Study On Spectral Characteristics And Estimation Models Of Different Nutrient Contents In Forest Soils Based On Hyperspectral Technology

Soil nutrient conditions is the basic factor of soil quality change. Understanding the soil nutrient status quickly have a great significance for studying the vegetation growth of forest land. The technology of hyperspectral remote sensing in soil is one of the hot spots in current research, which has the characteristics of rapid determination of soil nutrient content,and provides a new way to estimate soil nutrient content quickly and accurately.Taking Lushan District, Jiujiang and Wanli District, Nanchang in North of Jiangxi province as the study area, on the basis of considering soil properties, altitude, elevation and other factors,318 valid samples were collected according to soil sample distribution in 2014, based on the determination of the main nutrient of soil organic matter, total nitrogen,total potassium, total phosphorus, considering the abundance of iron in the soil of Jiangxi Province, the content of available iron in soil was also determined., and hyperspectral reflectance by ASD FieldSpec 4 spectroradiometer in the laboratory. On the basis of the analysis of the spectral characteristics of soil type and main nutrients in soil, the correlation between soil nutrients contents and soil spectral reflectance was analyzed by correlation coefficient method. Then, the hyperspectral direct estimation models of soil organic matter,the direct and indirect estimation models of total nitrogen and available iron were built and validated based on partial least squares regression ?PLSR?, BP neural network?BP?, support vector regression ?SVMR? and fixed, ariable weight combination models.The main findings were as follows:?1?The correlation between forest soil nutrientscontents and spectral reflectance was analyzedThe correlation between forest soil nutrients contents and spectral reflectance was analyzed by Correlation coefficient method. The correlation between the content of soil organic matter and the original spectral reflectance was negatively correlated, and the most obvious correlation was 600?800 nm, the maximum value of the relationship in the 650 nm band; the correlation between the content of soil total nitrogen and the original spectral reflectance was negatively correlated, and the most obvious correlation was also 600?800 nm, the maximum value of the relationship in the 649 nm band; the correlation between the content of soil total phosphorus and the original spectral reflectance was negatively correlated, but is not obvious;the correlation between the content of soil total potassium and the original spectral reflectance was not obvious, the value was very low; the correlation between the content of soil available iron and the original spectral reflectance was negatively correlated, and the most obvious correlation was 600?800 nm too, the maximum value of the relationship in the 673 nm band.Correlation analysis between soil organic matter content and total nitrogen, total potassium, total phosphorus and available iron content in soil by SPSS 17.0.There was a significant correlation between soil organic matter and total nitrogen, it can be fitted by power function; there was the positive correlation between the content of soil organic matter and total phosphorus content,but it could not be fitted by function; the correlation between soil organic matter content and total potassium content was very poor; there was a significant linear correlation between soil organic matter content and available iron content,it can be fitted by linear function. The results provide the basis for indirect estimation of soil total nitrogen and available iron content.?2? Hyperspectral single estimation models of major forest soil nutrients content were establishedOn the basis of correlation analysis, hyperspectral single direct estimation models of forest soil nutrients content, hyperspectral single indirect estimation models of forest soil total nitrogen and available iron content respectively were built by PLSR, BP and SVMR.The results show that BP is the best estimation model.In the direct estimation, the best estimation model of soil organic matter content was BP model, with the prediction R² to 0.77, RMSE as 11.96g kg^-1, RPD to 1.99; the best estimation model of soil total nitrogen content was BP model, with the prediction R² to 0.63, RMSE as 0.52g kg^-1 RPD to 1.64; the best estimation model of soil total phosphorus content was PLSR model, with the prediction R² only to 0.09, RMSE as 0.17g kg^-1, RPD to 1.05; the best estimation model of soil total potassium content was BP model, with the prediction R² to 0.20, RMSE as 3.23g kg^-1, RPD to 1.12; the best estimation model of soil available ironcontent was SVMR model , with the prediction R² to 0.39, RMSE as 27.84mg kg^-1, RPD to 1.24. In the indirect estimation, the best estimation model of soil total nitrogen content was BP model , with the prediction R² to 0.89, RMSE as 0.27g kg^-1,RPD to 3.02; as well as the best estimation model of soil available ironcontent was BP model, with the prediction R² to 0.52, RMSE as 26.03mg kg^-1, RPD to 1.44.The results showed that BP and SVMR were better than PLSR, and the indirect prediction of soil total nitrogen and available iron content was better than that of direct prediction.?3? Hyperspectral combination estimation models of major forest soil nutrients content were establishedOn the basis of single model calibration results, hyperspectral combination direct estimation models of forest soil organic matter, total nitrogen and available iron content,hyperspectral combination indirect estimation models of forest soil total nitrogen and available iron content respectively were built by six kinds of fixed weight and four kinds of variable weight method.The results show, In the direct estimation, the best estimation model of soil organic matter content was the binomial coefficient of variable weight combination model, with the prediction R² to 0.81, RMSE as 10.24g kg^-1, RPD to 2.32; the best estimation model of soil total nitrogen content was the simple weighted fixed weight combination model, with the prediction R² to 0.70, RMSE as 0.47g kg^-1, RPD to 1.84; the best estimation model of soil available ironcontent was theerror square sum reciprocal variable weight combination model, with the prediction R² to 0.49, RMSE as 24.97mg kg^-1, RPD to 1.39. In the indirect estimation, the best estimation model of soil total nitrogen content was the simple weighted variable weight combination model, with the prediction R² to 0.73, RMSE as 0.43g kg^-1 ,RPD to 1.91; the best estimation model of soil available ironcontent was the binomial coefficient of fixed weight combination model, with the prediction R² to 0.58, RMSE as 25.03mg kg^-1,RPD to 1.49. Compared with the single estimation model, the variable weight combination model is better than the fixed weight combination model,and the indirect combination model is better than the direct combination model.In this study, based on the soil hyperspectral data, the spectral characters of soil organic matter, total nitrogen, total phosphorus,total potassium and available iron were studied and the estimation models of soil organic matter, total nitrogen and available iron content was discussed. It offers a new research method for the spectral estimation of soil nutrient content, and offers a theoretical basis and foundation for the rapid determination of soil nutrient content by remote sensing.