| Soil properties in entire soil profiles or different soil horizons are of great importance in the studies of soil genesis, development, classification, and other soil processes. Traditional soil information acquisition is time-consuming and costly, but remote sensing technology can provide soil information at various scales rapidly and periodically, and has been widely applied in researches such as soil resource survey, land quality evaluation, soil classification, and soil mapping. Traditional methods for measuring soil physical and chemical properties are time-consuming, complicated and costly, and they cannot meet the needs of rapidly monitoring soil property changes. In recent years, the spectrometric technology, which is rapid, simple, and non-destructive in quantitatively analyzing samples, were extensively used in various fields and achieved good results. The data acquired through using imaging technology and spectroscopy technology together have both high spatial resolution and high spectral resolution, thus being able to provide very rich soil remote sensing information which builds a solid foundation for quantitative monitoring of soils and soil properties in the horizontal dimensions. However, studies in soil science thus far show that we still lack an imaging technique specifically for measuring soil total nitrogen (TN) contents in entire soil profiles with high spatial and spectral resolutions. At present, quantitative researches on soil properties mostly use samples collected in the top layer of0-15cm or0-20cm depth; studies on soil point samples in the profile depth of0-100cm were rarely reported, and we have not seen in literatue any study that mapped soil properties vertically along entire soil profiles.In view of this situation, this study used a hyperspectral camera (400-1000nm in753spectral bands) with a CCD of1004pixels×1002pixels for data acquisition. We first analyzed the effects of different pre-treatment methods and modeling methods on prediction results, and explored the optimal prediction model using the soil spectral reflectance data from0-100cm soil profiles collected from the Qianjiang area. Then we established a soil calibration model using the Vis-NIR spectral reflectance data of soil point samples collected from the Xianning area, and further applied this model to soil TN content inversion and mapping using the Vis-NIR hyperspectral image data from three complete soil profiles in the same area, so as to examine the capability of the imaging spectroscopy in predicting soil TN contents along vertical soil profiles. At the same time, we designed a Matlab-based soil hyperspectral image processing system for processing relevant data. Main achivements from this study are listed as follows:1. Soil TN prediction based on Vis-NIR spectraThe method of soil modeling is one of the main factors influencing the accuracy of soil property quantification with Vis-NIR spectroscopy. After drying, grinding and sieving the48point soil samples within the depth of0-100cm in12soil profiles collected from Qianjiang, we obtained their soil spectra. In this study we compared the performances of three calibration methods-PCR, PLSR, and BPNN, based on the Vis-NIR reflectance spectra for soil TN prediction. The spectra data of48soil samples in the470-1000nm wavelength range were processed using the method of first-order derivative transformation combined with second-order Savizky-Golay smoothing, and then leave-one-out cross validation was adopted to determine the optimal factor number. The results indicated that the two linear models (PCR and PLSR) were able to meet the general prediction requirement with little difference:their R2indices are0.74and0.8, respectively, and their RPD values are2.23and2.22, repsectively. The two nonlinear models, built by BPNN combined respectively with PCR and PLSR, are superior to the linear models of PCR and PLSR in prediction precision. The BPNN-PC model used the principal components (PCs) resulting from the principal component regression (PCR) as itsinput, while the BPNN-LV model used the first4latent variables (LVs) obtained from PLSR as its input. Among them, BPNN-LV has the best performance (R2=0.9, RPD=3.11). Therefore, BPNN-LV can be a good model for rapidly and accurately predicting the vertical spatial distribution of soil TN using Vis-NIR spectroscopy.2. Soil profile TN inversion and mapping based on Vis-NIR imaging spectroscopyPre-processing should be conducted on the spectral images of the complete soil profiles taken from Xianning before soil TN inversion. Geometric correction of profile hyperspectral images was performed using photos taken by a digital camera with a fixed-grid background. This method solved the deformation problem of spectral images caused by technological limitations in instruments and platforms, and adjusted the image precision to1mm. Soil image data (160pixels×980pixels) and effective spectral bands (470-1000nm) were kept through image scribing in spatial and spectral dimensions. Comparison of multiple supervised classification methods on the geometrically corrected and scribed images found out that the minimum distance method was the best in distinguishing invalid data (e.g., shadow and crack) from soil data. We proposed a "sampling panel" method, which make the template collect block or strip-shaped samples according to constraint conditions, then average the samples (similar to ROI), and finally solve the scale inconsistency problem between point samples and profile spectra. We further built a PLSR calibration model using the spectrum data of10spot soil samples so as to forecast the TN contents in three complete soil profiles using their spectral images. Results showed that Vis-NIR imaging spectroscopy technology could be used in soil TN inversion and mapping along vertical profiles, and was able to meet the general prediction purpose. Test with measured data obtained a R2vaule of0.56and a RPD value of1.41for the depth of0-100cm, which indicated that the prediction method reached the coarse estimation level. The R2and RPD values were better for the0-60cm depth, with the former being0.87and the latter being1.76; this also indicated that the Vis-NIR spectrum tecnology might have some limitation in the vertical direction. But checking the predictions of each profile separately found that the large forecast deviation of XL-1in the60-100cm depth lowered the total prediction effectiveness of the three profiles in the0-100cm depth; the test results for XL-1in0-60cm depth had R2=0.94and RPD=2.19, implying the model was excellent and could accurately predict, but those in the depth of0-100cm had R2=0.15and RPD=1.06, indicating a large decrease in prediction accuracy. The test results for XL-2and XL-3in the0-100cm depth showed that their models reached the level of rough prediction, with R2being0.91and0.93, and RPD being1.81and1.69, respectively. Above results concluded that this study had preliminarily established a set of procedures for soil TN inversion and mapping using the Vis-NIR spectroscopy technology, and the method was feasible for coarsely estimating soil TN contents in whole soil profiles.3. Design and implementation of the soil hyperspectral image data processing systemBecause of the huge soil hyperspectral image data, it is inefficient to deal with them directly. The existing image processing softwares can only achieve universal functions, and not have some functions, such as removing invalid values in image, sampling in the templates and so on. Based on Matlab2010b, a graphical user interface (GUI) was designed and compiled for processing soil hyperspectral image data. The system has the following functions:soil hyperspectral image data reading, image scribing, invalid value elimination, sample template, prediction value position reduction, accuracy evaluation, etc. It can meet the research needs and is easy to use. It improved our working efficiency and obtained good results in this study. At the same time, it also made up the deficiency of related professional software systems (such as ENVI4.7and Pls-Toolbox7.0.2) in processing methods. |
PDF Full Text Request