Geodeziya, kartografiya, geografiya

Artificial Neural net classification
Artificial neural network algorithms attempt to mimic the human learning process to associate the correct land cover class label to image pixels. Neural networks have their roots in the field of artificial intelligence and although the algorithm still pales in comparison with the human brain its use in image classification has been quite successful. One advantage to using neural networks is that it is easy to incorporate ancillary data in the classification process. This in itself can greatly improve classification accuracy.
The training process for a neural network classification can be time consuming and is not as simple as the supervised statistical approach. There are a number of parameters that are not very intuitive and experience plays an important part in learning how to use these methods effectively.
Binary decision tree classification
Decision trees are a common machine learning tool that have taken hold in the remote sensing arena. Decision trees are a set of binary rules that define how specific land cover classes should be assigned to individual pixels. (Figure 6) illustrates a simple decision tree that was used to classify forest and non-forest classes using two Landsat TM bands. At each node a true/false decision is made thereby creating a branch in the tree. The terminal node (the bottom tips of the tree) along each path defines an individual land cover class. This approach makes it easy to integrate ancillary data into the classification process.
Creating the rules for a decision tree requires decision tree generator software available in most popular statistics packages. This software takes sample data from training areas in the satellite imagery and ancillary data and works to define the most effective set of rules to define the nodes in the decision tree. Experience is helpful when working with decision trees as they require a good deal of editing (pruning) to produce suitable land cover maps.
Image segmentation
Image segmentation is not a classification tool; rather, it is a method to group contiguous pixels into areas (segments) that are relatively homogeneous. It can be thought of as a preprocessing step before classifying an image. The segmentation is done using sophisticated algorithms that compare a pixel's value with the values of the neighboring pixels. If neighboring pixels are relatively similar then they are added to the contiguous group and if they are not then another segment is started. These programs provide variables that allow an analyst to specify the relative size and sometimes even the shape of the segments.
Once an image has been segmented it can be classified at the segment level instead of the pixel level. There are several advantages to this approach:
• it runs much faster since the number of segments is much less than the number of pixels in an image,
• the relative scale of the segmentation output can be specified so different segmentation runs can be used to capture features of different sizes,

• the classification algorithm can utilize the spectral characteristics (the pixel values) of an image as well as a host of other segment characteristics that describe the segment, such as mean pixel value, standard deviation, shape of the segments, and dimensions of the segment, and

• the resulting image does not suffer from the "salt and pepper" effect common to the pixel-by-pixel classifiers mentioned above.

This approach holds a lot of promise for imagery with high spatial resolution although it is becoming quite popular for classifying moderate resolution imagery as well. With moderate resolution imagery the accuracy obtained using this approach is often similar to other methods but the advantages mentioned above make it an appealing alternative. As with the other newer methods, this classification approach requires experience to become proficient.
Manual
Manual, or visual, classification of remotely sensed data is an effective method of classifying land cover especially when the analyst is familiar with the area being classified. This method uses skills that were originally developed for interpreting aerial photographs. It relies on the interpreter to employ visual cues such as tone, texture, shape, pattern, and relationship to other objects to identify the different land cover classes. The primary advantage of manual interpretation is its utilization of the brain to identify features in the image and relate them to features on the ground. The brain can still beat the computer in accurately identifying image features. Another advantage is that manual classification can be done without a computer, instead using a hardcopy version of a satellite image.
The downside of manual interpretation is that it tends to be tedious and slow when compared with automated classification and because it relies solely on a human interpreter it is more subjective. Another drawback is that it is only able to incorporate 3 bands of data from a satellite image since the interpretation is usually done using a color image comprised of red, green, and blue bands. More information about using manual interpretation methods is available in the Justification for using photo interpretation methods to interpret satellite imagery guide.
The technique used in manual interpretation is fairly simple. The analyst views the image on either a computer screen or a hardcopy printout and then draws a polygon around areas that are identified as a particular land cover type (Figure 8). If the land cover delineations are done on a computer screen the land cover map is created during the delineation process. If the interpretation is done on a hardcopy image the resulting map will have to be digitized to convert it into a machine readable format.
Hybrid
A hybrid approach combines the advantages of the automated and manual methods to produce a land cover map that is better than if just a single method was used. One hybrid approach is to use one of the automated classification methods to do an initial classification and then use manual methods to refine the classification and correct obvious errors. With this approach you can get a reasonably good classification quickly with the automated approach and then use manual methods to refine the classes that did not get labeled correctly.
The editing process requires that the analyst be able to compare the classified map with either the original satellite image or some other imagery that can be used to identify land cover features. To compare a classified map with imagery it is helpful to have access to software that allows the analyst to flicker between two images (the land cover image and the original satellite image) or slide one image over the other on the computer display using a technique often called "swiping". By doing these comparisons the analyst gets a sense of the quality of the classification. When errors are spotted they can be edited using tools common in many image processing software packages.
It is surprising that many remote sensing practitioners will not edit a classified map even if it is obvious that a certain area is misclassified. If the purpose of the land cover classification study is to produce the best map possible then the analyst should use all means possible to meet that goal. In most cases, visually editing a classified map will improve the accuracy of the final product.