| Real time monitoring of soil physical and chemical properties is a prerequisite and inevitable requirement for the realization of precision agriculture.In the past 20 years,domestic and foreign scholars have established a relatively perfect hyperspectral prediction process of soil properties through continuous experiments and research and achieved high prediction results.However,most of the current studies obtain spectral data in the laboratory,and soil samples need to undergo a series of pretreatment,which undoubtedly reduces the efficiency of hyperspectral technology.Therefore,some scholars began to obtain spectral data in the field to predict soil properties.However,because the field environment is more complex than the laboratory environment,the prediction accuracy of the established model is usually low.Therefore,how to reduce the impact of the field environment,improve the prediction ability of the field spectrum,and meet the"precision"requirements of precision agriculture has become a research hotspot and difficulty in recent years.Based on the previous research results,this study takes the reclaimed soil in Huaibei City,Anhui Province as the research object to explore the impact of soil moisture on the spectrum and the feasibility of soil attribute prediction based on field spectrum,in order to provide technical support for rapid prediction of soil attributes.The main conclusions of this study are as follows:Affected by soil moisture,there are two obvious absorption peaks in the soil spectrum from 400 to 2450 nm,which are located near 1450 nm and 1950 nm respectively.By extracting eight characteristic parameters of the two absorption peaks,it is found that the characteristic parameters of the absorption peak at 1450nm have stronger correlation with soil moisture content,and the correlation between the wavelength position,left half area,symmetry and slope of the absorption band and soil moisture content is better than other absorption characteristic numbers.Through paired sample t-test combined with two-dimensional correlation spectrum analysis,it is found that the main influence bands of soil moisture are 1000～1100 nm,1400～2100 nm and 2200～2300 nm.When the soil moisture content is too low or too high,it will also affect the spectral characteristics of 500～800 nm.By comparing the aggregation degree of Soil Spectrum in principal component space before and after air drying,it is found that the spatial distribution integration degree of Soil Spectrum in wet soil is significantly higher than that in air dried soil,indicating that some soil spectral information affected by water is covered up,and soil moisture reduces the difference between sample spectra.After eliminating the influence of soil moisture by direct standardization algorithm and radiative transfer model,its aggregation degree in principal component space is significantly reduced;Comparing the average absorbance,it is found that the eliminated spectral data approximately coincides with the spectral curve of air dried soil,which shows that both algorithms can effectively eliminate the influence of soil moisture.The extraction ability of soil spectral information varies greatly under the combination of different spectral transformation and band screening methods,and the accuracy of the established soil organic matter prediction model is significantly different.The characteristic bands selected from the original spectrum and absorbance spectrum are mainly concentrated in 400～1500 nm,and the selected characteristic bands account for more than 95%of the whole band.However,due to the high redundancy between spectra,the highest prediction models are class B models,which can only roughly estimate the content of soil organic matter;After the First-order differential transformation,the selected characteristic bands are mainly concentrated in400～500nm,1885～1890 nm and 2300～2450 nm.The selected characteristic bands account for nearly 50%of the whole band,and the inter spectral redundancy is low.92%of the 24 models established reach the class C model,of which the highest RPD is 2.64and the highest R~2 is 0.86,which can accurately predict the content of soil organic matter;The characteristic bands screened under the continuum removal transformation are mainly concentrated in 410～560 nm.The selected characteristic bands are less than 30%of the whole band.25%of the established prediction models reach class C model,in which the highest RPD is 2.37 and the highest R~2 is 0.82.Among the three primary characteristic bands screening,PCC screening has the largest number of characteristic bands,large redundancy between bands,less screening results of spa and cars,but the correlation between some characteristic bands and organic matter is low,which limits the prediction accuracy of the model.In the three kinds of secondary band characteristic band screening,the number of characteristic bands screened by pcc-spa and pcc-cars is greatly reduced,and the accuracy of the established prediction model is improved.The characteristic bands screened by pcc-cars are the least,but the accuracy of some models in the established prediction model is reduced,which may be that the characteristic bands screened by pcc-spa and pcc-cars are not enough to characterize the spectral information of all organic matter,resulting in low accuracy of the prediction model.The spectra of soil samples are quite different under field and laboratory conditions.Compared with the indoor spectral curve,the field spectral curve is more rough,and there are two strong noise bands at 1350 nm and 1780 nm.However,the overall trend of the two spectra is the same.The spectral reflectance increases rapidly in the visible band and decreases gradually after slow growth in the near-infrared band.After filtering and first-order differential transformation,the soil organic matter R~2,RPD and RMSE established by the field spectrum are 0.51,1.43 and 3.24,which only reach the accuracy of class B model.The soil organic matter prediction model R~2 and RPD established by the spectrum after water impact correction by radiative transfer model are increased by 59.61%and 67.13%respectively,and RMSE is reduced by40.12%.The prediction ability of the model is significantly improved and reaches the accuracy of class C model,which can better predict the content of soil organic matter.When using spike algorithm to establish soil attribute prediction model,the number of transmission samples directly affects the ability of hyperspectral prediction of soil organic matter.When there are few transmission samples,its prediction accuracy is low.With the increase of transmission samples,the prediction accuracy of the model increases rapidly and gradually tends to be stable.The prediction model established by spike algorithm has the highest R~2 of 0.70 and RPD of 1.83,which only reaches the accuracy of class B model.Compared with the radiative transfer model,its prediction accuracy is still low.Figure[28]Table[4]Reference[134]... |
