Study On Estimation Model Of Soil Organic Matter Content In Plowed Layer Based On Convolutional Neural Network

Soil organic matter is an important component of soil solid components and an important source of nutrients for plants.Quick and accurate soil organic matter content is of great significance for the development of precision agriculture.Although the traditional soil organic matter test method has high accuracy,it takes a long time and high cost,and cannot meet the monitoring requirements of a large area.Hyperspectral remote sensing has become a new method for quickly obtaining a large range of soil organic matter content due to its advantages such as narrow and multiple bands,high spectral resolution,and rich information.But spectral estimation accuracy needs to be improved.Therefore,this article takes the soil samples of cinnamon soil in Zhangqiu District,Jinan City,Shandong Province as the research object.Based on the internal relationship between the soil organic matter content in the plow layer and the top soil,this study uses a convolutional neural network to establish a hyperspectral estimation model of the organic matter content in the plow layer.The main research contents and conclusions are as follows:(1).Characteristic factors of spectral data of soil organic matter in Zhangqiu city were determined.This study analyzes the spectral characteristics of cinnamon soil.The research uses first order differential,square root,reciprocal,logarithm and 8 methods to transform spectral reflectance data.By analyzing the correlation between the original spectral data and the transformed spectral data and the content of soil organic matter,the sensitive band of organic matter to the spectral signal was determined.Feature factors are extracted using the principle of maximum correlation.The results show that the sensitive bands of organic matter in the brown soil in the study area are 485 nm ~ 760 nm,1375nm ~ 1382 nm,2120nm ~ 2140 nm,2330 ~ 2350 nm.The first-order differential of the inverse square root and the first-order differential of the inverse logarithmic transform have better effects,and the correlation between the transformed spectral data and the organic matter content is significantly improved.The spectral data of the 845 nm,1474nm,1592 nm,2045nm and 2318 nm bands after the first order differential transformation are selected as the characteristic factors,and the absolute value of the correlation coefficient is greater than 0.65.(2).Determined the main components of soil organic matter spectrum data in Zhangqiu.In order to make full use of the spectral information,the principal component analysis method is used,and the SPA software is used for PCA operation to select the principal components of the spectrum.The results show that the cumulative contribution rate of the 32 principal components reaches a maximum of 100%;the cumulative contribution rate of the first 5 principal components extracted reaches 99%.(3).The hyperspectral convolutional neural network indirect estimation model of cultivated soil organic matter was established.Based on the extracted spectral feature factors and principal component data,use Python to programmatically process the data and build a network model(CNN).The generated four-dimensional spectral information array is used as input data,and the measured value of soil organic matter content in the cultivated layer is output data.Using BP neural network,PLSR and support vector machine models for comparative analysis.The results show that when using the characteristic factors for organic matter modeling and estimation,the decision coefficients of BP neural network,PLSR model,SVM model,and CNN model are 0.6374,0.5056,0.6272,and 0.8036,respectively.When the principal component factor is used for modeling and estimation of organic matter,the coefficients of determination are 0.4256,0.6574,0.6772,and 0.8227,respectively.The decision coefficient of the CNN model is significantly higher than the BP neural network,support vector machine,and PLSR model.This shows that the convolutional neural network is feasible and effective for hyperspectral estimation of soil organic matter content in cultivated layers.