Research On Near Infrared Prediction Methods Of Soil Available Phosphorus Based On Transfer Learning

Soil available phosphorus is an important nutrient component affecting crop growth and an important index reflecting the ability of soil to provide phosphorus and guiding the application of phosphorus fertilizer.So,rapid and accurate detection of its content is of great significance in precise fertilization of crops and environmental protection.Lab physical and chemical methods are usually used to detect the content of soil available phosphorus and they often cause problems such as long measuring period,complicated process and being prone to pollute the environment,which can not meet the requirements of the development of modern precision agriculture.Visible and Near-infrared(Vis-NIR)spectroscopy analysis technology has the advantages of high efficiency,simple operation and no pollution.Combined with machine learning method,it can detect the content of soil available phosphorus quickly and nondestructively.Transfer learning is an important research issue in machine learning,which has fundamentally relaxed the basic assumptions of traditional machine learning and has been widely concerned in the field of machine learning.Aiming at the problems of low prediction accuracy and model failure when the training samples and test samples of soil available phosphorus prediction model come from different regions.In this study,the yellow red loam sample in Southern Anhui was taken as the source domain sample,the Shajiang black soil sample in Northern Anhui was taken as the target domain sample,and the soil available phosphorus was taken as the research object.Based on the Vis-NIR spectroscopy analysis technology,a transfer learning method was proposed to deal with the different distribution of soil spectral data in different regions,and a model of soil available phosphorus migration in different regions was constructed.The main research contents and results are as follows:(1)The effects of different spectral pretreatment and machine learning modeling methods on soil available phosphorus prediction models are studied.The visible and near-infrared spectral data in the range of 350-1655 nm are collected from the samples in the source domain and the target domain.The prediction models of soil available phosphorus in the source domain and the target domain are constructed by combining the original spectrum and the pretreated spectrum with Partial Least Squares Regression,Support Vector Regression and Classification and Regression Tree algorithms respectively.The results show that the prediction accuracy of the model can be effectively improved by choosing the appropriate preprocessing and modeling methods,after the spectrum is processed by Savitzky-Golay convolution smoothing and the standard normal variate transformation,the Support Vector Regression algorithm is used to correct the modeling.The prediction accuracy of the source domain(R2 and RPD of the model were 0.85 and 2.59)and the target domain(R2 and RPD of the model were 0.61 and 1.60)is effectively improved.(2)Feasibility of sharing models for predicting soil available phosphorus in different domains was studied.Firstly,the target region is predicted based on the prediction model of soil available phosphorus in the source region;Then,the combined samples are constructed as training samples,and the soil available phosphorus prediction model in the transfer source domain is used for the prediction of the target domain.The results show that,based on the assumption that training samples and test samples are distributed independently and identically by traditional machine learning,the source domain model cannot be used directly in the target domain,and the model failure problem occurs(R2 and RPD of the model are-0.19 and 0.92).Combinatorial samples add the target domain samples on the basis of the source domain samples.When it is used as training samples,compared with the former,the distribution differences between the model training samples and the test samples reduce,and it alleviates the problem of model failure and improves prediction accuracy significantly(R2 and RPD of the model are 0.54 and 1.47).(3)Transfer learning algorithm is used to construct soil available phosphorus transfer models in different areas.Aiming at the impact of different distribution of samples in the source domain and the target domain on the feasibility of sharing the source domain model in the target domain,Two-stage Tr Ada Boost.R2 and Transfer Component Analysis(TCA)are used to process the spectral data in the source domain and the target domain with different distribution,and the soil available phosphorus transfer model is constructed.The results show that the feature-based TCA transfer algorithm has better transfer ability than the instance-based Two-stage Tr Ada Boost.R2 transfer algorithm when the sample distribution of the source domain and the target domain is very different.The transfer learning algorithm can solve the problem of different distribution of samples in the source domain and the target domain,realize the use of the source domain samples efficiently and transfer the source domain model to the target domain.The obtained TCA transfer model R2 and RPD are 0.79 and 2.18 respectively,which improves the prediction accuracy of soil available phosphorus content in the target domain.(4)The spectral detection system of soil available phosphorus has been developed and improved.The soil available phosphorus spectrum detection system is developed with Python language,which realizes four functions: sample data uploading,sample spectrum preprocessing,transfer model construction and transfer model prediction.The core of the system is the built-in transfer learning algorithm,which enables the input of a small amount of soil sample data of the target area to obtain the prediction model of soil available phosphorus with relatively high accuracy in the area,and reduces the modeling cost and time of the user.The system provides users with convenient and efficient tools for model construction and sample detection,and provides help for the promotion of the large-scale construction of soil available phosphorus prediction model.In summary,this study solved the problems of low prediction accuracy and model failure when the model training samples and test samples came from different regions,and improved the accuracy of soil available phosphorus content prediction in northern Anhui province.It provides theoretical basis and technical support for low-cost and efficient construction of soil available phosphorus prediction models in more areas,and provides a new idea for promoting the prediction of soil available phosphorus content in a large area.