Multi-temporal Remote Sensing Image Classification Based On Conditional Random Fields

With the development of remote sensing technology,the spatial,temporal and spectral resolution of remote sensing images are increased constantly.Massive multi-temporal remote sensing images are further accumulated compared with the past,which supply abundant information for the analysis of spatial-temporal behaviors of land cover objects.The comprehensive use of multi-temporal remote sensing images to improve classification accuracy has become a new research hotspot.Conditional Random Field(CRU)model treats the image as a probability map,and directly models the posterior probability of class labels,and obtains the optimal classification result through learning and inference algorithm.Therefore,we use CRF to model the spatial and temporal context to classify the multi-temporal remote sensing images,and a better classification accuracy is obtained.The specific work of this dissertation is as follows:1)This dissertation studies the conventional image classification methods and their latest applications in the field of remote sensing.The background and basic theoretical framework of the CRF model are studied.The construction methods of potential functions,learning algorithms and inference algorithms of CRF models are summarized,and some probabilistic frameworks of CRF models commonly used in image classification are illustrated.2)A multi-temporal remote sensing image classification method based on CRF is proposed.The conventional methods for classifying multi-temporal remote sensing images applied transfer matrix as the temporal context information which was set manually.Therefore,it is difficult to obtain the transfer matrix correctly and cannot make full use of the temporal context information.To address this problem,this dissertation proposed a novel multi-temporal remote sensing image classification method based on conditional random fields.Firstly,we applied Expectation maximization algorithm to create the temporal potential that describes the temporal context information.Then,constructed the conditional random field model by means of combing the spatial and temporal context information.Finally,used the model to classify remote sensing images.Extensive experiment results demonstrate that the proposed method effectively improves the classification performance on remote sensing images.3)A multi-temporal remote sensing image classification method based on high order CRF is proposed.The high order potential energy can increase the regional constraint and the ability to preserve the boundary of objects.Aiming at the creation of effective representation of high-order potential energy in high-order CRF model,we use super-pixel segmentation algorithm to get super pixels,and extract high-order potential energy from super pixels.The experimental results show that the high order CRF model can improve the classification accuracy of remote sensing images.