One of the key issues of land use/cover change detection using remote sensing images is the threshold determination.This paper introduces the histogram approximation method based on Expectation Maximization(EM) algorithm and Bayes Information Criterion(BIC) into unsupervised change detection.EM algorithm is an iterative algorithm that has many advantages in estimating the statistical values.BIC is always used for evaluating a statistical model on the aspects of accuracy and complexity.Difference is acquired by applying the change vector analysis(CVA) technique,which can magnify the difference between the two-temporal images.The probability distribution function(PDF) of its histogram can be modeled as a mixture of M Gaussian distributions.Different values of M will get different models.The best one will make the value of BIC minimum.According to this criterion,the statistical values of the mixture Gaussian distributions can be estimated using the EM algorithm and BIC.Then the threshold of the change detection will be obtained by finding the intersection point of two neighbor Gaussian distributions.M Gaussian distributions will gain M points that are the M thresholds.M is regarded as the number of the changed types.The estimated values including means and variations and the prior distributions of every Gaussian distribution have definite physical meanings.The means indicate the values of images of those changed types based on which difference image can be classified quickly.The variations illustrate the difference in one changed types,and the percentage of every change types can be given by the prior distributions of every Gaussian distribution.The traditional methods of change detection by remote sensing based on EM algorithm is often assuming the difference image containing two types of pixels.Which are changed pixels and unchanged pixels.But when there are more than one changed types and the difference images' histogram becomes complex,this method is proved not accurate.To compare these two methods and validate this method,this paper chose the area around the Miyun Reservoir as the experiment area.There are more than one types changed including water,bare land and vegetation and so on,so this area is representative for change detection study.The experiment data is 2001 TM image and 2004 ETM+ image.The difference image's histogram is modeled by 4 Gaussian distribution,according to the models the difference image is classified 4 types.Then the difference image is processed by traditional method of change detection based on EM algorithm.The entropy is introduced to evaluate the two experimental results,which is usually used to evaluate the uncertainty of one pixel belonging to one classification.Its advantage is that it can make the pixels' uncertainty visible in the image.Results show that the histogram approximation based on EM and BIC method is credible and effective in change detection from the remote sensing images,especially when the changed types are more complex.

It is usually assumed that the prior distributions of parameters and error are Gaussian distribution in remote sensing inversion.This assumption seems to be impractical in many cases.Prior distribution of parameters and error are very important in remote sensing inversion since many remote sensing inversion strategies take advantage of prior knowledge.We present a bootstrap method for estimating the prior distributions of parameters and error in this paper.This method relaxes the distribution assumption of parameters and error,and obtains those approximately exact distributions by means of prior data.Moreover,we classify prior data since they are collected from different classes,and implement statistical test for classified prior data.Results show that proper classification of prior data is reasonable.Finally,we take RossThick-LiTransit linear kernel-driven model as an example,and make a comparison of our method with usual Tikhonov regularizing inversion and Bayes inversion under normal hypothesis with NOAA-AVHRR observations.The result shows that classifying prior data and using the prior distribution obtained by bootstrap method can significantly decrease uncertainty of parameters.