Rice Yield Estimation With Multi-temporal Remote Sensing Abundance And Vegetation Index

Rice is the largest grain crop in China,which occupies a major position in ration consumption,accounting for up to 60%.With the gradual development of China’s economy and the continuous growth of the population,the total amount of rice needed will show an upward trend.In the context of increasing demand,achieving rapid and accurate production estimation is of great significance to food price regulation,agricultural policy formulation,and national food security.Based on the remote sensing typical vegetation index and abundance,this paper used multi-temporal data to conduct a comprehensive study on rice yield estimation.Vegetation index can effectively express spectral information at the level of pixels,and the abundance of end-members can reflect the image features at the level of sub-pixels,and can reflect the distribution of target features in the changing background.In order to effectively combine the advantages of the two parameters,this study used their product as a new parameter.By observing the growth of rice in the experimental area of Lingshui County,Hainan Province,the sensitivities of the three types of elements to yield were compared.Based on the multi-temporal data,this study established a multivariate linear yield estimation model,and recommended the optimal observation time scheme.The main research contents are as follows:（1）According to the spectral characteristics of rice,eight typical vegetation indices were selected,and correlation analysis was used to measure the sensitivity of vegetation indices to yield during each growth period.The results showed that the vegetation indices and yield of the jointing stage and booting stage had the highest correlation among the four periods,and the correlation coefficient were mostly above 0.6,while the correlation coefficient in the tillering stage and the heading stage were mostly in the range of 0.4 to 0.6.（2）Three end-member extraction methods of PPI,N-FINDR and SMACC were used to obtain the end-member spectrum of the image,and the reliability of the end-member spectrum was obtained by comparing the three methods.Based on the results of three end-member extraction methods,the mixed pixel decomposition was carried out respectively,and the correlation analysis between the rice abundance and yield was carried out.The analysis results showed that the PPI method was less affected by noise and the obtained end-member spectrum was more stable and more reliable than the N-FINDR method and SMACC method,and the sensitivity of the obtained rice abundance to yield was also higher.The highest correlation was at the jointing stage,and the correlation coefficient was 0.699.Compared with the vegetation index,the sensitivity of the end-member abundance to yield had no clear advantage.（3）In order to combine the advantages of vegetation index and end-member abundance,this study constructed a product parameter,which multiplied the rice abundance and vegetation indices under the PPI method,and used correlation analysis to measure the sensitivity between the product parameter and yield.After that,compared with the results of vegetation indices and abundance respectively.The comparison results showed that the product parameter can effectively improve the sensitivity of the vegetation index and end-member abundance to yield,and the improving capacity remained stable during different growth periods and facing different vegetation indices.It indicated that the product parameter can effectively combine canopy spectral information of the vegetation indices and sub-pixel level information of end element abundance.The product parameters that were most sensitive to yield during the four growing periods were:rice abundance multiplied by CI_green（550,800）at tillering stage,rice abundance multiplied by CI_rededge（720,800）at jointing stage,rice abundance multiplied by CI_green（550,800）at booting stage and rice abundance multiplied by NDRE（720,800）at heading stage.The correlation coefficients were 0.593,0.714,0.712 and 0.649,respectively.（4）Using the sensitive elements selected from the four reproductive period,the multivariate linear regression model was established by leaving one method for cross validation.The collinearity analysis before modeling showed that it is necessary to screen out the product independent variable in jointing stage,which was serious collinearity with independent variables in booting stage and heading stage.The R² of estimated yield model constructed by the data of tillering stage,booting stage and heading stage was 0.589.The product parameter collinearity between the tillering stage and the jointing stage had little influence on the yield estimation model.When it was necessary to estimate the yield in advance for subsequent field management,we can select the data of the tillering stage and the jointing stage to modeling,and the R² was0.534.If the manpower and material resources are very limited,the data of jointing period can achieve a good yield estimation effect,and model R² was 0.509.