A Model For Soil Iron Oxides Estimation Based On Machine Learning And Spectral Information

Iron oxides are an important component of rocks and soils,a source of nutrition for crop growth and food production,and can be used as a diagnostic indicator in the classification of Chinese soil systems.The soil of Jiangxi Province is strongly weathered and leached,and the biological cycle is active,so the iron and manganese are relatively enriched in the B layer,and the iron oxides content is high.Traditional laboratory analysis of soil iron oxides is time-consuming and laborious,and the operation steps are cumbersome,which can no longer meet the needs of modern agricultural development.The continued development of portable soil proximal sensors has improved our ability to use spectral information to assess soil conditions in the field.In the process of obtaining soil spectral information by portable proximal sensors,there is variability between the soil spectral information obtained by the sensors due to the complexity of the soil composition.This may affect the spectral response between the imaging results and the soil iron oxides,thus affecting the model accuracy.A detailed study of the causes affecting the spectral information in the B-layer soil has not been conducted yet.Smartphones and imaging spectrometers,as portable proximal sensors for acquiring light multispectral and hyperspectral data,respectively,are also gradually being used in the field of soil property prediction,but there are no studies that have explicitly analyzed the intensity of the spectral response of soils acquired by smartphones.The selection of modeling approaches has been a hot topic in the field of soil property estimation research.It is beneficial to clarify the influencing factors of different modeling approaches to better combine the number of soil samples,color space models and soil property characteristics to select the corresponding modeling approaches.With the rise of machine learning algorithms in the field of soil property estimation,there is much interest in predicting soil property content based on machine learning.However,a single machine learning algorithm is used in the current study,and no systematic study of machine learning algorithms in soil iron oxides estimation has been conducted.These problems need to be solved urgently for the estimation of iron oxides in B-layer soils using visible multispectral and hyperspectral data using machine learning algorithms.In this paper,150 soil samples from layer B in 70 profiles in Jiangxi Province were studied based on soil iron oxides(free iron oxide(Fed),total iron(Fet),and freeness of iron(Fed/Fet*100%))with soil spectral information obtained from 5smartphones and the SOC710 VP imaging spectrometer.An inversion model of soil iron oxides based on spectral information input variables was developed,combining Determination Coefficients(R2),Root Mean Squared Error(RMSE)and Relative Percent Deviation(RPD)and the spectral response between chemical properties and soil.To reveal the influencing factors of soil iron oxides spectral properties;to elucidate the variability of spectral information acquisition by different proximal sensors;and to clarify the influencing factors of machine learning on soil iron oxides inversion results.The main findings are as follows:(1)In the visible range,the presence of organic matter(Soil Organic Matter,SOM)affects the spectral response of soil iron oxides.Among them,V,i,s,a*,b*,and c* all showed highly significant negative correlation with SOM,and the correlation between V and SOM was the largest.V indicates the brightness of the soil,the higher the SOM content,the darker the soil color;conversely the brighter the soil color.SOM showed a highly significant negative correlation with R and a*,with R denoting the red light band and a* denoting the soil redness.That is,the lower the SOM content,the redder the soil.SOM decreases the redness of iron oxides in soils.Since iron oxides in Jiangxi soils mainly exhibit red color,SOM can suppress the spectral response of soil iron oxides by decreasing the redness.(2)In the visible range,the smartphone has a spectral information acquisition capability equal to or slightly inferior to that of the SOC710 VP imaging spectrometer.Smartphones can be applied to soil spectral data acquisition and quantitative inversion of soil iron oxides.The results of the models constructed by the five smartphones using the four modeling methods are: Phone 5 > Phone 4 > Phone 2 >Phone 3 > Phone 1,and the accuracy of the model constructed by Phone 5 is the best among the five smartphones.Combining the spectral response intensity and modeling results of smartphones and imaging spectrometers,it can be concluded that the stronger the spectral response of the instrumentation is,the more beneficial it is to obtain soil spectral data,and the spectral response intensity has a direct impact on the modeling accuracy.(3)The accuracy of iron oxide model estimation is: BPNN > PLSR > SVM >RF,and the machine learning algorithm is slightly better than PLSR.Whether modeling with visible multispectral or hyperspectral data,BPNN demonstrates its extreme fault tolerance,and together with thousands of training steps to reduce the variability between actual inputs and outputs,its comprehensive performance is superior to that of SVM,RF,and PLSR.Although the overall accuracy of the BPNN model is better than that of SVM,RF,and PLSR,both RF and PLSR have certain modeling accuracies as well as the highest stability in the entire model,and PLSR has a model accuracy slightly better than or equal to that of SVM.Therefore,when evaluating the accuracy of the model,it is necessary to consider a variety of factors in model construction: the model’s own algorithm characteristics,the number of soil samples,spectral data,and soil properties.