Utilizing geometric attributes of spatial information to improve digital image classification

Authors:
Publisher website
Google Scholar
Abstract
Multispectral classification techniques are the most commonly used methods for extracting thematic information from remotely sensed data. These techniques analyze the spectral reflectance properties recorded by the sensors using special algorithms that quantitatively assess the relative multi‐band brightness values (BVs). Unfortunately, due to the coarse band widths employed by the frequently used sensor systems (e.g., Landsat and SPOT), and the fact that various terrestrial objects may exhibit similar spectral responses, there is significant confusion in the land cover information extracted, which in turn reduces the classification accuracies of remote sensing‐derived image‐maps. This research focuses on incorporating the geometric attributes of detected ground objects to resolve some of the spectral confusions that occur when pure per‐pixel classification algorithms are used. The method identified homogeneous areas by using an edge detection algorithm and stratified the image into two strata based on the shape index of the patches. One stratum was composed of patches with regular shapes and large sizes, while the other included highly fragmented patches. By stratifying the image, the classes which frequently caused the spectral confusion, such as grassland and wet meadows, were assigned to different strata, and a spectral classification algorithm was separately applied to each stratum to derive their land cover information. The classifications were subsequently merged to produce a composite land cover map of the study area. Traditional unsupervised and supervised multispectral classification algorithms were also applied to the data in order to perform a comparative evaluation of the accuracies of each technique. The accuracy assessment revealed that by considering the geometric attributes of various terrestrial features, the digital image classification was significantly improved.

This publication has 1 reference indexed in Scilit: