Research On Detection Methods Of Orchard Soil Nutrient Based On Visible/Near Infrared Spectroscopy

Many provinces in China have the situation of blind and excessive fertilization.This cancause the waste of production resources and the increase of farmer production cost, and willcause the pollution of the environment, especially the eutrophication of the lake river.Inaddition, due to lack of technical guidance in the application of chemical fertilizer, theproportion of the problem of imbalance, nitrogen, phosphorus fertilizer and potassiumdeficiency, how to guide farmers reasonable fertilization is imminent.In modern agriculture,the important part of variable fertilization is to understand the distribution of soil nutrient, andthe quick analysis and test of the nutrient content in soil has been the bottleneck of modernagriculture.This paper takes Jiangxi Wanan a navel orange orchard soil as the research object.Amodular small soil nutrient detection device was designed and soil spectra were collected.6different preprocessing methods are compared, and3kinds of band filter algorithm are usedto simplify the variables. Finally, the different forecasting models were expected to develop aportable instrument for the rapid detection of soil nutrients, contributing to the developmentof agriculture. To verify the accuracy of the instrument, A comparative study of soil nutrientcontent using Brook TENSOR37type spectrometer and the establishment of the forecastmodel.The main conclusions of this work are as follows:(1) TENSOR37（BRUKER）spectrometer is better to normalization method use as datapretreatment of the soil nutrients original spectral, this method not only can eliminate certainnoise, but also makes the characteristic peaks prominent, and can reflect the real situation ofsoil nutrient near infrared spectroscopy. For the total nitrogen (TN) model, the best result wasthe partial least squares regression (PLS) model, the factor number is5, at the same timemakes the prediction correlation coefficient increased (Rp) to0.9494, and the root meansquare error of prediction (RMSEP) reduced to0.0118%. PLS model of the Total phosphorus(TP) gives the best result with the correlation coefficient (Rp) of0.6562and the root meansquare error of prediction (RMSEP) of0.1173g/kg with the factor number is2. For the totalpotassium (TK) model, the best result was the partial least squares regression (PLS) model,the factor number is5, at the same time makes the prediction correlation coefficient increasedto0.9760, and the root mean square error of prediction reduced to1.8528g/kg. PCR model ofthe soil Organic matter (SOM) gives the best result with the correlation coefficient (Rp) of0.9787and the root mean square error of prediction (RMSEP) of4.1543g/kg with the factornumber is4. (2) The self design portable soil nutrient instrument is based on visible light andshortwave near infrared, this part include design of scheme, selection of main parts used onthe portable instrument, collecting process of the soil spectra. This paper also discussed theproblem of original spectral preprocessing, and the results showed that the Baseline method isbest.using baseline can remove baseline drift with time caused by instability of light sourceand makes the soil nutrient spectrum more close to the real situation.(3) This paper also establishes models and optimize model of the self design portable soilnutrient instrument, and use Genetic Algorithm（GA），Successive Projections Algorithm（SPA）and Competitive Adaptive Reweighted Sampling (CARS) to select variables of thesoil nutrients original spectral, and makes the model simply and rapidly. We have4indicators,each group of3through the first3kinds of algorithms, and a total of12.(4) For the portable instrument, we need the simultaneous detection of nitrogen,phosphorus, potassium and organic matter, and also require the speed fastly, so we chooseCARS-PLS model as the model of the portable. For the total nitrogen (TN) model, theCARS-PLS model, it makes the prediction correlation coefficient increased (Rp) to0.8969,and the root mean square error of prediction (RMSEP) reduced to0.0133%. CARS-PLSmodel of the Total phosphorus (TP) gives the best result with the correlation coefficient (Rp)of0.8187and the root mean square error of prediction (RMSEP) of0.0890g/kg. For the totalpotassium (TK) model, the CARS-PLS model that makes the prediction correlationcoefficient increased to0.9719, and the root mean square error of prediction (RMSEP)reduced to2.2291g/kg. The CARS-PLS model of the soil Organic matter (SOM) gives thebest result with the correlation coefficient (Rp) of0.9734and the root mean square error ofprediction (RMSEP) of4.6377g/kg.(5) In summary, the modeling precision is decreased for the portable, but the difference issmall compared to the laboratory instrument. so it can meets the requirement of application.