Hyperspectral Estimation Of Major Soil Nutrient Content

Hyperspectral remote sensing is an important development direction of quantitative remote sensing, soil remote sensing is one of the hot spots of remote sensing applications. At present soil testing and formulated fertilization (Balanced fertilization) has been put into practice in the nationwide. The quick detection of soil nutrient content, especially the fast and accurate estimation of the major soil nutrient content has been necessary to the precise agriculture. So the study on applying hyperspectral technology to gain soil nutrient information has already been focused on.Taking Qihe County of Dezhou City in Shandong Province as the pilot area, the study was carried on to collect soil samples in the end of September2009, then to mensurate their nutrient(soil organic matter, alkali hydrolysable nitrogen, available phosphorus and potassium) content and hyperspectral reflectance using ASD FieldSpec3spectroradiometer in the laboratory. On the basis of screening the study samples of soil organic matter, alkali hydrolysable nitrogen, available phosphorus and potassium respectively, the spectral properties of soil and major nutrient were analyzed. After derivative and wavelet transformation of the reflectance, characteristics wave bands of major soil nutrient were extracted by correlation analysis, stepwise multiple linear regression (SMLR) and genetic algorithms in combination with partial least squares (GA-PLS) method. Then, the hyperspectral estimation models of soil organic matter, alkali hydrolysable nitrogen, available phosphorus and potassium content were built and validated based on characteristics wave bands and wavelet coefficients respectively. The mail results were as follows:(1) Characteristic wave bands of major soil nutrient were determinedBased on soil original spectral reflectance and its various transformations, characteristics wave bands of major soil nutrient were extracted using correlation analysis, SMLR and GA-PLS method. The characteristics wave bands of soil organic matter content were415,456～469,482,806～819,832,1051～1064,1415～1428,1684,1813,1920,1969,2185～2198,2283～2296,2356,2372,2423～2436nm. The characteristics wave bands of soil alkali hydrolysable nitrogen were413,449～469,908～1001,1023,1065～1078,1716～1736,1912～1925,2213～2233,2262～2275nm. The characteristics wave bands of soil available phosphorus were477～490,561～581,813～827,983,1469,1499～1512,2234～2254,2311～2331nm. The characteristics wave bands of soil available potassium were463～469, 917,1275-1288,1317-1330,1730,2120,2258,2280,2360-2375nm.(2) Hyperspectral estimation models of major soil nutrient content were established based on characteristic bandsOn the basis of characteristic wave bands, hyperspectral estimation models of soil major nutrient content respectively were built, and the models based on characteristic wave bands selected by GA-PLS method were considered as the best.The best estimation model of soil organic matter content was the partial least squares (PLS) regression model based on the characteristic wave bands of first derivative of reflectance, with the calibration R2to0.97, RMSE as0.48g kg-1, and the prediction R2to0.95, RMSE as0.48g kg-1. The best estimation model of soil alkali hydrolysable nitrogen content was the PLS regression model based on the characteristic wave bands of first derivative of reflectance, with the calibration R2to0.97, RMSE as4.78mg kg-1, and the prediction R2to0.95, RMSE as5.49mg kg-1. The best estimation model of soil available phosphorus content was the PLS regression model based on the characteristic wave bands of first derivative of logarithm of reflectance, with the calibration R2to0.94, RMSE as3.00mg kg-1, and the prediction R2to0.93, RMSE as3.32mg kg-1. The best estimation model of soil available potassium content was the PLS regression model based on the characteristic wave bands of first derivative of logarithm of reflectance, with the calibration R2to0.78, RMSE as28.49mg kg-1, and the prediction R2to0.91, RMSE as21.39mg kg-1. The prediction R2of the best models was greater than0.90, which could indicate that this type of models had the high prediction accuracy for soil nutrient content.(3) Hyperspectral estimation models of major soil nutrient content were built based on wavelet analysisBase on the reflectance and its5kinds of first derivative transformation spectrum of the soil samples, the low frequency wavelet coefficients of1-5level were obtained by using bior1.3wavelet function respectively to build the SMLR and PLS regression estimation models of major soil nutrient content. Studies showed that the calibration accuracy of the SMLR models were generally lower than the PLS regression models. For soil organic matter content, the low frequency wavelet coefficients of1-3level can replace the original spectrum in the PLS regression modeling, for the content of soil alkali hydrolysable nitrogen, available phosphorus and potassium, the low frequency wavelet coefficients of1-2level could substitute for the original spectrum. Considering the calibration R2and the principal component numbers of the models, the best estimation models based on wavelet analysis of soil nutrient content were determined, and had good prediction accuracy by validated. Therefore, based on the wavelet analysis and the PLS regression, the spectra data could be compressed significantly (the compression ratio of the third level coefficent was13%, the second level as25%), which is feasible to forecast soil major nutrient content.(4) The technical process of the major soil nutrient estimation based on Hyperspectra was optimized and put forwardWith a view to the non-destructive, accurate and rapid monitoring of soil nutrient content, the article compared and analyzed the key technique, and developed the integrated technical process of hyperspectral estimation of soil nutrient. Firstly, the first derivative transformation of soil reflectance or its logarithm was obtained. Secondly, the spectra was divided into some regions according to appropriate interval and the characteristic wave bands were selected by GA-PLS. Finally, the hyperspectral estimation models were built based on the characteristic wave bands making use of PLS regression.In this study, based on the soil hyperspectral data, the spectral characters of soil organic matter, alkali hydrolysable nitrogen, available phosphorus and potassium were studied and the estimation models were established, and the rapid and non-destructive estimation technical process of the major soil nutrient content was optimized and put forward. It offers a theoretical basis and foundation for monitoring soil nutrition diagnosis and the guidance of agricultural production using satellite remote sensing in the future, which can accelerate the development of precision agriculture and the applications of hyperspectral remote sensing.