Classify Multi-Spectral Data Classify Geologic Terrains on Venus Apply Multi-Variate Statistics Operations What Do I Need? Classify Merge Combine Cross Scan Score Warp Respace Cover Subscene Rotate Translators Recode To classify multi-spectral data, you need a minimum of one and up to five map layers with the same coverage, cell resolution, and orientation. Also, you will require a training map which provides information to the Classify operation on the location of the various class types that you are seeking in the multi-spectral image. Refer to the Example section below for a full explanation of the training map layer. If one or more of your multi-spectral or multivariate map layers has cells with the value VOID, VOID will be assigned to those cells in the output map layer. Do not allow your training map class polygons to overlap cells in the multi-spectral or multivariate map layers that have the value VOID. Example Multi-spectral satellite imagery can be used to derive thematic map layers, perform change detection, and prepare maps for base mapping, planning, land use, resource management, geology, hydrology, geomorphology, and many other applications. Two fundamental questions that GIS are designed to answer are: "Where is what?" and "What is where?"; that is to say: " Where on my map layer does HD-CLS-1
a particular phenomenon occur?" and "What phenomenon is found at a given location on my map layer?" Multi-spectral imagery allows you to answer these kinds of questions about the surface of the Earth or other planetary bodies. Remote sensing image analysts have found that similar reflectance and emission values for each sensor band are produced by similar landcover types and that signatures can be derived for each landcover type. For example, an alfalfa field reflects a certain range of infrared light, red light, and green light. Other landcover types may have the same infrared range of reflectance values but different ranges for red and green light. By applying statistical analysis techniques to the data values in a set of multi-spectral images, it is possible to assign a unique class identifier to particular combinations of value ranges. HD-CLS-2
These LandSat Thematic Mapper images of the Mississippi were obtained from the NASA website (www.nasa.gov) to demonstrate land cover classification. LandSat images are available in seven spectral bands: Band Spectrum Micrometres Resolution 1 Blue 0.45-0.53 30 m 2 Green 0.52-0.60 30 m HD-CLS-3
Band Spectrum Micrometres Resolution 3 Red 0.63-0.69 30 m 4 NIR 0.76-0.90 30 m 5 MIR 1.55-1.75 30 m 6 Thermal 10.40-12.50 120 m 7 MIR 2.08-2.35 30 m Bands 2 (Green), 3 (Red), and 4 (Near Infrared - NIR) will be used to classify the land cover for this area. Satellite images are captured by exposing a sensor to the reflected and emitted wavelengths of energy from the surface of the Earth or other planetary body. The energy is passed through filters that isolate certain energy wavelengths. The energy intensity is quantized (i.e., quantified and grouped) and stored as digital values ranging from zero (no energy) to 255 (maximum measurable intensity). Digital data of different wavelengths are made available by satellite data distributors as sets of multi-spectral image maps for a common area. Each image has the reflectance or emission values stored on separate image layers, each registered to a common grid of cells or pixels. These separate layers can be imported into MFworks and viewed individually or combined and analyzed using MFworks operations such as Classify, Merge, Combine, Cross, Filter, Scan, and Score operations, or the mathematical and Boolean operators, or the functions. Multi-spectral image classification from satellite imagery has been most successful when applied to agricultural and natural vegetation land cover mapping. The majority of multi-spectral satellite imaging platforms were designed to maximize the discrimination of these classes of land cover. Other types of surfaces, such as rock and ice, are much more difficult to discriminate and classify from multi-spectral satellite imagery. Rock outcrops are rarely homogeneous or extensive. Commonly they are covered by debris, soil, and/or vegetation. Radar imagery, such as that provided by RADARSAT, are better suited for geologic applications. You can apply the MFworks Classify operation to RADARSAT imagery. The LandSat 5 TM (Thematic Mapper) images the surface of the Earth continuously in seven energy wavelength bands from blue to the far (or thermal) infrared. (Source: Lillesand & Kieffer, 1987, Remote Sensing and Image Interpretation, John Wiley & Sons.) HD-CLS-4
Band 1 (blue) is useful for coastal water body mapping as well as soil, vegetation, forest type, and cultural features. Band 2 (green) is designed to measure the green reflectance peak of vegetation. Band 3 (red) is designed to sense chlorophyll absorption which is useful for plant species differentiation and cultural feature identification. Band 4 (near infrared) is designed to allow you to determine vegetation type, vigor, and biomass, as well as delineating water. Band 5 (mid infrared) is useful for detecting vegetation moisture content and for distinguishing snow from clouds. Band 6 (far infrared/thermal) can be used to analyse vegetation stress and thermal mapping. Band 7 (mid infrared) is useful for discriminating rock and mineral type. Multi-spectral imagery can be classified in two ways: unsupervised classification and supervised classification. Unsupervised Classification Unsupervised classification applies statistical analysis to the multi-spectral image layers and assigns a class to each cell automatically, based on value combination patterns in the data and the probability of class membership. You then have to examine and "ground truth" the classification to assign meaning to the classes. For example, class "1" might be bare ground, class "2" might be "paved surfaces", class "3" might be conifers, and so on. Unsupervised classification answers the question "what is where?" At this time, the MFworks Classify operation does not perform an unsupervised classification. Supervised Classification Supervised classification is similar to unsupervised classification in that multi-variate statistical analysis is applied to a set of multi-spectral images, or other multi-variate datasets, and a class is assigned to each cell based on the statistical probability that combinations of value ranges represent a single phenomenon. However, instead of returning a set of values to which you apply class names, you provide the Classify operation with the classes and expected value ranges for each class, and the Classify operation will assign these classes to the value range combinations that it recognizes in the input map layers. Supervised classification answers the question "where is what?" HD-CLS-5
To perform a supervised classification you must decide or determine what classes of land cover exist in your area of interest. This can be done by consulting airphotos, existing maps, high resolution imagery, field surveys, or by visiting the area and recording the classes by location. For this last method, use a GPS (Global Positioning System) unit or topographic map to record the locations of the various land surface classes. The Training Map Layer Once you have decided what classes you are looking for and have located some of those classes within the study area, you can make the training map. The training map should have the same coverage, cell resolution, origin, and geometry as the multi-spectral data layers. Creating the Training Map Layer Creating the training map layer is the most difficult and time consuming part of supervised image classification. To develop an adequate training map layer, you must have detailed knowledge of the area that you want to classify. The better your training map layer, the more successful your classification will be. The process of creating the training map layer involves careful marking of cells that represent each class on the training map layer. Your training areas must be representative of the actual classes in the data and of the range of values within each class. You may need more than one training area to completely define the spectral characteristics of a given class. For example, in the training sites for a category called "ice", you may want to delineate cells that contain glacial ice, snow, frozen lakes, ice rafts, and pack ice. There are many ways that you can create a training map layer. You can begin with a set of airphotos, a topographic map, a DEM, a panchromatic satellite image, a file of XYZ coordinates and values, a copy of one of your multi-spectral images, or, as in the example below, a false colour composite and a true colour composite made from the multi-spectral dataset. The MFworks Merge operation was used to make these composite images of the Mississippi study area from the multi-spectral image dataset. The HD-CLS-6
false colour composite map layer will be used as the training map layer for the Classify operation. The true colour composite will be used to visually inspect any areas in the false colour composite that require clarification. The MFworks Merge operation was also used to create the true colour map layer. HD-CLS-7
Use the drawing tools to carefully mark representative areas for each class in the study area. Choose a unique value for each class and draw on top of the cells containing the class. The value range of the map layer named FalseColour is 0 to 32767; values outside this range were specified for the training sites. For each class polygon that you mark, ensure that you do not include cells that do not belong to the class and that you do not include cells right at class boundaries. Boundary cells may contain mixed information from the adjacent classes due to the low resolution of the multi-spectral images. Cells with mixed information are sometimes referred to as "mixels". HD-CLS-8
If mixels are included within the definition of the class, then the statistical algorithm may misclassify cells from the adjacent class. Your choice of training cells should be broad enough to cover the variance and covariance of each class. As the number of training cells increases for each class, the estimates of the mean for each class improves. It is better to have many locations containing a few cells for a given class than to have one location containing many cells. Choose locations that are well dispersed across the study area whenever possible. Refer to Remote Sensing and Image Interpretation by Thomas Lillesand and Ralph Kiefer (John Wiley & Sons, 1987) for an excellent review of the training stage and training set refinement. Other Sources for the Training Map Layer Airphotos: scan and import hard copy originals or import digital files. If your airphotos are not orthorectified, you can use the Warp operation to georectify them to a topographic map or other reference map layer. Next, use the Cover operation to mosaic the airphotos. Use the Respace operation to rescale the airphoto mosaic map layer to the same cell resolution as your multi-spectral image map layers. Finally, use the Subscene or Cover Mosaic operation to crop or expand the boundaries of the airphoto mosaic to match the coverage of the multi-spectral image map layers. It may also be necessary to use the Rotate operation to adjust the azimuth of your training map layer. Panchromatic Satellite Images: a similar technique as above can be applied to panchromatic imagery. Often a satellite dataset will include a panchromatic image of the multi-spectral coverage. Panchromatic images are pre-orthorectified saving you the step of georectification using the Warp operation. As well, a panchromatic image is supplied in digital form, preregistered to match the multi-spectral dataset. However, panchromatic imagery is significantly more expensive than standard orthorectifed and unrectified airphotos. Multi-Spectral Image: to save the cost of a panchromatic image, you can use the Merge operation, as in the example above, to create a colour composite from your multi-spectral dataset. Alternatively, you can create a copy of one of the images from your multi-spectral dataset and use that to define the locations of your class training sites. Topographic Maps and DEMs: scan and import a paper topographic map or download and import a digital topographic map or DEM of your study area. If necessary, use the Cover operation to mosaic your topographic map scans or your DEMs. Use the Respace operation to rescale your topographic map layer or DEM to the same cell resolution as your multi- HD-CLS-9
spectral image map layers. It may also be necessary to use the Rotate operation to adjust the azimuth of your training map layer. Use the Subscene or Cover Mosaic operation to crop or expand the boundaries of the topographic map layer or DEM to match the coverage of the multispectral image map layers. XYZ Coordinate File: An XYZ file contains a list of coordinate points and an associated data value for each cell. The coordinates can be row/column, XY, Latitude & Longitude, or UTM metres. The Z value for each coordinate pair should be the value that you have assigned to represent the associated class. GPS (Global Positioning System) units are commonly used to collect spatial data and generate XYZ type files. Import the XYZ file, entering the coordinate range and cell resolution of the multi-spectral image map layers in the appropriate fields of the Import XYZ File dialog box. The resulting file can be used as your training file directly. Extract training sites Once you have marked training sites for each class on your training site base map, use the Recode operation to create the training map layer by extracting the training sites that you designated. In this case, all cells but the training areas are assigned the value VOID: If you were to perform this operation from the Script window the statement would be: TrainingMap = Recode "False Colour + Training Sites" HD-CLS-10
Assigning VOID to 0É32767 CarryOver BuildText; The training map layer contains only the polygons that were assigned as training areas. The non-training areas are assigned the value VOID. The colours that you assign to the class zones will be the colours that get assigned to the zones in the final classified map layer. A simple brown to sand colour sequence is applied to the training sites: You are now ready to classify the land cover in the multi-spectral dataset. Use the Classify operation to perform a maximum likelihood classification HD-CLS-11
of the multi-spectral images based on the classes that you specified in training map layer: If you were to perform this operation from the Script window the statement would be: "LandCover" = Classify Green With Red With NIR UsingÊ"TrainingMap"; The resulting map layer named LandCover is comprised of cells that have been assigned to one of the classes specified in the training map layer based HD-CLS-12
on the maximum likelihood that the cell is a member of one of the given classes. HD-CLS-13
Each class can be assigned an appropriate colour to create the final thematic map layer. HD-CLS-14