| In recent years,with the continuous increase of the global population and the continuous deterioration of the environment,food security has become one of the most important security issues of concern to all countries in the world.As an important ration crop in China,rice growth monitoring during its whole growth period and accurate estimation of rice yield have always been the focus of researchers.Monitoring the dynamic change of rice above ground biomass(AGB)during its whole growth period will help agricultural technicians to judge the growth of rice so that they can adjust the field management measures timely to ensure that the rice is always in the best growth condition.Monitoring the dynamic change of rice AGB also provides reliable data for the agricultural technicians' accurate estimation of rice yield.Unmanned Aerial vehicle(UAV)remote sensing technology has become the first choice for researchers to estimate AGB due to its real-time,convenient,and non-destructive observation characteristics.This research is closely integrated with the needs of precision agriculture.The UAV remote sensing platform is equipped with multispectral camera and RGB camera to conduct high-frequency observations of rice growing in a normal field planting environment from the tillering stage to the wax maturity stage.The relationships between various remote sensing data and biomass was analyzed with regard of the field collected rice biomass.Difference estimation methods are proposed based on the difference of data types.Biomass model of the whole growth period is established to provide quantitative decision-making reference for precision agriculture.The main research work includes:(1)The relationships between multispectral vegetation indices and rice biomass during the whole growth period were analyzed.The separation phenomenon between the VIs and the biomass was studied.It was found that the separation phenomenon was caused by the ridge closure under the condition of large ridge double row planting and the heading of rice.The whole growth period was divided into three sections by the time of ridge closure and heading as the boundary.Then the estimation models were separately built.A multispectral image-based biomass estimation model for the whole growth period of rice was established.Most VIs can estimate rice biomass with good accuracy.The best VI is EVI2 RE,and the verification accuracy is R ~2 =0.91,RMSE=185.35 g/m~2,and n RMSE=8.43%.All VIs are severely saturated in the middle section(between ridge closure and heading).Before ridge closure and after heading,the estimation accuracy of the red edge index is better than that of the red band index.Using fully constrained mixed pixel decomposition for the VIs,the relationship between endmember abundance and biomass is also separated into three sections due to the influence of ridge closure and heading.The endmember abundance was used to estimate rice biomass separately.The optimal endmember abundance is the bright leaf abundance,and the verification accuracy is R ~2 =0.89,RMSE=204.50 g/m~2,and n RMSE=9.30%.After combining the VIs and endmember abundance using stepwise multiple linear regression,the estimation accuracy of rice biomass was further improved,and the verification accuracy was R ~2=0.92,RMSE=175.70 g/m~2,and n RMSE=7.99%.For multi-spectral images,using a multiple linear regression model to combine the VIs and endmember abundance can achieve the best estimation accuracy,which can be used to estimate the biomass of rice throughout the whole growth period in a normal planting environment.(2)Based on the RGB indices and the texture feature indices,two models were constructed to estimate biomass throughout the whole growth period.The relationships between RGB indices and biomass during the whole growth period were analyzed.The normalized b-band reflectance is not affected by ridge closure,saturation and heading,and can be directly used for biomass estimation throughout the whole growth period.The verification accuracy is R~2=0.74,RMSE=321.52 g/m~2,and n RMSE=14.61%.The other bands and RGB indices are separated into two sections affected by ridge closure,but they are not affected by saturation and heading.When the models were built separately,the RGB index with the highest verification accuracy is VARI,and the verification accuracy is R~2=0.92,RMSE=117.23 g/m~2,and n RMSE=8.06%.Of the texture indices,the highest accuracy is the CON of the green band along the ridge direction(D10 direction)at a resolution of 6.4 cm.This texture feature is not affected by ridge closure,saturation and heading.The verification accuracy is R ~2 =0.88,RMSE=220.60 g/m~2,n RMSE=10.03%.Combining the RGB index and texture features can further improve the estimation accuracy.The optimal multiple linear regression combination is the combination of GDIS?D-11,b-band and g-band at a resolution of3.2 cm.The verification accuracy is R ~2 =0.89,RMSE=211.19 g/m~2,and n RMSE=9.60%.This model can be applied to the estimation of rice biomass of the whole growth period,without the consideration of ridge closure or heading as a boundary for segmented modeling.(3)The resolution of RGB images and the calculation direction of Gray Level Cooccurrence Matrix(GLCM)texture features have a significant impact on the relationship between texture features and biomass,and the impact varies with the change of the form of texture features.VAR,ENT,and SEC are almost not affected by the change of calculation direction,while HOM,CON,and DIS have differences between different calculation directions depending on the image resolution.When the image resolution is centimeter-level(0.8 cm,1.6 cm,3.2 cm),the change in the calculation direction of texture features has little effect on the relationship between texture features and biomass.When the image resolution is sub-decimeter-level(6.4cm),the direction change has little effect on the texture of the green band,but has a greater effect on the red and blue bands.The correlation between the textures along the ridge direction(D10 direction)and the biomass is greater than other directions.When the image resolution is decimeter-level(12.8 cm,25.6 cm),the texture features of the three bands have differences between different directions.The correlation coefficients between HOM,CON,DIS along the ridge direction(D10 direction)and biomass are greater than other directions,and the difference in the direction of the texture features of the red and blue band is larger than that of the green band.When the image resolution is sub-meter-level(51.2 cm),the change in the calculation direction has a little impact on the texture features of all bands,but the impact is limited and there is no obvious rules.(4)The new indices based on the three-dimensional structure of the canopy can better estimate the biomass of rice throughout the whole growth period,and it is not affected by ridge closure,saturation and heading.Among them,the performance of the canopy structure index,which reflects the three-dimensional fine structure of the canopy,is better than the DSM texture index,which reflects the distribution of the threedimensional structure of the canopy on the two-dimensional plane.The best canopy structure index are the height-correlated indices.Taking HMean as an example,the verification accuracy is R~2=0.83,RMSE=258.95 g/m~2,and n RMSE=11.77%.The most accurate index of the DSM texture indices is DMEA.The verification accuracy is R~2=0.72,RMSE=327.20 g/m~2,and n RMSE=14.87%.(5)Using machine learning algorithms to combine multi-dimensional data can effectively improve the estimation accuracy of biomass.The two machine learning algorithms used can construct models that are not affected by ridge closure,saturation,and heading of various combinations.The model with the highest accuracy is the model constructed using the RF algorithm with a combination of "one-dimensional + twodimensional + three-dimensional" data.The verification accuracy is R ~2 =0.94,RMSE=156.81 g/m~2,and n RMSE=7.13%.It can estimate the biomass of the whole growth period of rice with the highest accuracy. |
PDF Full Text Request