following meaning (class and cluster are used as synonyms, explanations are based on the U.S. Army CERL, 1993 tutorial):

Minimum class size: minimum number of pixels to define a cluster;

Class separation: minimum separation below which clusters will be merged in the iteration process. It depends on the image data being reclas-sified and the number of final clusters that are statistically acceptable. Its determination requires experimentation, usual values range between 0.5 to 1.5. Note that as the minimum class separation is increased, the maximum number of iterations should also be increased to achieve this separation with a high percentage of convergence (see percent convergence);

Percent convergence: point at which cluster means become stable during the iteration process. When clusters are being created, their means constantly change as pixels are assigned to them and the means are recalculated to include the new pixel. After all clusters have been created, i.cluster begins iterations that change cluster means by maximizing the distances between them. As these mean shift, a progressively higher convergence is approached. Because means will never become totally static, a percent convergence and a maximum number of iterations are supplied to stop the iterative process. The percent convergence should be reached before the maximum number of iterations. If the maximum number of iterations is reached, it is probable that the desired percent convergence was not reached. The number of iterations is reported in the cluster statistics in the report file;

Maximum number of iterations: determines the maximum number of iterations which is greater than the number of iterations predicted to achieve the optimum percent convergence of the Chi-square test. If the number of iterations reaches the maximum designated by the user; the user may want to rerun i.cluster with a higher number of iterations;

Sampling intervals: simplifies the calculations by grouping the pixels into blocks. If the system resources are about to be depleted due to a too small block size, i.cluster will send an email containing a warning. These numbers are optional with default values based on the size of the data set such that the total pixels to be processed is approximately 10,000 (consider round up). With appropriate hardware, the unrecommended sampling may become unnecessary.

After starting i.cluster, first enter the group and subgroup names. Then, provide a filename for the Result signature file (which will store the cluster information for i.maxlik. Only if present (not in the first run), a filename for a Seed signature may be specified. This allows you to use cluster information from a previous run or to use spectral signatures from another partial

supervised reclassifications using i.class. Then enter a filename for the Report file which will be written to the current directory. It contains statistical information about the clustering process. If desired, the module can Run in background .

The following screen allows us to modify the parameters as described above. For a LANDSAT-TM5/7 scene, the Number of initial classes should be initially set to 20. This is a test case - the number has to be changed depending on the results, especially when the convergence is not reached. The other default parameters may be accepted for now. To continue, enter <ESC><ENTER>. Now the cluster analysis is running, generating the cluster statistics and the report file. After checking the quality of the clustering process in the report file, an eventual modification of the parameters and one or more new runs of i.cluster are required.

Second step: Unsupervised reclassification of image data. Finally, the unsupervised classification based on the MLC algorithm can be started with i.maxlik. The module will assign all pixels in the satellite image to the spectral signatures (classes) derived by the previous clustering process. After starting i.maxlik, the image group has to be selected. Then the Result signature map which is the result of the clustering process performed with i.cluster is queried. Second, a name for the Classified map layer (the new reclassified image) which will be created by i.maxlik has to be specified. Finally, a name for the Reject threshold map is needed to store the pixel assignment confidence levels. As described above this map represents the spatially localized errors which occurred when assigning each pixel to a class. After specifying all parameters, GRASS will compute the unsupervised reclassification. These maps can be displayed now with d.rast. The reject threshold map contains one calculated confidence level for each classified cell in the classified image. In case the quality of the reclassification process is not acceptable, the number of classes or other parameters need to be changed subsequently, and the clustering and MLC analysis must be repeated with the new values.

The classes in the reclassification map are then manually assigned to the appropriate land use types in the verification. The assignment of categories can be done with r.support ( Edit categories ) or r.reclass. To change the map colors to more intuitive ones (water colored blue, etc.), the module r.colors can be used. A command line based example for reclassifying the SPOT-1 HRV/PAN data into 5 land use classes is as follows:

i.group group=spotmss sub=spotmss\

in = spot.ms.1, spot.ms.2,spot.ms.3,spot.p

#clustering:

i.cluster group=spotmss sub=spotmss classes=5 sigfile=cluster #MLC:

i.maxlik group=spotmss sub=spotmss sig=cluster\ class=mlc.unsup rej=mlc.unsup.re j d.rast.leg mlc.unsup d.rast mlc.unsup.rej

tselect all areas with confidence level >= 90% #of correct assignment: r.report mlc.unsup.rej un=h

r.mapcalc mlc . unsup. qual = if (mlc . unsup. re j >= 12, 1, nullO) r.report mlc.unsup.qual un=h d.erase

d.histogram mlc.unsup.rej

The filtered rejection map mlc.unsup.qual can be used as MASK to select the pixels with a high confidence level of assignment.

9.8.2 Supervised radiometric reclassication

In a supervised reclassification, the classification process is supported by an interactive selection of known areas (for the general workflow see Figure 9.16). Using visual inspection in the field or auxiliary training maps, areas with known land cover are selected and stored in a training map, which is used to identify the spectral signatures for the reclassification process. These known areas are also called ground truth areas . It is important that the training areas are homogeneous samples. Since training areas cover several pixels, small local variations are included for the definition of the classes. For verification, the module i.class supports analyis of channel-wise histograms. A Gaussian distribution of the spectral responses is assumed and standard deviations are displayed in the histograms. These standard deviations can be modified to change the cluster statistics. The spectral signatures (grouped later into classes) are computed from the regional mean values of the training areas and their co-variance matrices.

The training areas can either be digitized within the module i.class (covered in the first part of the following description) or prepared from auxiliary maps such as already available land use maps (second part of the following description).

Interactive selection of training areas. The manual vectorization of training areas is accomplished with i.class. First the satellite channels have to be joined into an image group and subgroup using i.group. Creating a natural or false color composite (see Section 9.7.2) which will be helpful for identification of training areas is recommended.