Rice Growth Status Monitoring Using Imagery From Unmanned Aerial Vehicle Platform

Rice is one of the most important crops and major food for people around the world,and the development of unmanned aerial vehicle(UAV)has provided a new platfonn for rice growth status monitoring.Four field experiments involving different varieties,nitrogen rates,and planting density were carried out at the paddy field located in RuGao,Jiangsu.Multispectral(MS)camera,digital camera and modified NDVI camera were mounted in the UAS,and autopilot to take imagery of the study area through the whole growth period of rice.In this study,we explored the image pre-processing methods,and proposed a new method to mapping the vegetation fraction from UAV imagery with texture and DN value extracted from the imagery.In addition,a comparative analysis of the relationship between the vegetation index and LAI,yield were taken,and quantitative regression models for rice LAI and yield estimation were established.The prospective results would provide a nev technical support for a rapid and non-destructive monitoring of rice growth.In this study,we explored the method of mapping the rice vegetation fraction from UAV imagery,and proposed to combine DN value with texture for classification pixel by pixel.Two texture features(mean and coefficient of variation(C.V))were calculated in this study,and image was classfied into pure classes of rice,road and mixed classes using the C.V value.Then the background was distinguished from the mixed classes by comparing between mean value and sensitive parameters.Several classification methods,such as parallelepiped,minimum distance(MinDist),Mahalanobis distance(MahaDist),maximum likelihood(MaxLik),support vector machine(sSVM),and neural network(NN)were tested on the three types of image for rice plant detection.The results showed that the method proposed in the study achieved more stable and higher detection accuracy than other classification methods,with an average overall accuracy of 93.53%,91.25%and 92.55%on the MS image,digital image and NIR-G-B image,respectively.And the optimal band or vegetable index for classification of the three sensor images were 800nm,ExG,NIR.Window size was influenced by the resolution of the image and the altitude the image acquired,the best detection result would get with window containing a line of rice and a line of background.In summary,these results might provide a technical approach to mapping the vegetation fraction rapidly from a low altitude UAV image,and lay a foundation for reducing the influence on spectrum from background and improving prediction accuracy of the crop growth by image processing.Furthermore,we gave an exploration for the quantitative relationships between rice leaf area(LAI)and the vegetation indices.The spectral index and color index were calculated from the MS image and digital image,respectively.Color index showed a high correlation with the LAI at the periods from initial tillering to late booting stage,and vegetation index VARI obtained highest correlation(R2=0.77)based on digital image.While the correlation decreased significantly from late booting stage to filling stage for the reason that these indices saturated to high LAI from late booting stage.Spectral index calculated from MS imagery showed a higher correlation with LAI than color index,and the index DVI[800,720]estimated LAI from initial tillering to filling stage with a highest accuracy(R2=0.78).Late booting stage with high LAI value was identified as the hardest stage for LAI estimation,Index VARI and NDVI[800,720]were used to estimate LAI based on the 2 types of image.According to the results,index containing red edge band and near red band showed a high correlation with LAI.Six color indices and six spectral indices were obtained from the MS imagery and digital imagery.The relationships between rice yield and vegetation indices at different growth stage were analyzed.The results indicated that vegetation indices at different stages were highly correlated with yield.The highest correlation was achieved at the booting stage with correlation coefficient of 0.68 and 0.75 for VARI for digital imagery and NDVI[800,720]for MS imagery,respectively.Further analysis was conducted on the relationships between yield and the multi-temporal vegetation indices.The highest correlation coefficients reached 0.79 with MLR(NDVI[800,720])at booting stage and heading stage for three experiments based on the MS images,while the highest R2 for digital image was 0.73 with MLR(VARI)from jointing stage and booting stage for exp.3 and exp.4.The ?-CLAI-YIELD model was established for yield prediction with the cumulative LAI(CLAI)estimated from remote sensing data.And the DVI[800,720]-CLAI-YIELD model from jointing stage to filling stage,VARI-CLAI-YIELD model from jointing stage to booting stage showed a highest correlation with grain yield for the MS images and digital image,respectively.In conclusion,three types of models were suitable for yield prediction,and the model based on the booting stage index was a simple and effective method,while the ?-LAI-YIELD model had more mechanism.