| 【Objective】The determination of soil total nitrogen has always been dominated by chemical analysis,which is time-consuming and labor-intensive.The development of spectroscopic technology provides new ideas for rapid monitoring of soil total nitrogen,but due to the weak penetration of soil by the spectrum itself,monitoring deep soil total nitrogen using spectral technology has become a difficult and painful point in the current research field.Therefore,this article establishes a monitoring model for topsoil total nitrogen through indoor and outdoor soil spectral feature analysis,in order to provide theoretical basis and scientific basis for meeting the requirements of rapid total nitrogen nutritional monitoring and precise fertilization management of cotton fields at each tillage layer.【Method】The study area was the Karamay area in Sinkiang,China.Six cotton fields with different soil texture types were selected as research areas,and a total of 120 sampling points were collected(20sampling points in each cotton field).In-situ spectral data of soil surface and 20 soil depth profile samples(collected every 3 cm from 0 to 60 cm depth)were collected at each sampling point,and the collected 2400soil samples were brought back to the laboratory for chemical analysis and indoor spectral acquisition.The soil was divided into topsoil(0-3 cm),shallow tillage layer(3-21 cm),middle tillage layer(21-42 cm),and deep tillage layer(42-60 cm).By analyzing the vertical variation trend of soil total nitrogen content in soils with different texture types and different tillage layers,as well as the correlation between spectral data and soil total nitrogen content,a model for monitoring deep soil total nitrogen was established.In order to improve the accuracy of the model,different spectral data pre-processing methods were combined,and four machine learning models were used to establish direct and comprehensive monitoring models,and the prediction accuracy of direct and comprehensive models was compared and analyzed.【Results】(1)Nonlinear fitting and model establishment of total nitrogen content and soil depth on soil surface:The soil texture types in the study area can be divided into three categories:sandy clay loam,loam sandy soil and sandy loam soil.The total nitrogen variant fitting equations were y=0.8609+0.0116x-7.7846×10-6x~2,respectively.y=0.0149x+1.0242×10-4x~2;y=-0.08461-0.00314x-1.6331×10-4x~2。The coefficient of determination R2 of the fitted equation of the three soil types was greater than 0.7,and the total nitrogen content decreased with the increase of soil depth,and the variation rate in the spatial distribution of soil total nitrogen content was 24.1%～42.3%.Under different soil texture types,the total nitrogen content of soil surface layer and the total nitrogen content of different deep cultivation layer had a strong correlation,and the correlation coefficient was between 0.727～0.886.The correlation analysis results between porosity and soil total nitrogen content of three soil types showed that the two showed a strong correlation coefficient(0.783～0.864),indicating that the soil total nitrogen monitoring model was established by soil quality.(2)Screening for the optimal preprocessing combination and regression modeling method for estimating soil total nitrogen content:The comparison of the accuracy results of monitoring soil TN through selecting different preprocessing combinations and different model methods shows that different preprocessing combinations can improve the corresponding model accuracy,but there are differences in performance.For example,combination 2 has the best effect on improving the accuracy of the random forest regression model(R~2 is 0.858),but the performance is the lowest in the support vector machine regression model accuracy(R~2is 0.501);therefore,in the selection of combinations and models,appropriate combinations need to be carefully selected based on the adaptability of the model to improve the monitoring accuracy of soil TN.(3)Research on the optimization performance of NGO algorithm in optimizing soil TN estimation model:On the basis of selecting appropriate preprocessing combinations,the optimization algorithm Northern Goshawk Optimization(NGO)can improve the accuracy of direct and indirect monitoring models constructed based on in-situ field spectroscopy.Both showed that NGO-GRNN had the best monitoring model performance.In the direct monitoring model,compared with the GRNN model,the R~2 of the surface,middle and deep soil layers increased by 60%,12%,and 19%,respectively.In the comprehensive monitoring model,compared with the GRNN model,the accuracy of soil TN monitoring in the surface layer improved by 15%.(4)Accuracy comparison and applicability study between direct and comprehensive monitoring models:The monitoring model for soil total nitrogen content based on field in-situ spectroscopy is established by two modeling methods:direct and comprehensive monitoring models.The predictive accuracies of the direct and comprehensive monitoring models in each tillage layer are 0.62-0.75 and 0.54-0.87,respectively.Validation and analysis of the two monitoring models with a new soil dataset collected in 2023 show that,compared with the unclassified direct monitoring,the comprehensive monitoring has significantly improved monitoring accuracy overall.The most significant improvement is in the monitoring of sandy loam soil,with an increase in monitoring accuracy of 25%and 8%in the middle and deep tillage layers,respectively,compared to direct monitoring.[Conclusion]:The study mainly concluded that three soil types,sandy loam,loam and sandy loam,were selected,and TN monitoring models of different cultivated soils were established based on field in-situ spectra.The results show that the classification of soil texture can improve the fitting accuracy of TN vertical scale abnormality trend.Different machine learning methods should adapt to different preprocessing methods for the same estimation object.The Northern Goshawk(NGO)optimization algorithm can improve the estimation accuracy of soil TN by random forest regression and generalized propagation neural networks.The comprehensive estimation of the TN model of each cultivated layer by means of surface TN model based on spectrum combined with nitrogen vertical abnormality equation has higher accuracy and better effect than direct estimation of TN model of each cultivation layer based on spectrum,and has certain universality in Shihezi reclamation area,which can provide a theoretical basis for future research on soil TN deep monitoring model and provide technical support and theoretical decision-making for precision agriculture to quickly obtain deep soil nutrients. |
PDF Full Text Request