Diagnosis Of Nitrogen.Phosphorus And Potassium Deficiency Based On Temporal Dynamics Of Rice Leaf Image

Image processing technology has been widely used in plant nutrition diagnosis,but most studies are mainly focus on the canopy and leaf information at different growth stages,ignoring the dynamic characteristics which is valuable for the exploration of character symptoms and improvement of diagnostic effect.Therefore,scanning was applied in this study to acquire the time series images of rice leaf under nitrogen(N),phosphorus(P)and potassium(K)deficiency,and image processing technology was used to extract the entire leaf features and local features for dynamic analysis.Based on this,feature selection was taken to optimize the data set which would be used in model establishment and validation.The detailed contents and conclusion of this study are as follows:1.Establishing the database of leaf dynamics under NPK deficiencyLeaf images in this study were acquired in 2014,2015 and 2016 by scanning.To reveal the temporal dynamics of rice leaf,the first incomplete leaf and three fully expanded leaves were marked and scanned every 3days from 20 days after seedling transplantation(DAT20)to DAT44.2.Temporal dynamics of rice leaves under different levels of NPK deficiencyLeaf growth is a dynamic process,dynamic changes of morphology and color are the mainly features.Therefore,dynamic analysis is mainly focused on these two aspects.For leaf morphology,the leaf area was used to reveal the leaf extension rate and senescence rate;for leaf color,different color indices(red,green,normalized red index et al.)were chosen to reveal the color dynamics under NPK deficiency.According to the results,higher nutrient supply resulted in a faster leaf extension rate and a lower senescence rate,leaves with N deficiency presented the lowest extension rate and the fastest senescence rate,followed by P and K deficiencies.Therefore,N deficiency has the biggest effects on leaf growth,followed by P deficiency and then by K deficiency.In order to apply the dynamic leaf features in the establishment of diagnostic model,quantification of dynamic characteristics was carried out by calculating the relative growth rate(RGR)of different indices.Moreover,RGR was calculated with different time interval(data sets calculated by 3days interval named P1-P7,and for 6days interval named P1'-P3')to establish dynamic index data set for following analysis.3.Identification of NPK deficiency based on dynamic leaf characteristicsIn this study,identification of different kinds of NPK deficiency is the first step,and then the stress of NPK deficiency would be identified.Stepwise discriminant analysis(SDA)was used to optimize data set,fisher discrimination with leave one out cross validation(LOO-CV)method were taken to identify NPK deficiencies.The feature selection results showed that the optimal data set contained different types of indices for different leaf positions.RGR(leaf area,leaf perimeter,eccentricity,principle component analysis index,normalized red index,green)were the major constituents of the optimal data set of the first incomplete leaf,RGR(chlorotic leaf area,leaf perimeter,leaf length,blue et al.)were the major constituents of the optimal data set of the 3rd fully expanded leaf.For the 1st and 2nd fully expanded leaves,color indices were the main components of the optimal data set.The discrimination results showed that the diagnostic effects of different leaf positions changed with time,and the first incomplete leaf and the 3rd fully expanded leaf were the optimal leaf positions in diagnosis:the diagnostic accuracy of the first incomplete leaf decreased with time,but for the 3rd fully expanded leaf the accuracy increased with time;the diagnostic effects of the 1st and 2nd leaves presented the state of fluctuating,the accuracy were around 50-60%.Moreover,the first incomplete leaf performed better at early stage(training accuracy 77.30%,validation accuracy 69.30%at DAT26),and the 3rd fully expanded leaf generated the highest accuracy at later stage(training accuracy 100.00%,validation accuracy 95.00%).4.Identification of different levels of N,P and K deficiency based on dynamic leaf characteristicsThe optimal leaf positions for identifying the level of N,P and K deficiency were same:the first incomplete leaf was the ideal indicator for early stage diagnosis and the 3rd fully expanded leaf was the optimal leaf position for later stage diagnosis.In the identification of different levels of nitrogen deficiency,the RGR(leaf perimeter,eccentricity,kawashima index,normalized red index et al.)of the first incomplete leaf outperformed the other indices;the RGR(leaf area,leaf width,normalized red index,green,red of chlorotic part)of the 3rd fully expanded leaf outperformed the other indices.The first incomplete leaf performed better than the other at early stage,achieving training accuracy 69.80%and validation accuracy 61.90%at DAT26;at later stage,the 3rd fully expanded leaf would be the optimal leaf position,the best diagnostic accuracy is achieved at DAT35(training accuracy 100.00%,validation accuracy 95.20%).To identify the phosphorus stress level,RGR(leaf area,leaf length,red,blue and normalized red index)of the first incomplete leaf and RGR(leaf area,perimeter,leaf length,leaf width,chlorotic area et al.)of the 3rd fully expanded leaf were effective indices in identification.Likely,the first incomplete leaf achieved higher accuracy at early stage(data sets P2 and P1' produced training accuracy 70%,validation accuracy 65%at DAT26);and the 3rd fully expanded leaf achieved the highest accuracy at DAT38:training accuracy 100.00%,validation accuracy 91.70%.During the identification of potassium stress level,the optimal indices of the first incomplete leaf were RGR(leaf area,leaf perimeter,green,blue,normalized red index and principle component analysis index),and RGR(leaf area,leaf perimeter,chlorotic area,red and principle component analysis index)of the 3rd fully expanded leaf would be the optimal indices.The optimal leaf in identification at early stage was the first incomplete leaf(training accuracy 78.70%and validation accuracy 68.50%at DAT26),and it was the 3rd fully expanded leaf performed better at later stage(achieving the highest accuracy at DAT35:training accuracy 100.00%and validation accuracy 97.80%)Given the above,this study analyzed the temporal dynamics of rice leaf at extension,stable and senescence stage,making a deep exploration of character symptoms for diagnosis.The NPK deficiency had been successfully identified by using the dynamic characteristics of the first incomplete leaf and the 3rd fully expanded leaf,providing new thoughts and methods for nutrition diagnosis study.