Copyright © Panorama Group 1991 - 2021
Complex of automated decryption and vectoring the remote sensing data
Complex of automated decryption and vectoring the remote sensing data is intended for an automatic vectoring of linear and areal objects by panchromatic, color and multispectral images of the Earth surface.
The process of automatic vectorization consists of the following stages:
At the step of classification there is carried out the definition of belonging the pixels of an initial raster to classes of recognizable objects. To do this, the user specifies patterns - the area uniquely belonging to recognizable objects.
On patterns there is going a training of the classifier - calculation and memorization of interpretive features for each position of scan window within the pattern. As the interpretive features there are used the statistical (average color, standard deviation) and textural characteristics (contrast, energy, correlation). Statistics can be calculated both on an arbitrary spectral channel, and on a combination of channels (an average or NDVI index). The calculated features are stored into a n - dimensional weights array (n - number of features) of belonging a feature set to a class.
At classification by a feature set for each position of a window there is accumulated the weight of belonging all pixels of a raster to each of classes. As a result, pixels get a sign of a class with the maximum total weight.
Raster of classification contains noise - incorrectly classified pixels. For a filtration of noise there is carried out a postprocessing the raster of classification - deleting of areas by size, erosion, building of low-contrast areas, deleting of shades, smoothing the edges of areas. Steps of erosion and building of low-contrast areas allow to specify borders between areas. Deleting of areas is carried out by a joining of small areas to neighbours by the maximum border and by the priority set by the user. Deleting of shades carries out joining of shades to the specified classes by a shade's angle of incidence. Smoothing the edges of areas is performed for generalization of the areas form.
After processing, the classification raster will be transformed into a set of vector linear or areal objects.
At the final step the vector objects are united or deleted on the basis of the analysis of their relative positioning. The united network of objects together is smoothed and filtered before saving into a created map. Results of recognition can be exported into formats SIT, SXF, SHP for their subsequent editing by means of the GIS.
In addition to a topographical decoding the complex allows to carry out multispectral classification. For these purposes, the user can use statistical and textural features onto any raster channel or onto virtual channels, calculated according to the formula NDVI.