Study On Soil Physical And Chemical Information Detection Methods Based On Spectral Technology

In view of that soil is not only the natural resource that human beings depend on,but the most basic means of production in agriculture,it is of great significance for understanding and improving soil production potential by using advanced technical means to accurately measure its various physical and chemical indices.The type of soil texture is an important indicator of soil fertility,water retention,and aeration capacity.Therefore,the type determination of the soil texture can accurately provide a scientific basis for the rational use and development of soil resources.In addition,considering that the growth of crops is inseparable from the supply of soil nutrients,the accurate detection of soil nutrients has important guiding significance for determining the type and level of fertilization,which even can play an irreplaceable role in avoiding the abuse of chemical fertilizers and better protecting land resources.The detection of heavy metal content in the soil can not only provide a guarantee for the safe production of crops,but effectively prevent it entering the human food chain through the enrichment of crops,thus avoiding endangering human health.What's more,the growth of crop roots in soil directly reflects the sensitivity of the crops to nutrients,and on this basis,the accurate and dynamic monitoring of soil roots can be of great significance for screening predominant gene crops.This paper studies the detection methods on soil physical and chemical properties based on various spectral and spectral imaging technologies,and further develops the soil physical and chemical property detection system,which mainly includes the following aspects:(1)Based on the near-infrared hyperspectral and XGBoost algorithm,the paper predicts the total nitrogen,organic matter content,p H and EC of soil samples,and further establishes a quantitative prediction model of near-infrared hyperspectral.XGBoost.The XGBoost algorithm is applicable to the spectrum analysis and can realize the contribution rankings of the modeling effect in the feature dimension,thereby helping to filter the feature wavelength.The results show that the near-infrared hyperspectral can have good performance on predicting the soil total nitrogen and organic matter content.Specifically,the determination coefficients of the best models in the prediction set are more than 0.84 and with over 2.0 of RPD,but the detection accuracy for the soil p H and EC still needs to be further improved.(2)Based on the ensemble learning method,the paper adopts Raman spectroscopy and microscopic imaging to judge and predict the soil texture.And the ensemble learning can effectively eliminate the impact of random samples on the premise of small samples.According to the related results,there is a certain correlation between the Raman spectrum and the soil texture under the influence of fluorescence,especially 0.64 score of F1 further indicates that there is still a long way to achieve the accurate prediction.In addition,the combination of CNN model with ensemble learning method can comprehensively determine the soil texture,and further improve the prediction accuracy,especially the highest score of F1 is 0.50.(3)Under the premise of soil matrix effect,the paper adopts LIBS and various data analysis methods to quantitatively detect multiple heavy metal elements(Cu,Ni,Cr,Pb)in soils in agricultural areas.Based on the results,the modeling and forecasting effect based on multivariate analysis has a greater advantage than univariate analysis,especially the PCR and LASSO algorithms have the prominent representations due to their feature dimension reduction.Considering the four elements,the optimal prediction determination coefficients R2 are: Cu,0.94;Ni,0.93;Cr,0.91;Pb,0.89,respectively.In addition,LASSO can also select the characteristic emission line in the spectral dimension with the regularization penalty term,which can provide more interpretation for LIBS spectrum based on the results of data analysis.(4)Based on the deep convolutional neural network,the paper proposes the segmentation model Seg Root for crop roots under the complex soil background,sets two hyperparameters of width and depth in the design of the CNN model,as well as further puts forward the most cost-effective model Seg Root-8-5 by comprehensively weighing its overall parameter amount and the segmentation accuracy.And this model can only use CPU for the fast segmentation prediction on the host computer without advanced GPU.Thus,a prediction binarization mask of Seg Root-8-5 can help to establish an estimation model for the total length of soil roots,especially the determination coefficient of the total length of predication and artificial counting can reach to 0.98.(5)The paper develops a low-cost and portable wireless near-infrared detection system See Soil for detecting the physical and chemical properties of soil by using the cost-effective near-infrared spectroscopy technology.With the small size(110 mm × 80 mm × 50 mm)and only 500 g weight,the detection equipment of this system has the highly integration characteristics and can work without any external equipment(power supply,light source and data processing PC),especially suitable for in-situ detection of soil.The system can be connected to the Internet through Wi-Fi,and users can control the detection system through the physical buttons,as well as local and remote web pages.In addition,according to evaluation and detection results for the soil physical and chemical properties in the soil data set with the See Soil system,the prediction performance of the organic matter content is good,and the determination coefficient of the model in the prediction set is 0.64,while the prediction performance of soil total nitrogen is excellent and even reaches 0.84 of the determination coefficient.