Research On Inversion Model Of Soil Nutrients In Orchard Based On Hyperspectral

The overall status of soil nutrient content is the basic factor that affects the changes in soil quality.A quick understanding of the soil nutrient status of orchards is of great significance for studying the growth environment of fruit trees.While the soil provides a carrier for the growth of fruit trees,the nutrient content also plays a decisive role in the growth cycle,fruiting rate and quality of the fruit trees.In this paper,the apple orchard soil of Penglai and Qixia in Yantai City,Shandong Province was selected as the research object,and the soil nutrient organic matter,total nitrogen,available phosphorus and available potassium content and soil hyperspectral data were measured successively.Pearson correlation analysis and segmented extreme value method were used to screen the characteristic wavelengths of the spectral data of soil organic matter,total nitrogen,available phosphorus and available potassium.For problems such as poor estimation effects of traditional models,convolutional neural networks were used to establish A hyperspectral inversion model of orchard soil organic matter,total nitrogen,available phosphorus and available potassium content was developed.The main research contents and results are as follows:(1)Collection of soil samples,nutrient data,and hyperspectral data.Soil samples were collected in 106 representative apple orchards including Penglai and Qixia,the main apple producing areas of Yantai,Jiaodong,and a total of 151 valid soil samples were collected.The organic matter,total nitrogen,available phosphorus,and available potassium content of the soil samples in the study area were successively determined by laboratory chemical methods;the ASD Field Spec 4 ground object spectrometer was used to perform hyperspectral detection of the soil samples,and the visible light-near-infrared continuous spectrum curve was detected.20 spectral curves are measured for each soil sample.(2)A piecewise extreme value method is proposed to select the characteristic factors of the spectral data of different soil nutrients.Perform breakpoint correction on the original spectral data,remove abnormal samples,and then perform S-G smoothing and 9 spectral transformations.Use the principle of correlation to analyze the correlation between the original spectral data and the transformed spectral data with the content of soil organic matter,total nitrogen,available phosphorus and available potassium.Consider the representativeness of the number of characteristic factors,and propose a classification based on the principle of covering the entire band.The method of selecting the extreme value of the correlation coefficient to determine the characteristic factors of different nutrients is used for modeling.(3)The feedback mechanism function is introduced to improve CNN,and a soil nutrient content inversion model based on CNN is established.The characteristic factors are determined as modeling data through various methods of spectral mathematical transformation,maximum correlation coefficient and segmented extreme value method.When modeling,SVR is used to assist CNN for training,and a new training set is constructed according to the training results of SVR.Introduce the feedback mechanism function,use this mechanism to promote the convergence of the model,and then continuously adjust the learning rate and the number of iterations to complete the optimization of the model.The results show that when the characteristic factors are used for organic matter modeling and inversion,the determination coefficients of the PLSR model,SVR model,BPNN model and CNN model are 0.47,0.58,0.61,and 0.9,respectively;when the total nitrogen modeling inversion is performed,the PLSR model The coefficients of determination of PLSR,SVR,BPNN,and CNN models are 0.53,0.42,0.52,and 0.87,respectively;when inversion of available phosphorus modeling is performed,the coefficients of determination of PLSR,SVR,BPNN,and CNN models are 0.62,0.41,respectively,0.51,0.94;in the inversion of the memory quick-acting potassium modeling,the determination coefficients of the PLSR model,the SVR model,the BPNN model and the CNN model are 0.57,0.79,0.63,0.95,respectively.Therefore,the prediction effect of the CNN-based inversion model is better than the commonly used models in terms of absolute coefficient performance.