Cotton Nitrogen Nutrient Estimation And Application Based On Chlorophyll Fluorescence

【Objective】Nitrogen is an important factor influencing the growth and development of cotton and yield formation.Real-time and accurate information on plant nitrogen content is the key to efficient nitrogen fertiliser use.Compared to traditional plant nutrient monitoring methods,remote sensing technology has the advantages of being fast and non-destructive.However,the current nutrition monitoring based on single remote sensing information is easily affected by the external environment,and the accuracy and stability of the model is not high enough to meet the demand for accurate monitoring.Chlorophyll fluorescence is known as an ideal"probe"of plant health and can directly reveal the photosynthetic status of the plant.Therefore,in this study,we determined the most suitable time for collecting leaf fluorescence parameters and canopy SIF by analysing the correlation between leaf chlorophyll fluorescence parameters,canopy SIF and plant nitrogen content,and constructed a cotton nitrogen nutrition monitoring model based on leaf chlorophyll fluorescence parameters,canopy SIF and"fluorescence parameters+SIF"fusion using machine learning and integrated learning methods.The model is based on leaf chlorophyll fluorescence,canopy SIF and"fluorescence+SIF"fusion.A linear nitrogen estimation model based on the fusion of leaf chlorophyll fluorescence parameters,canopy SIF and"fluorescence parameters+SIF"was also constructed using subjective and objective weighting methods,with a view to improving the intuitiveness and ease of use of cotton nitrogen monitoring by applying a simple linear model.The results of the study provide a new and effective means to further improve the accuracy of cotton nitrogen nutrient monitoring and the rapid and effective scientific assessment of cotton nitrogen content.【Methods】This study was conducted on Xinjiang drip-irrigated cotton from 2020 to 2022 at the Second Company of Shihezi University Teaching Experimental Farm,with the variety’Xinluzao 53’.Five nitrogen treatments were set up,N0(0 kg/hm~2),N1(120 kg/hm~2),N2(240 kg/hm~2),N3(360 kg/hm~2)and N4(480kg/hm~2),and chlorophyll fluorescence parameters and canopy SIF were obtained from the top four main stem leaves(L1-L4)at flowering,blomming,blomming and boll,boll and boll open stages.Four machine learning and integrated learning methods,four subjective weights,two objective weights and combined subjective and objective weights were applied to establish a cotton nitrogen monitoring model based on fluorescence parameters,SIF and"fluorescence parameters+SIF"fusion.The models were validated and evaluated for accuracy.【Results】(1)The chlorophyll fluorescence parameters Fv/F0,Fv/Fm,ΦPSII,Fv’/Fm’and Fv of the top four leaves of cotton showed a decreasing trend with the advancement of the reproductive period,and the overall performance was L2>L1>L3>L4 among leaf positions and N3>N4>N2>N1>N0 among nitrogen treatments;the correlation analysis between each fluorescence parameter and plant nitrogen content was used to screen out the parameters Fv/Fm,qP,qN,ΦPSII and Fv’/Fm’that correlated well with The leaf positions L2 and L3 and fluorescence parameters Fv/Fm,qP,qN,ΦPSII,and Fv’/Fm’that correlated well with the nitrogen content of cotton were screened out by analyzing the correlation between each fluorescence parameter and plant nitrogen content,and it was found that the estimation model accuracy and stability of RF was the best among the four machine learning and integrated learning methods(R~2=0.912 in the training set and R~2=0.681 in the validation set),but the fluorescence model based on The accuracy of the linear estimation model based on subjective and objective weights of fluorescence parameters and nitrogen was lower.(2)The SIF of cotton canopy showed a trend of increasing and then decreasing with the advancement of fertility,and the overall performance among nitrogen treatments was N3>N4>N2>N1>N0;the correlation coefficient between SIF and plant nitrogen content observed at 16:00was the highest at 0.720;machine learning and integrated learning methods to construct cotton nitrogen content estimation models showed good estimation results,among which the RF model performed(R~2=0.938 in the training set and R~2=0.647 in the validation set);the linear estimation model of SIF and nitrogen in cotton canopy based on subjective and objective weights was significantly better than the estimation model of chlorophyll fluorescence parameters,among which the model of XGBoost combined with Critic method was the best(R~2=0.658 in the training set and R~2=0.725 in the validation set).(3)Comparative analysis of machine learning and integrated learning models based on the fusion of fluorescence parameters with SIF feature levels resulted in the best RF model with L2 fluorescence parameters fused with SIF,with R~2=0.957 and RMSE=0.779 g/kg in the training set and R~2=0.791 and RMSE=0.155 g/kg in the validation set;the L3 fluorescence parameters fused with SIF Light GBM model had the best results with R~2=0.958,RMSE=0.739 g/kg in the training set and R~2=0.741,RMSE=1.969 g/kg in the validation set;the accuracy of cotton nitrogen estimation model based on feature-level fusion improved by 60.79%on average compared with the accuracy of leaf chlorophyll fluorescence parameter model,and improved by The accuracy of cotton nitrogen estimation model based on feature-level fusion improved 60.79%on average compared with the leaf chlorophyll fluorescence parameter model and20.31%compared with the canopy SIF model.(4)The accuracy of the linear model based on the combination of subjective and objective weights improved significantly compared with the model based on the subjective or objective weights alone.The results of the linear model for nitrogen content estimation of cotton plants with the fusion of two data sources of"fluorescence parameters+SIF"are as follows:the optimal linear model of"L2 fluorescence parameters+SIF"is XGBoost-entropy weight model y=25.822x+7.748,with R~2of 0.776 and RMSE of 1.726 g/kg in the training set and R~2of 0.777 and RMSE of1.314 g/kg in the validation set;the optimal linear model for"L3 fluorescence parameter+SIF"is the RF-Critic model y=28.999 x+5.781,with R~2of 0.857 and RMSE of 1.379 g/kg in the training set and R~2of0.837 and RMSE of 1.121 g/kg in the validation set.(5)Based on the existing cloud platform for accurate monitoring and intelligent diagnosis of cotton nutrition,the two types of cotton nitrogen optimal estimation models based on machine learning and one-dimensional linearity constructed in this study were integrated and applied to develop The fluorescence monitoring system was tested and applied in Shihezi General Farm,and the results showed that the application reduced nitrogen fertilizer application by 12.78%,increased yield by 6.55%,and increased revenue by 152 Yuan per mu in cotton fields.The application of the platform is of great significance for the precise management of cotton field fertilization,improving the efficiency of fertilizer utilization and the improvement of the overall benefit of cotton production.【Conclusion】In this study,we compared the effects of different modelling methods on the accuracy of nitrogen estimation models,and concluded that the accuracy of the model with the fusion of fluorescence parameters and SIF feature levels was significantly higher than that of the single leaf fluorescence parameter and canopy SIF model,and the RF model with"L2 fluorescence parameter+SIF"was the best.