Research On Soil Organic Matter Inversion In Mountainous Cultivated Land Based On Hyperspectral Remote Sensin

The mountainous area of Guizhou Province accounts for nearly 90%.The fragmented distribution of cultivated land in mountainous areas leads to low utilization rate of some cultivated land and uneven distribution of soil texture,which limits agricultural production to a certain extent.Soil organic matter(SOM)content is an important index to measure soil fertility.It is of great significance to monitor SOM content quickly and accurately for scientific management of cultivated land in mountainous areas.In 2022,the national development document No.2 focused on Guizhou in depth,emphasizing that the vigorous development of modern high-efficiency agriculture with mountain characteristics is an important guarantee for the implementation of the rural revitalization strategy.Therefore,it is of great significance to explore a more effective algorithm to estimate the SOM of cultivated land in mountainous areas to improve the efficiency of agricultural production in mountainous areas.In recent years,hyperspectral remote sensing technology has been gradually applied in SOM rapid detection due to its advantages of high timeliness,large amount of information and no pollution.In this paper,the hyperspectral data of 100 soil samples in the field of Guiyang pepper demonstration base in Guizhou Province were taken as the research object,and the SOM estimation model of Guiyang pepper base was constructed based on spectral transformation method and machine learning.The soil samples collected in the field were detected to obtain the visible-near infrared band spectral information,and the soil spectral data were smoothed and denoised.Four spectral data transformations(first-order differential,second-order differential,first-order differential of reciprocal logarithm,continuum removal)and four types of models(partial least squares regression,support vector machine,random forest and BP neural network)were used to combine different models for predicting SOM content.The soil between the carriages was used as the training sample,and the soil on the carriage was used as the verification sample.By analyzing and comparing the accuracy of the model,the model combination with the highest comprehensive stability is selected.In order to ensure the universality of the model for SOM inversion of cultivated land in mountainous areas,this paper takes the cultivated land soil within the boundary of 13 counties(districts and cities)in Guizhou Province as the test set to test the model operation effect.The conclusions are as follows:(1)The models constructed by airborne hyperspectral and ground spectral data sources have the ability to estimate SOM content.The airborne hyperspectral data modeling has the basic ability to estimate the SOM content in mountainous areas,and the ground object spectrum has excellent ability to quantitatively predict the SOM content.(2)The hyperspectral data of cultivated soil in mountainous areas improved their correlation with SOM to varying degrees by spectral transformation.The first-order differential data transformation can better highlight the information of soil organic matter in mountainous areas.The number of sensitive bands that pass the significance test is the largest.Among them,the airborne hyperspectral data performs the first-order differential transformation through the significance test.There are up to 28bands,and the ground object spectrum reaches up to 941.At the same time,the model established by using the first-order differential has the highest accuracy;(3)The number of sensitive bands and the span of band range obtained by correlation analysis of SOM in mountainous areas are positively correlated with the accuracy of the prediction model.The sensitive bands of SOM in mountainous areas exist in visible light-near infrared.The visible light is concentrated in 480～780nm,and the near infrared is concentrated in 800～853nm,1032～1115nm,1180～1223nm,1303～1435nm,1486～1598nm,2249～2283nm.(4)In estimating the SOM content of cultivated land in mountainous areas,the partial least squares model has a rough estimation ability;the SOM estimation ability of the support vector machine model was significantly improved compared with the partial least squares model.Random forest is better than the first two,but the accuracy of the verification model is not the best;the BP neural network in the nonlinear model is suitable for SOM estimation of cultivated land in mountainous areas with its high accuracy and good stability.Among them,the first-order differential-BP neural network has the best prediction effect(Training set:R~2=0.852,RMSE=2.534;the validation set:R~2=0.878,RMSE=3.315,RPD=2.425),which is more universal for SOM monitoring in Guizhou.