Object-based Change Detection With Optical Satellite Imagery Based On Multi-features Information Mining

With the continuous development of remote-sensing satellite platform and sensor technology,the earth observation technology and means have become increasingly mature.The system has the capability to offer huge amount of multi-source and heterogeneous remote-sensing imagery.Therefore,how to extract useful information from remote-sensing data and avoid the phenomena of "information islands" are of great significance for the discovery of image potential and the transformation from "data" to "information".High-resolution remote-sensing imagery(HSRI)is rich in geometry and structure information,which greatly facilitates the interpretation,classification,and change analysis of satellite imagery.Change detection(CD)technology is capable of discovering the change process of surface features,providing decision-making information for resource investigation,environmental protection,urban planning and disaster assessment.Among the CD techniques,object-based change detection(OBCD)methods have become the main trends of CD in HSRI for its capabilities of making full use of spatial context information and avoiding the interference of "salt&pepper" noise.However,due to the decline of spectral separability and the aggravated phenomena of "same spectrum from different objects" and "same objects with different spectrum",OBCD remains a challenging issue,which mainly lies in the following aspects:firstly,multi-temporal image segmentation(MTIS)is the foundation and basis of OBCD,and the accuracy of OBCD is seriously restricted by MTIS;secondly,the usage of scale information is mostly implemented by scale fusion,which neglects the consideration of scale sets constraints;thirdly,objects in the MTIS result are treated independently,ignoring the effect of neighboring objects interaction on the change property.Therefore,the intensive studies on MTIS,scale sets constraints and neighborhood information constraints play an important theoretical role in improving OBCD accuracy.Based on the above key problems,three corresponding OBCD methods are studied and presented to improve OBCD accuracy from the viewpoints of OBCD process,namely MTIS,multi-features extraction,multi-scales fusion and post-processing.The main work of this article can be concluded as follows:1)Research on CD methods based on segmentation optimization and multi-features fusion.Object change is ignored in existing MTIS and the object sizes are distributed uniformly,which affect the accurate extraction of object boundary and the statistical stability of feature extraction.To this end,the method of segmentation optimization based on object intersection and merging is presented,which on the one hand establishs the spatial relationship between objects from two periods of images by the intersection process,and on the other hand produces image objects of reasonable size and uniform distribution by the merging process utilizing spectral and shape features.In order to make full use of rich detail information in HSRI,a multi-features fusion method based on spectral,textual and spatial features is presented,which effectively compensates the insufficient detection performance using single feature and achieves a preferable balance in terms of accuracy and efficiency for its flexible parameter setting and high degree of automation.2)Research on CD methods based on scale sets constraints.Scale sets can be regarded as a collection of image sequences of different scales.CD results from different scales are usually treated independently,and multi-scale fusion is thus implemented to combine different CD results.However,contextual information in scale space is ignored when incorporating scale factors only through multi-scale fusion.To address the above issue,two OBCD methods based on scale sets constraints are presented,namely OBCD method based on multi-scale propagation and OBCD method based on multi-scale bag-of-words model,which introduce scale factors and the contextual information of scale space into OBCD from different perspectives respectively.In multi-scale propagation method,multi-scale representation is established from the point view of multi-scale segmentation,where scale sets constraints are constituted by using father-child mapping relationship between scale layers,then uncertain information is reduced and CD results are refined by comprehensively analyzing the information from both coarse scale and fine scale.While multi-scale representation is constructed from the perspective of image pyramid in multi-scale bag-of-words model method.First,pyramid images are generated based on raw HSRI and segmented label images;second,multi-scale bag-of-words model is established through which multi-scale histograms are generated and cascaded for the multi-scale representation of each segmented object;finally,similarity analysis of histograms is implemented to determine the change category of each segmented object.3)Research on the CD result refinement methods based on label smoothing.Acting on the assumption that objects are independent to each other,similarity analysis of objects is implemented to generate final CD results in existing OBCD methods,which ignores the spatial interaction between neighboring objects caused by the factors of segmentation errors and image quality.Smoothing prior constraints are implemented on the initial CD results by label smoothing methods,which can make full use of spatial contextual information.Therefore,the mostly used label smoothing methods based on local filtering and Markov random field(MRF)are summarized and a refined MRF model is presented,which is capable to delineate spatial interaction between neighboring objects more accurately by making full consideration of spectral and shape features between neighboring objects.First,initial label image is generated from initial CD result generated using OBCD method.Second,spectral and shape features(mean value,the length of shared boundary,area)between neighboring objects are extracted from the difference image,and prior distribution knowledge of objects in the difference image is considered to construct the label field and feature field in MRF energy function;Finally,the optimized label of each object is determined by solving the energy function using Graph Cuts algorithm,namely the change property.