Research On Hyperspectral Inversion Of Soil Salt Content In Tumochuan Plain

The soil is closely related to the survival of human beings and the sustainable development of society.The parameters of the content of all elements in the soil determine the life of plants.The soil has a certain degree of self-purification ability like rivers and lakes,but the purification ability is relatively speaking.If it exceeds a certain limit,it will not be restored to its original stable state,such as excessive use of soil,salinization,and chemical pollution.This dissertation studies the salt content of soil.Excessive salt concentration in the soil will cause plants to fail to grow in a normal form,destroying its original physiological parameters and structure,and not conducive to its growth and nutrient absorption,thereby affecting the regional Economic development and ecological protection have also affected the output of food crops in rural areas,threatening global food security and environmental quality.The Tumochuan Plain in Inner Mongolia is one of the main food producing areas in the province.However,due to the immature irrigation conditions and unreasonable development and utilization in the north,the area of saline-alkali land in this area is getting larger and larger,resulting in a serious shortage of grain production in the area.The quality is getting worse and worse,and the lives of farmers who make a living from farming are getting harder and harder.Quantitative prediction of the soil salt content in the Tumochuan Plain and real-time knowledge of the soil salt content in the area are of far-reaching significance for controlling soil salinization in the area and guiding agricultural production in the area.Therefore,this study used hyperspectral technology to invert the salt content of the soil.In this experiment,sampling points were uniformly arranged in the Potato Demonstration Base in Tumd Left Banner,Hohhot,and the salt content of each sampling point was measured and recorded by the HM-WSYP soil salt rapid measuring instrument.A total of 60 soil samples were collected.The sample soil in this paper was put into a sealed bag and brought back to the laboratory.In this paper,the handheld hyperspectral camera Specim IQ was used to collect the target hyperspectral image data sample library.Secondly,this research preprocessed the collected hyperspectral image data,and used the black and white correction method to normalize the data.Used one-dimensional convolutional neural network to classify images,the hyperspectral image of the soil was extracted to obtain the spectral data of the soil,and the spectral data was spectrally transformed.Then this paper used continuous projection algorithm and random leapfrog algorithm to extract the characteristic bands of the transformed spectral data,and analyzed the correlation between the extracted characteristic bands and the salt content of the soil and selected the ones that pass the P=0.01 significance test.The characteristic band was regarded as the sensitive band.Finally,an inversion model was established to predict the actual content of soil salinity.The results showed that:(1)Spectral transformation could highlight the hidden details of the spectrum,and could also improve the correlation between spectral data and soil salt content.(2)Performing two feature extractions on the spectral data could better remove the redundant information in the spectral data and extracted the sensitive bands.From the extracted sensitive bands,it could be found that the sensitive bands of the soil salt content were concentrated around 500nm and 800nm.(3)Compared with multiple regression,one-dimensional convolutional neural network regression,partial least square regression,and random forest regression,the prediction accuracy of the combination of the two models were significantly improved.The soil spectral data undergoes logarithmic first-order differential transformation,and then the sensitive bands were selected through the continuous projection algorithm combined with the correlation analysis method.The finally established partial least squares and random forest(PLSR-RF)inversion model has the best effect.The determination coefficient R_c~2 of the modeling set was 0.889,and the root mean square error RMSE_c was 0.096 g/kg,relative analysis error RPD_c was 3.131,validation set determination coefficient R_v~2 was 0.941,root mean square error RMSE_v was 0.049 g/kg,relative analysis error RPD_v was 4.117.