| With the development of remote sensing, spectral resolution and signal to noise ratio is getting higher and higher, near-infrared spectroscopy (NIRS) equipment is developed from the filter, grating dispersion to fourier transform, spectral resolution of hyperspectral remote sensing reaches within10nanotechnology, all these make them not only have the advantages of convenience, non-destructive testing, but also improve precision and accuracy of monitoring by using their sophisticated spectral edge. Soil organic matter (SOM), total nitrogen (TN) content and so on are regarded the most important information in agricultural production. Many studies have approved that soil spectral characteristic are closely related to its physical and chemical properties. So, it is an important basic work that studies on non-destructive spectral analysis technology of estimating soil nutrient, which will provide a quick and accurate method on estimating these soil properties in precision farming.The primary objective of this study was to explore the sensitive wavebands, spectral indices and quantitative models for estimating soil major nutrient information through systematic analysis of hyperspectral and NIR spectral information, using three methods of regression analysis based on characteristic spectral parameters, partial least square (PLS) and back propagation neural network (BPNN), respectively, on the basis of5different soil types with varied fertilities from middle and eastern of China. The anticipated outcome of the study will provide theoretical basis and technical support for estimating soil nutrients information with visible/near-infrared spectra.Spectral characteristics of five different soil types were analyzed comprehensively. Using reflectance of400-1000nm for hyperspectra and the first derivative of1000-2500nm with Norris smoothing filter for NIR spectra, which were compressed from600to6,1501to5using PLS method, respectively, and compressed spectral data was as the input of BPNN, soil type as output. PLS-BPNN models of the soil classification were established. The results showed that both of models were reliable and practicable, this identification method could be used for classification of soil. Using analysis technology of hyperspectra, the relationship between SOM content and the combined spectral indices of two-band in350-2500nm was studied. The relationship between spectral indices based on the first derivative spectra and SOM content were obviously stronger than those composed of original reflectance, with exceptional performance from derivative with Norris smoothing filter. The correlation sequence of SOM with different index types was difference index (DI)>ratio index (RI)>normalized index (ND), regardless of the composition of the original spectral reflectance or the first derivative spectra. DI composed of the first derivative of554nm and1398nm [DI(DR554, DR1398)] performed the best as index correlations with SOM. Testing of the monitoring model based on DI(DR554, DR1398) with independent datasets from different types of soil samples resulted in R2, RMSE and RPD of0.84,3.64and2.98. Results of PLS and BPNN analyses for the first derivative spectra suggested that both methods had a good potential for estimating OM content. Comparing the three methods, the sequence of predictive accuracy was PLS-BPNN model> PLS>DI(NDR554, NDR1398). It also suggested that PLS-BPNN model had a great potential for estimating SOM content, and resulted in the highest prediction accuracy. The model based on DI(NDR554, NDR1398) was somewhat inferior, but this model was simpler, it needed only a few band information, it might provide effective information for estimating SOM content in portable instruments and demonstrated a great future in application.Further analysis was conducted on the correlations between SOM content and various spectral indices including RI, DI and ND derived from NIR spectra. All RI, DI and ND were analyzed for all available combinations with two-band original reflectance and/or the first derivative values of reflectance between1000and2500nm. Therefore a novel spectral index for estimating SOM content using NIRS was established. Results showed that spectral indices composed of reflectance corrected with multiplicative scatter correction (MSC) and Savitzky-Golay (SG) smoothing method were highly correlated with SOM. Spectral indices composed of two bands of first derivative spectra or original reflectance had somewhat lower correlation coefficients. The correlation sequence of SOM with different types of spectral indices was DI>RI>ND regardless of the composition of the original spectral reflectance or the first derivative spectra. DI composed of the reflectance from1883and2065nm corrected with MSC and SG smoothing methods [DI(CR1883, CR2065)] performed the best as linear correlations with SOM. Testing of the monitoring model based on DI(CR1883, CR2065) with independent datasets from different soil types resulted in R2and RMSE of0.837and4.06. Compared with the result from the PLS method, the monitoring model based on DI(CR1883, CR2065) was somewhat inferior, but since it needed only two reflectance bands, and the monitoring model was simpler, DI(CR1883, CR2065) might be a good spectral index for estimating SOM content in portable instruments.Soil TN estimating models were established by using simple regression analysis based on characteristics spectral parameters, PLS and PLS-BPNN, respectively. Results showed that hyperspectral data in500-900and1350-1460nm of the first derivative spectra with Norris smoothing filter and NIR spectral in4000-5500cm-1(1818-2500nm) of reflectance corrected with MSC were the best ranges for estimating soil TN content, respectively. Analysis results of PLS and PLS-BPNN on the two type of spectra data suggested that both methods had a good potential for estimating TN content. R2of calibration were0.81,0.98and0.90,0.98, respectively. R2, RMSE and RPD of validation were0.81,0.219and2.28,0.93,0.149and3.36,0.91,0.147and3.40,0.95,0.121and4.13, respectively. DI composed of the reflectance from1884and2071nm corrected with MSC and SG smoothing method [DI(CR1884, CR2071)] performed the best as linear correlations with TN. Tests of the estimating model based on DI(CR1884, CR2071) with independent datasets from different soil types resulted in R2, RMSE and RPD as0.79,0.24and2.08, respectively. These results suggested that the models established based on the three methods could meet the needs of estimating soil TN content.The quantitative relationship between soil total phosphorus (TP) content and two different sources of spectral data was studied, and performances of simple regression analysis based on characteristics spectral parameters, PLS and PLS-BPNN for estimating TP content was analyzed and compared. Soil TP calibration models based on different methods appeared great differences. For hyperspectral, the results of calibration using PLS and PLS-BPNN based on reflectance of500-1500nm showed that R2of calibration were0.70and0.92. R2, RMSE and RPD of validation were0.64and0.68,115and113.11,1.58and1.61, respectively. For NIR spectra, the results of calibration using PLS and PLS-BPNN based on4000-7500cm-1of the first derivative spectra with Norris smoothing filter showed that R2of calibration were0.79and0.88. R2, RMSE and RPD of validation were0.77and0.84,91.6and78.03,1.98and2.33, respectively. DI(NDR1134, NDR1372) of hyperspectral and DI(R2237, R2261) of NIR spectra performed better as index correlations with TP. R2of calibration were0.61and0.67, R2, RMSE and RPD of validation were 0.27and0.64,161.07and110.49,1.29and1.65, respectively. These results suggested that BPNN model calibrated by NIR spectra could be used to estimate soil TP content precisely. However, models calibrated under other conditions resulted in RPD<2and could only be used to estimate TP content roughly.The calibration models were established using PLS and BPNN techniques to relate spectral data to the content of available N (AN), available P (AP)and available K (AK) for five dried soil types. For hyperspectral data, R2between results from chemical analysis and spectral predicted content, based on calibration of PLS, were0.83for AN,0.60for AP and0.53for AK, R2, RMSE and RPD of validation were0.90,0.28and0.53,15.4,22.4and41.3,3.08,1.17and1.35, respectively. R2based on calibration of BPNN were0.97,0.90and0.88, R2RMSE and RPD of validation were0.90,0.26and0.73,16.19,23.05and30.53,2.50,1.31and1.72, respectively. For NIR data, R2between results from chemical analysis and NIR predicted content, based on calibration of PLS, were0.83for AN,0.61for AP and0.77for AK, R2, RMSE and RPD of validation were0.85,0.46and0.67,19,20and32.4,2.50,1.31and1.72, respectively. R2based on calibration of BPNN were0.97,0.83and0.94, R2, RMSE and RPD of validation were0.91,0.57and0.80,14.61,16.31and25.23,3.25,1.61and2.21, respectively. The results showed that BPNN technique was better than PLS in spectral analysis, and the method was feasible to estimate AN content, but the mean relative errors of test samples for AP and AK were high relatively, therefore, further study should be done in this field. |
PDF Full Text Request