Improving classification accuracy using knowledge based approach

Alesheikh, Ali A.
Assistant Professor
Email: ali_alesheikh@hotmail.com

Fariba Sadeghi Naeeni Fard
Graduate Student
Dept. of Geodesy and Geomatics Eng.,K.N. Toosi University of Technology
Email: sadeghi@ncc.neda.net.ir
Tel: (+98 21) 8779473,Fax: (+98 21) 8779476

Ahmad Talebzadeh
Applications & GIS Director, Iranian Remote Sensing Centre


Abstract
Remotely sensed images are major sources of information, and as such, are used in many fields like morphology, geology, and agriculture. So far, many efforts have been performed to extract information from satellite images, and various methods have been developed in the field. Two main approaches are visual interpretation and quantitative analysis (digital interpretation). Among digital techniques, classification is a common and powerful information extraction, which is used in remote sensing. There are many classification methods that have their own advantages and drawbacks. Standard classification methods usually take pixels as fundamental elements and try to label the pixels based on their spectral properties. It is clear that using spectral properties alone, may not lead to adequate accuracy. Maximum Likelihood Classification (MLC) is perhaps the most widely used classification method. The underlying assumption on performing MLC is that the prior probability of land cover is equal, due to insufficient information. However, a-prior occurrence probability gives a crucial effect to classification results. As long as the class showing the highest likelihood is allocated to a pixel, misclassification errors are unavoidable. The objective of this paper is to improve the accuracy of Maximum Likelihood Classification method using a-priori information. Estimates of a-prior probability through, crop areas, crop calendar, soil type information and some a-prior probability about agricultural practices have been used in assigning the probability to pixels before classification. Methods of gathering information on a-prior probability, together with the results of the research have been presented and analyzed in this paper. An industrial agricultural field, Moghan Plain, in North Western Iran has been selected for testing the methods. Validation results demonstrate that this way is effective to improve classification errors.