Exploration Of The Influence Model Remote Sensing Image Uncertainty On Classification

Uncertainty as an inherent property of the objective world and the entity itself,can reflect the error between the acquired data and the real data.In recent years,with the wide application of remote sensing data in land planning,disaster monitoring,target detection and other fields,remote sensing data has become the main information source to deal with geographic information problems.The uncertainty of remote sensing data is restricting the further development of productization and application of remote sensing information.As one of the main application fields of remote sensing data,the uncertainty of remote sensing data exists in the whole process of classification,including data acquisition,transmission,processing and information extraction,precision evaluation,and continuous dissemination and accumulation.Therefore,to understand the sources and nature of uncertainty in remote sensing image data and explore the impact mode of image uncertainty on the classification results of land cover are of great significance for proposing effective schemes to avoid and reduce image uncertainty and developing reliability classification methods.This article analyses the characteristics and formation mechanism of random uncertainty and fuzzy uncertainty in remote sensing images based on information entropy model and fuzzy entropy model,additionally using evidence combing theory to improve these two models,and establishes evaluation index model in pixel scale to describe these two kinds of uncertainty..Then by combining with the reliability evaluation index and classification accuracy evaluation index,this article construct the index system of reliability evaluation of classification results,including the overall classification accuracy,consistency,integrity and reliability.On this basis,this article selects four groups of remote sensing data with different resolution including BJ2,GF1,Landsat-8 and Worldview 2 as data source.The experiment use SVM classifier,minimum distance classifier,MPs classifier based on morphology to classify the image,and use regression analysis to explore how the random uncertainty and fuzzy uncertainty effects on the reliability of the image surface coverage classification model by using different images and different classification method.The experimental results show that:(1)Uncertainty evaluation index to a certain degree can reflect the uncertainty of image pixel size.The distribution of the random uncertainty and fuzzy uncertainty in the same image is different in categories and space.And pixels with higher random uncertainty and fuzzy uncertainty are more likely to be gathered in ground object boundary by linear distribution.(2)During the classification results of minimum distance classifier,SVM classifier,MPs classifier in different images,the rate of misclassified pixels which contains random uncertainty and fuzzy uncertainty is much higher than pixels with certainty.At the same time,the numeric value and mean value of random uncertainty and fuzzy uncertainty in misclassified pixels are higher than those in correctly classified pixels.(3)Random uncertainty and fuzzy uncertainty in images have a certain influence on the reliability of the results of different classification methods,more specifically,in the same image,the overall correctness and consistency have strong negative correlation with the two kinds of uncertainty,which means the greater uncertainty will be,the lower overall correctness and consistency of classification results will have,and specific downward trends are distinguished between different classification methods;However the random uncertainty and fuzzy uncertainty effect the classification integrity and reliability differently in categories and methods.During the results of classification,the classification integrity and reliability of cultivated land and forest land are both strongly effected by these two uncertainty,while the classification integrity and reliability of buildings are more relevant with fuzzy uncertainty than that with random uncertainty.