On Multi-Granular Labeled Classification For Spatial Remote Sensing Data

Rough set theory, proposed by Pawlak in the early1980s, can be used to analyze andprocess imprecise, inconsistent and incomplete information for data analysis. The rough setdata analysis uses only internal knowledge, avoids the external parameters, and does notdepend on the previous model or assumptions such as the probability distribution of thedata method, the membership function of the fuzzy set theory. Its basic idea is to unravel anoptimal set of decision rules from an information system via an objective knowledgeinduction process which determines the features constituting the optimal rule set forclassification. The one of the core issue of access to knowledge based on rough set theory isattribute reduction. Knowledge gained from the knowledge systems only after reducing hasthe value of promotion and application. Because of the unique advantages of rough setmethod of attribute reduction and rule extraction, the method has been successfully used ingeographic information systems and spatial remote sensing data analysis.In the problem of remote sensing image classification, the display of surface (e.g.,vegetation) data is the gray value of the different spectral bands reflected. The type of dataclassification task is to select a small set of spectral bands makes the same classificationability with the original spectral bands set. It should be noted that the gray value of thesurface reflected by the spectral band is taken a real number in the interval [0,255].Therefore, the original data (training set) is a real value information system, but if we seethe real value information system as training samples, we will get too many rules, andthe generalization ability is poor. So, we need have an appropriate transform or discreteprocessing to the real value information system. Leung etc. proposed the data analysis andprocessing model combining rough sets and statistics. This method transforms a real valueinformation system into an interval-valued information system via statistical method for theprocessing of remote sensing image classification. And it is based on the followingassumption: the gray value of the surface reflected by each spectra band is a normaldistribution. So we can transform the real value information of each type of surface into aninterval value μ±2σ(μ is mean, σ is variance), thus we transform the originalsample data set into an interval-valued information system. In this dissertation, we willpropose a rough set model to solve the problem of remote sensing data classification ofspatial surface.Granular Computing, sees the granular as the basic unit and simulates the naturalmodel of human thinking is used to handle large and complex data sets and information to establish an effective computational model. Although many granular computing modelshave been proposed, the information multi-granular computing of same object andmulti-granular problem solving is still lack of specific model. Recently, Wu and Leungproposed the theory and applications of granular labelled partitions in multi-scale decisiontables.This dissertation is mainly to study the problem of multi-granular labelled spatialremote sensing data. According to the different physical meaning of the surface material,we first get a multi-scale decision table (interval-valued information system). We thenpropose a rough set approach to find the “if-then” classification rules hidden in themulti-granular decision table through knowledge reduction, and we further calculate theclassification accuracy for the rule set. For the class which we can not classify without error,we define a misclassification rate, and then find the “if-then” classification rules within themisclassification rate and calculate the classification accuracy. Finally, we can obtainconclusion via our calculation.