Rice Planting Region Extraction Based On Temporal And Spatial Feature Fusion

The purpose of this paper is to explore the strategy of rice information extraction based on the fusion of temporal and spatial characteristics.Specifically,"super-pixel" will be introduced into the current classification of low spatial resolution sequential image.Based on the super pixel segmentation method,the high spatial resolution image will be segmented to obtain the super-pixel image with high internal homogeneity,and then the terrain texture information with high-precision will be obtained;Subsequently,the superpixel will be used as the basic unit to obtains the average time series curve of each super pixel based on the weighted average method;Finally,the similarity between each superpixel and the standard time series curve of the target figure is calculated by using the curve similarity measurement method,and the final classification result is obtained by reasonable threshold division.In general,it is divided into three parts:(1)Research on cloud processing method based on GF-1 visible bandAlthough the revisit period of GF-1 can reach 4 days,but its PMS sensor imaging width is small(60km),so to obtain the GF-1 PMS image covering the whole research area,it is necessary to select the image with good quality from a large time range,especially the low cloud image.In order to make the most of the existing GF-1 image,this paper explores the cloud processing algorithm of GF-1 image,and proposes an Object-oriented Cloud Extraction with Space and Spectrum information,OCESS)to obtain the distribution of cloud areas in the study area,in the validation of the method,the mapping accuracy is 93.68% and the user accuracy is 99.40%,so the performance is good.(2)Remote sensing data preprocessing and super-pixel segmentation optimizationThe preprocessing operation of image with high spatial resolution is mainly to carry out radiometric correction to obtain the real reflectance data of ground objects.After that,because SLIC algorithm is selected for super-pixel segmentation in this paper,in order to meet the data requirements of this algorithm,panchromatic fusion processing is carried out for GF-1 PMS image in panchromatic and multispectral bands based on Pansharping method.The image generated by this fusion method has no color distortion,image quality reduction and other problems.For MODIS NDVI time series image,in order to improve the quality of MODIS NDVI time series data as much as possible,it is necessary to smooth it.In this paper,three kinds of time series curve smoothing methods are compared and analyzed.Based on variance analysis,correlation analysis and other methods,the smoothing effect of each filtering method is judged,and finally EMD algorithm is selected as the best filtering algorithm.For the preprocessed GF-1 PMS panchromatic image,in order to ensure the quality of super-pixel segmentation,based on the principle of OSTU algorithm,this paper proposes a super-pixel scale optimization algorithm.On the basis of determining the appropriate scale,the super-pixel segmentation is completed by simple linear clustering algorithm(SLIC).(3)Rice information extraction based on morphological similarity distance.In this paper,based on Euclidean distance and considering the shape distribution characteristics of the curve,an algorithm is proposed to enlarge the difference between features according to the feature distance of corresponding elements(Morphological Similarity Measurement Algorithm,MSMA),to better extract rice information,and the effectiveness of the algorithm is verified.In order to verify the reliability of the algorithm,this paper will compare the performance of two different algorithms(DTW method and KMeans method)in rice information extraction.The results show that the overall mapping accuracy of the simple MSMA method is 93.29%,the overall area accuracy is 80.57%,the overall mapping accuracy of the super pixel based MSMA method is 90.70%,and the overall area accuracy is 89.48%.In general,the effect of the latter is better.At the same time,it is found that the mapping accuracy of DTW method based on super-pixel fluctuates too much,the spatial-temporal structure of single cropping rice is irregular,and the extraction area deviates from the actual situation greatly,which shows that,firstly,the feature of DTW algorithm--the way of time axis distortion is not conducive to the extraction of single cropping rice;secondly,The fluctuation range of the spectral response of rice phenological characteristics in the same time period in different years is small,so the influence of this factor can be ignored.Therefore,the effect of extracting single cropping rice information by DTW based on super-pixel is inferior to MSMA based on super-pixel.For KMeans method based on super-pixel,it is found that although its performance in extracting single cropping rice information is better than that of DTW method based on super pixel,the spatial distribution of single cropping rice extracted by KMeans has the characteristics of concentration,and one of the disadvantages of KMeans is still unavoidable,which is easy to fall into the problem of local optimization.In conclusion,compared with DTW method based on super-pixel and KMeans method based on super-pixel,MSMA method based on super-pixel is more effective in extracting single cropping rice information.