The classification algorithm is a supervised process. First the operator must construct typical examples for each type of sediment (sand, gravel, rock ...). Then SeaClass will compute the neural net that will separate each class. When classifying an entire image, small portions are extracted, parameters computed and fed into the neural net. The output of the computation will give the most probable class for each area in the image. If the texture in the image does not match any of the samples then it will not be classified and will be tagged as "Unknown".
This guide can be read from the start following each step or jump to a heading using these links:
.tmap_moz (Perspective Sidescan mosaic)
.DDS_VIF (Triton Map Visual Information File)
.TIF (GeoTIFF)
The procedure is the same for any of these file types, the following guide uses a GeoTIFF image.
Both the .tmap_moz and DDS_VIF files have full GIS information within the file. The story is slightly different for a GeoTIFF:
The type of GeoTIFF that is supported in Triton Perspective is an "Externally Referenced" GeoTIFF, meaning that the GIS information about the image is contained in the associated .TFW (World File), there is more information about the world file format here:
https://www.openjump.org/wiki/show/TFW+(World+File)+Format
Some GeoTIFF images are "Internally Referenced" meaning that the GIS information is written to the TIF header; however this type is not currently supported in Triton Perspective. There are free applications available on the web for creating TFW files from GeoTIF internal tags:
https://www.mentorsoftwareinc.com/freebie/free0699.htm
NOTE: Only Projected, Un-rotated GeoTIFFs are supported in Triton Perspective.
Triton Perspective is a multi-processor capable application, if the PC you are running has 2 or more cores processing will be faster if you inform Perspective of this, go to View>Settings Info...
And select the number of cores:
Start Triton Perspective and choose File>Import>GeoTIFF... (or Mosaic File.. for a Perspective .tmap_moz or TritonMap File.. for .DDS_VIF):
Choose the correct Projection: (the import of files with unusual or "Custom" projections and Datums is covered in a separate guide):
And UTM Zone:
And Datum:
The image is displayed in the main map view window of Triton Perspective:
Use the zoom tool to enlarge a portion of the image where we may have grab samples or ground truth for the bottom types we wish to classify elsewhere in the image.
This is an example of a sidescan mosaic generated from Triton Map™ and exported as a GeoTIFF image. For this document we will consider only two types of seabed, sand and mud.
Click the Classification icon in the Perspective tool bar to open the Select/Create Training Set dialog:
The default name for this training set is "Training Set 1" - you may change that name if needed.
Enter the Sample Size you need, (Note that choosing small values on a large image can take a very long time to generate the classification image, but see the section on Quick Classify).
Hit Next to open the Edit Classes and Samples dialog:
Click Add Class and enter a name for the first seabed type you want to classify (sand) then click OK.
You will then be prompted to enter a color for the class:
Hit OK, from this point the Wizard leads you through the steps:
"Click on the Record button to begin recording samples".
"Double-Click in the Perspective Map View on areas you want to use as samples for Training the Neural Net, Click on the Pause button when you are done".
Double click on the imagery in the main map view window, a box of the chosen color for this seabed type will appear at the location of the sample.
In this case we selected 4 samples for sand, and then clicked the Pause button which allows us to click Add Class again enter the new class (mud):
We chose the red color for the mud samples:
When all samples have been chosen click Finish:
Confirm that you want to Quit the Neural Net Creation:
The main map view window now displays all chosen samples in their associated color:
In the tree view, Right-Click on the node "Training Set 1" and select Train Neural Net...
In the Train Neural Net dialog click Create:
And enter a name for this Neural Net (demo).
If there were existing Neural Nets you would have the opportunity to choose one from the list, click Finish.
The following message may be displayed:
Click OK.
A new file with the extension .CNX is created in the working folder:
See Quick Classify at this point if desired.
The .CNX file is now been associated with the image (TIFF, DDS_VIF or tmap_moz), the next task is to create another layer that represents the results of that classification.
In the tree view Right-Click on the Image layer node to be classified and choose Classify:
In this case seaclass.TIF is the name of the image file, and the neural net to use is demo.cnx, Click OK.
In the dialog enter the name of the classification file, note that this is another Perspective image layer with a .tmap_cla extension.
Click Save.
Depending on the size of the file the generation of this layer may take some time:
Note the progress bar in this image showing that two cores are available (See View>Settings above).
A new classification layer is created:
Hit View>Zoom Home to see the whole image:
At this point it would be a very good idea to save the project, hit File>Save Project As and create a project.
The classification layer may be exported in two ways, as a .DXF file (R13) and as a GeoTIFF (.TIF plus .TFW)
To export a DXF file Right-Click on the Layer Name under the Classification node and chose Export:
Here is a view of the DXF file in Adobe® Illustrator:
Opacity - note that by default the classification layer is displayed in Perspective with 50% opacity, you can change this setting by Right-Clicking on the image layer and choosing Color Settings:
Use the Opacity slider to adjust the opacity of the layer allowing the underlying sidescan image to be seen "through" the Classification layer:
Before exporting the image you should be aware of this setting, if you need a dense image make sure the slider is all the way to the right (non-transparent), otherwise the generated TIF will be "dimmed":
Click on File>Export; the only option is "Composite GeoTIFF" which means that everything currently visible in the map view window will be exported:
If you need to export only the Classification layer (or any other layer) you need to uncheck the layers that are not needed, for example the Annotation, the Navigation (if any) and the sidescan mosaic:
The GeoTIFF Exporter window gives a preview of what will be in the final image:
Use the slider to "chop" the image if needed into 1, 4, 8, 16, 32 etc smaller TIFs (each has a .TFW) this can help with handling large high resolution images, you can also downsample the image using the pixels/meter setting:
Hit Export to open the Save Geo TIFF dialog and give the image a name
Note that if the slider was used to generate a number of images then each is automatically named along with the corresponding .TFW file:
As mentioned earlier the creation of the Classification image can take some time, it is worth noting that the actual classification of the image does not take very long. There is a Tool Bar Button called Quick Classification ID that can be used to check if the Classification image will be what was expected BEFORE the actual image is created. To activate the Quick Classify button you must complete all steps up to the end of Training the Neural net, once the .CNX file has been generated left clicking on the individual Neural Net in the tree view will activate the button. Modality>Quick Classification ID will also activate the button:
When the Quick Classify button is activated the mouse cursor becomes a Query pointer; double clicking anywhere on the image will briefly (3 secs) show the current classification for that point in the image:
It is quite likely that as more work is done in a survey area that more classes will need to be defined, use the Edit Training Set option to add more classes to a project:.
Right-Click on the node Training Set 1 under Manual Training and select Edit Training Set...
Click Add Class to add a new class name and color then click as before to define the new class:
The same .CNX file can be updated or a new one created, the classification can then be done with either of the available .CNX files.
