Rice Area Extraction Studies Based On Oriented-object Method

Rice is one of the most important crops in China,which has the largest planting area and highest yield.Extracting the information of rice production and rice planting distribution is an important realistic significance for food security and social stability.Remote sensing technology has the characteristics of fast speed,wide range and periodicity.Based on the technical characteristics of remote sensing technology,it has obvious advantages in applying rapid acquisition of rice information.This study is based on medium-resolution multi-source remote sensing satellite data(Landsat 8,Sentinel-1)and detailed ground data.Using remote sensing data fusion method combined with object-oriented crop classification.Solve the specific problems in the process of rice information extraction,quantitative analysis of the factors affecting the extraction accuracy of rice planting areas.The purpose is to improve the understanding of multi-source data,object-oriented methods and data mining algorithms in rice remote sensing.This study selects the Dengta as the research area,which is located in Liaoning Province.Comprehensive use of Sentinel-1 radar data and Landsat 8 optical data.Combined with the October 2016 survey of the study area and Digital Globe data,this paper conducts the following two aspects of research :1)Study the effect of improving classification accuracy under multi-source data,Determine the appropriate features and segmentation scales under this data to support more accurate remote sensing identification of rice areas;2)Based on the screening of object features under multiscale segmentation,study the classification accuracy of rice and the resolution of rice recognition under different classifiers,and determine the optimal rice extraction algorithm under different data characteristics.Provide necessary technical support for national food production macro decision-making based on rice planting regional distribution.Two experiments were designed according to the research objectives: 1)Identify rice growing areas under multi-source data Constructing feature space using multivariate features such as spectral data,vegetation index,radar data,and texture data,based on the univariate analysis method,using the object-oriented method to construct the feature space,extract the rice planting area 2)Identification of rice remote sensing multi-algorithm based on multi-space scale Multi-segment scale extraction of rice planting areas for multi-source data,And use the random forest,decision tree,support vector machine to extract rice planting areas at each scale.Compare the accuracy of rice classification for each data under different classifiers.The main conclusions of the paper are as follows: 1)Different multi-feature spaces have different importance in rice classification. Among the spectral data,vegetation index,radar data and texture data in the multivariate feature space,the vegetation index and radar data have significantly improved the classification accuracy of rice.2)In the classification of rice,the support vector machine and the random forest classifier are relatively stable compared to the decision tree classifier.Support vector machine and random forest can still perform excellent in overall accuracy even in the case of less feature space,Increase the feature space,the improvement of the decision tree is obvious,the overall accuracy is increased by 7.9%,kappa coefficient increased by 8.1%.Compared with the decision tree,the support vector machine and the random forest classifier are improved by 3.6% and 2%,and the kappa coefficient is increased by 3% and 2.12%.