Classification And Evaluation Of Land Cover Based On Time Series Remote Sensing Images

With the increase of satellite remote sensing image acquisition methods and the reduction of purchase cost,it is possible to construct multi-source time series remote sensing image data integration in the same region,which makes the research on land cover classification based on remote sensing images become a new research hotspot.Accurate time-series land cover products can better reflect the dynamic changes of regional land use,and provide strong data support for the monitoring of land use and future development planning of the region.Probability graph model is the most popular research tool in the field of land cover classification.This paper selects two of the most representative models: Markov Random Field(MRF)and Hidden Markov Model(HMM).In order to compare the classification performance between these two models and the traditional supervised classifier,this paper constructs the time series remote sensing image dataset from 2007 to 2015 with the Poyang Lake Ecological Economic Region as the research area,and designed six experimental schemes for analysis.The main research contents of this paper include:(1)In order to improve the spatial continuity of the initial classification results,a context MRF image classification model is constructed based on MAP-MRF framework.The multi-dimensional Gaussian distribution is used to fit the distribution characteristics of various feature categories,and the Potts model is used to describe the relationship between the two double-point groups in the second-order neighborhood system.Then the specific form of the energy function is determined by this method.Finally,the ICM algorithm is used to solve the energy function minimization problem.(2)In order to improve the time consistency of the initial classification results,a HMM-based time series surface coverage classification model is established.The initial probability distribution is calculated by statistical initial classification results,the transition probability matrix is determined by considering the difficulty of transition between different land cover labels,and the observed probability distribution is calculated based on the probability output of the basic classifier combined with the Bayesian formula.After determining the three elements of the model,the Viterbi algorithm is applied to solve the optimal state chain.(3)In order to evaluate the classification performance of the six experimental schemes more comprehensively,four evaluation indexes were constructed for the analysis of experimental results.The index of pixel label change can reflect the stability of classification results,the index of illogical transition can reflect the time consistency of classification results,the index of unreasonable spatial distribution can reflect the spatial continuity of classification results,and the index of classification accuracy reflects the correctness of classification results.(4)Based on the probabilistic calculation method of hidden state chain in hidden markov model,a reliability evaluation strategy for sequential land cover classification products is proposed.In view of the two situations with or without probability output,the transition probability matrix and the observation probability distribution involved in the state chain probability calculation are processed differently to obtain the probability value of each pixel state chain of the image,and the reliability evaluation is conducted by probability classification statistics.