| Nowadays, precision agriculture is at the leading place in the development of agriculture all over the world. To accelerate the process of precision agriculture, it is needed to fully understand the soil productivity and spatial variability, which helps adjust the amount of input into soil, reduce the environmental pollution and improve the quantity and quality of agricultural products. The technology for real-time and quantitatively monitoring and diagnosing soil fertility and crop growth is the very prerequisite to realize precision agriculture. One of the key areas in the outline of "National medium- and long-term program for scientific and technological development (2006-2020)" is the precision agricultural operationalization and informatization, with the content involving the focus of the future research on information digitization gathering technology in growth of fauna and flora and ecological environment and on real-time monitoring technology for soils properties, such as water, fertility and photoelectricity. The Proximal Soil Sensing (PSS) technology which could obtain real-time soil information from close by (say within 2m), or within, the soil body by utilizing field-based sensors provides a possibility to realize this aim. Overall, Proximal Soil Sensing with visible and near-infrared (vis-NIR) spectroscopy is most widely used because of its numerous untraditional notable features, such as non-contact, free electronic jamming and high sensitivity, which has a huge potential of application in digital soil mapping and precision agriculture etc. Laboratory-based vis-NIR spectroscopic measurement overcomes the disadvantages in the traditional laboratory chemical analysis such as time and cost consuming, and has been widely accepted nowadays. However, the process of sample collection, transportation and preparation for laboratory-based spectral measurement weakens the advantages of vis-NIR spectroscopy technology, while field-based vis-NIR spectroscopy technology could conduct rapid spectral measurement in situ in the field with soil structure undamaged. It is a relatively simple, high-efficient and labor-saving method, and can acquire high-density soil spatial information when moving across the field, which meets the needs for precision agricultural developments. Meanwhile, different to laboratory condition, there are a lot of environment factors exist in the field, such as soil water, surface conditions and ambient light and so on, which might affect the spectra, mask or partly mask the characteristics of some soil properties, increasing the difficulty in extracting the effective information, and reducing the prediction accuracy of soil properties. To offer real-time and accurate soil information for precision agriculture, improving the prediction accuracy of soil properties with field vis-NIR spectra has become a hot but also difficult research point in recent years in soil spectroscopy technology and related fields.In the present study, nine paddy fields from close vicinity to six cities in Zhejiang province, China, were chosen as study area. The water in the paddy fields was drained and left to dry before sensing and sampling. A total of 104 sensing sites were randomly chosen in the nine paddy fields. We recorded the spectra of the soil by proximal in situ and stationary vis-NIR sensing and also sampled them for measurements using traditional chemical methods for seven important soil properties including soil organic carbon (OC), organic matter (OM), total nitrogen (TN), available nitrogen (AN), available phosphorus (AP), available potassium (AK) and pH. The feasibility of predicting seven soil properties using field spectra was studied, and different chemometric methods were used to increase their prediction accuracies, the results of which was used to provide technical support for precision agriculture and auxiliary decision making guidance. The main results are as follows:(1) Analysis of the main environmental factors which affect the field vis-NIR spectraField-based spectroscopic measurement improves the efficiency of soil data collection by avoiding tedious sampling and preparation procedures (i.e. drying, grinding and sieving). However, the environmental factors such as soil moisture, texture, the shadow of the microcapsules in soils, the condition of the soil surface, the heterogeneous of soil properties distribution, macro void, residuals such as plant root, stones and dust, temperature and ambient light and so on, might affect the prediction accuracy of soil properties. In this study, the field-based spectroscopic measurement was firstly performed on the 104 sensing sites, after which, the samples were collected and taken back to laboratory, air-dried, ground and sieved to 2 mm in particle size to measure the spectra again, obtaining the laboratory spectra. In the field-based spectroscopic measurement, the high density contact probe was used to create the artificial dark room condition in the field to prevent the ambient light. The probe has a 2 cm diameter window at the front, which is small enough and convenient to choose the relatively smooth soil surface where there are no macro voids, plant roots or stones. In this study, the main environmental factors which affect the field vis-NIR spectra and the mechanism were studied based on the spectral characteristics. The continuum removal was used to enlarge the spectral characteristics and the wavelength specific t-test was used to compare the field spectra with laboratory spectra. The comparisons were carried out between field spectra and laboratory spectra and their continuum removed spectra, respectively. The results show that the difference between the field and laboratory spectra were mainly at the absorption band representative of soil water at around 1450 and 1940 nm. The water absorption valleys were much larger than laboratory ones both in width and depth, which reflects the permanently waterlogged condition of paddy soils.(2) The feasibility study of using field vis-NIR spectra for the prediction of important soil propertiesThe overtone and combination vibration of different molecules of soil constitution occurred in the vis-NIR range is the theoretical basis of using vis-NIR spectroscopy to predict the soil properties. Due to the existence of soil moisture, surface soil condition and other environmental factors, the spectral characteristics of some soil properties might be masked or partly masked, increasing the difficulty to extract the effective soil information from the field spectra. And it becomes more obvious for paddy soils because of their permanently waterlogged condition, especially from the impact of soil moisture. This study analyzes the feasibility to use partial least squares regression (PLSR) algorithm and least squares support vector machine (LS-SVM) algorithm to predict seven important soil properties(OC, OM, TN, AN, AP, AK and pH) with the in situ field vis-NIR spectra, and the results were compared to that of laboratory spectra. The results found that soil TN, AN, OC, OM and pH can be quantitatively predicted while AP and AK cannot. Successful prediction of TN, AN, OC and OM are mainly due to the direct spectral response of them in the vis-NIR region, while AP and AK is unpredictable because of no direct spectral response. Successful prediction of pH may be due to its relation to wavelengths indicating minerals. Compared with the prediction with laboratory spectra, the prediction accuracy of soil properties with field spectra is lower no matter using PLSR or LS-SVM algorithm.(3) The comparison of using linear and non-linear multivariate calibration methods to predict with field spectraThe overtone and combination bands of the molecules of soil constitution in vis-NIR are usually broad, relatively weak and overlapped. It is difficult to distinguish their spectral characteristics visually. Molecules do not behave totally harmonically, resulting in the location of which is often slightly shifted from the exact expected location. It is thus often difficult to analyze vis-NIR spectra using conventional methods. With the development of chemometric methods, more and more multivariate data mining techniques are applied to the extraction of spectral characteristics bands of soil properties from spectra and the calibration of soil properties with vis-NIR spectra. In this study, seven important soil properties were predicted by the in situ field vis-NIR spectroscopy using linear PLSR algorithm and nonlinear LS-SVM algorithm respectively. Compared with linear PLSR algorithm, there is of great improvement of prediction accuracy on OC, OM, TN and pH using LS-SVM method, except that the prediction accuracy of the AN does not improve significantly. Neither of the two algorithms can predict AP and AK quantitatively. The results concluded the use of nonlinear LS-SVM algorithm in the prediction of soil properties with the in situ field spectra in the vis-NIR range.(4) Remove the effects of soil water from in situ field measured spectraFrom the previous study conclusions, it is clear that compared with laboratory-based measurements, field-based spectroscopic measurement improves the efficiency of soil data collection by avoiding tedious sampling and preparation procedures (i.e. drying, grinding and sieving), which is labor-, money-and time-efficient. However, the existence of environmental factors especially soil moisture in this case makes it difficult to extract the effective soil information from the field spectra, decreasing the prediction accuracy of soil properties using field spectra. It is thus of great importance to find a chemometric method that removes the effects of soil moisture from the field spectra. To date, there are not many algorithms related existing. In this study, the external parameter orthognolization (EPO), direct standardization (DS) and piecewise direct standardization (PDS) algorithm were used to remove the effects of soil moisture and other environmental factors from the field spectra. Among them, the EPO algorithm projects the soil spectra orthogonal to the space of soil moisture, the effects of which were thus removed. The DS algorithm analyses the difference between the field and laboratory spectra and builds the relations between them to calculate the transfer parameters. The environmental effects can thus be removed by multiplying the transfer parameters with field spectra. PDS is an improved version of DS. Similar to DS, PDS relates the absorbance of selected standard’ spectra recorded in the laboratory to their field spectra using the same model structure. But it does so more locally by using only neighbouring wavelengths that are within a small window of the field spectra to fit each spectral wavelength on the laboratory spectra, instead of the whole wavelength region used in the DS. The results show that all the three methods, i.e. EPO, DS and PDS can remove the soil water effects from the field spectra and improve their prediction on OC. With the non-processed field spectra, the soil OC was only quantitative predicted (1.4<RPD< 2.0) while using the processed field spectra with the previous three methods, the calibration of OC with high accuracy (RPD> 2.0) can be obtained. However, each of the three methods has its advantages and disadvantages. The premise to perform all three methods was to select a subset of samples from the whole dataset as transfer samples, and all transfer samples need to be measured both in situ in the field and then air-dried, ground and sieved to re-measure the spectra in the laboratory. Only in this way the relations between the field and laboratory spectra can be developed and the transfer parameters for the three methods can be calculated. The differences among the three methods are, The’local’character and multivariate nature of the PDS method enable the local rank of each window in PDS was smaller than the other two methods, which means the number of transfer samples can be smaller. Smaller number of transfer samples brings a save of labor, time and also cost-efficient. But the field spectra need to be first derivatived before performing the PDS algorithm for the dataset used in the study. It is not required for DS and EPO, though. We also compared the prediction accuracy of OC using field spectra processed with three methods, and the result shows that the DS algorithm performs best.(5) Using the Chinese Soil Spectral Database to predict soil OC with field vis-NIR spectraThere are many soil types included in the Chinese Soil Spectral Database (CSSD). The soil OC calibration model based on the CSSD spectra can be generalized to predict with local regional or field samples. However, this model is not suitable to predict with in situ field vis-NIR spectra because the CSSD spectra is measured in the laboratory. Three algorithms, EPO, DS and PDS were used in this study to remove the moisture effect from the 104 field spectra. Then the CSSD calibration model was used to predict the soil OC with the processed field spectra. The result shows that by using the three algorithms, the accuracy of soil OC prediction with field spectra was greatly improved from unpredictable (RPD=0.23) with the original field spectra to roughly quantitatively estimated (for EPO and PDS:RPD>1.40), or even predicted with high accuracy (for DS:RPD=2.06). The prediction with the DS-transferred field spectra is closed to that with the laboratory spectra (RPD=2.11). Based on this, the study offered a suggestion on the procedures for practical use of the CSSD to predict the soil OC with field vis-NIR spectra. Using the procedures suggested, the soil OC can be quantitatively predicted rapidly with only a subset of soils sampled and taken back to laboratory and none of the samples measured their properties using the tranditional chemical methods. Meanwhile, spiking method was used in this study by adding a subset of local field samples into the CSSD, a new calibration of OC was built with the combined samples to predict with the field spectra of the rest local samples. Compared with the traditional prediction (i.e. all the soil samples in the local field were divided into calibration and prediction set), spiking algorithm only needs to measure soil OC of a few samples (n=15-25).As the object of this study, the paddy soil was characterized. However, the flooded soil condition in paddy fields makes it difficult to perform soil sampling and analysis. The best time to sample is during the later stages of cultivation, before harvesting, when water has been drained and the soil is drying. It is difficult to acquire soil information in this short window of opportunity using conventional soil sampling and laboratory analysis while proximal soil sensing (field-based) with vis-NIR spectroscopy provides a potential solution to this problem. However, compared with the irrigated soils, the environment effects, especially the impact of soil moisture, become more obvious for paddy soils because of their permanently waterlogged condition. It thus appears more urgent to remove the effects of environment and improve the prediction accuracy of soil properties of paddy soils. Based on the research content, this study has basically arrived the expected research goal, and new progress has been made in the following aspects:(1) There have been many researches on the use of different linear and non-linear data mining algorithms to predict soil properties with laboratory vis-NIR spectra. However, few researches used non-linear algorithm to predict with field spectra. The non-linear LS-SVM algorithm was used in this research to predict soil properties, which has inhanced ability of extracting useful information from the field spectra, and the prediction accuracy were thus increased. The research provides algorithm support for the quantitative prediction of soil properties with field spectra.(2) How to remove the environmental effects from the field spectra in order to improve the prediction accuracy of soil properties has been the hot but difficult topic in the soil spectroscopic stud, however, related research is very rare. The DS and PDS algorithm were put forwarded in this study, to remove the effects of soil moisture from the field spectra of paddy soil, as well as the existed EPO algorithm. The field spectra processed by each of the three algorithms can be predicted with soil properties successfully. This study provides algorithm support for quantitative prediciton of soil properties with field spectra.(3) This study used the CSSD OC calibrated model, and the algorithms which were used to remove the environment effects from field spectra, to obtain soil OC content rapidly in the local field or region. The procedures for practical use of the CSSD to predict soil OC with field spectra were also given in the study, which offers a way for precision agriculture to rapidly obtain soil information. |
PDF Full Text Request