST DenSeagMin2 = 5.07 ; CONST DenSeagMax2 = 5.35 ;

# Sparse seagrass class boundaries

CONST SpSeagMin1 = 6.22 ; CONST SpSeagMax1 = 6.53 ;

CONST SpSeagMin2 = 5.86 ; CONST SpSeagMax2 =6.10 ;

# Sand class boundaries

CONST SandMin1 = 6.89 ; CONST SandMax1 = 7.26 ;

CONST SandMin2 = 6.32 ; CONST SandMax2 = 6.48 ;

# Lobophora dominate algal area and coral patch reef class boundaries

CONST LobCoralMin1 = 5.19 ; CONST LobCoralMax1 = 6.16 ;

CONST LobCoralMin2 = 5.42 ; CONST LobCoralMax2 = 5.87 ;

# Dense Montastraea reef class boundaries

CONST MontMin1 = 3.80 ; CONST MontMax1 = 4.42 ;

CONST MontMin2 = 4.76 ; CONST MontMax2 = 5.25 ;

# Gorgonian plain class boundaries

CONST GorgMin1 = 6.03 ; CONST GorgMax1 = 6.69 ;

CONST GorgMin2 =5.03 ; CONST GorgMax2 =5.50 ;

# Sand box-classifier

IF ( (@1 >= SandMin1) AND (@1 <= SandMax1) AND (@2 >= SandMin2) AND (@2 <= SandMax2) ) 32 ELSE 0 ;

# Sparse seagrass box-classifier

IF ( (@1 >= SpSeagMin1) AND (@1 <= SpSeagMax1) AND (@2 >= SpSeagMin2) AND (@2 <= SpSeagMax2) ) 16 ELSE 0 ;

# Gorgonian plain box-classifier

IF ( (@1 >= GorgMin1) AND (@1 <=GorgMax1) AND (@2 >= GorgMin2) AND (@2 <= GorgMax2) ) 8 ELSE 0 ;

# Lobophora dominated algal areas and coral patch reef box-classifier

IF ( (@1 >= LobCoralMin1) AND (@1 <= LobCoralMax1) AND (@2 >= LobCoralMin2) AND (@2 <= LobCoralMax2) ) 4 ELSE 0 ;

# Dense seagrass box-classifier

IF ( (@1 >= DenSeagMin1) AND (@1 <= DenSeagMax1) AND (@2 >= DenSeagMin2) AND (@2 <= DenSeagMax2) ) 2 ELSE 0 ;

# Dense Montastraea reef box-classifier (sets this class to value of 1)

IF ( (@1 >= MontMin1) AND (@1 <= MontMax1) AND (@2 >= MontMin2) AND (@2 <= MontMax2) ) 1 ELSE 0 ;
 

Back to main page

Back to Module 7

USDOC | NOAA | NESDIS | CoastWatch