Change Detection From Heterogeneous Remote Sensing Images With The Scarcity Of Prior Information

With the development of remote sensing technology,heterogeneous remote sensing data show unique complementarity and superiority in practical change detection applications.Direct comparison of heterogeneous data for change detection always has a poor detection accuracy.In addition,heterogeneous remote sensing data often contain a weak priori information,so how to use this under-exploited priori information for change detection is also a major challenge.In this paper,we research the heterogeneous remote sensing images change detection under the different conditions of priori information scarcity.The main research includes the following three aspects:(1)For the weak priori information contained in sequence heterogeneous remote sensing images is not fully utilized,we propose a semi-supervised label propagation(SSLP)approach.First,a clustering label propagation(CLP)is designed to cluster pre and post images,respectively.It assigns pseudo labels to unlabeled pixel pairs that have similar mapping relationships to the labeled pixel pairs.Second,a pixel density metric is investigated to filter out the data with low density,which can ensure the reliability of the propagated data.Finally,a secondary expansion method based on pixel neighborhood is used to generate enough training data for training a classifier.Validating the effectiveness of the proposed method on three real datasets,it can be seen that the accuracy of SSLP is higher than 97.74%in all datasets,and K_a is improved by at least about 7%.(2)For the inconsistent distribution of heterogeneous remote sensing image features due to different imaging mechanisms,a multioutput image regression and association-based feature fusion(IRAF)approach for heterogeneous change detection is proposed based on the unchange weak priori information available in heterogeneous remote sensing data.First,IRAF determines the adaptive regression direction based on the information entropy,which utilizes the difference of information between heterogeneous data.To transform heterogeneous data into a common feature space,the regression image will be obtained via a multioutput multilayer perceptron image regression model.Then,the fuzzy local information C-means algorithm is used to identify the fuzzy region in the difference image,which further ensures the reliability of significant sample pairs.Finally,an association-based fusion method is applied to the source dataset by simultaneously exploiting the high-order information of heterogeneous data and the association information between features.The binary change map is obtained via training a classification model with the boosting dataset.Experiments conducted on three real datasets(Sardinia,Yellow River and Texas)shows the effectiveness of the IRAF by comparing it with seven related methods,K_a was improved by about 10%,5%and 6%respectively.Experimental results shows that IRAF can suppress the effect of noise and effectively improve the change detection accuracy.(3)For the difficulty of establishing connections between heterogeneous data domains without prior information,we propose a self-supervised transformation(SST)change detection approach,which automatically mine some reliable changed/unchanged pixel pairs and make images transformation.First,an image pixel discretization(IPD)method is developed to strengthen the difference between ground objects,which can improve the quality of the superpixel segmentation.Second,a difference learning method based on spectral and textural features is designed to select some significantly changed/unchanged pixel pairs by quantifying the difference between superpixel pairs.Third,a multi-medium bidirectional pixel transformation(MBPT)method is introduced to transform the heterogeneous images to a common mapping feature space via the labeled changed/unchanged pixel pairs.Finally,an evidence fusion method is defined for expanding labeled data thereby training a classifier to achieve accurate detection results.The effectiveness of SST was verified on four heterogeneous datasets,accuracy was higher than 96.91%,and K_a was improved by at least about 2%.