Predicting Soil Total Nitrogen Based On The Soil Visible And Near-infrared Spectral Library

Nowadays, precision agriculture is at the leading place in the development of agriculture in China. As one of the important research directions of precision agriculture, getting real-time and accurate soil information such as total nitrogen is the key basis for the agricultural production decision. The diffuse reflectance spectroscopy (DRS) techniques are rapid, relatively cheap and more efficient for obtain data of soil information than conventional laboratory analysis, which can not satisfy the demands of improving economic returns and preserving the ecological environment of precision agriculture. The prediction of soil total nitrogen by DRS has the great application value and the broad application prospect for guiding N fertilization in the field, by which can reduce the pollution to the environment effectively because of the excessive fertilizer application, protect the agricultural ecological environment system in China via sustainable healthy development. However, many of the researches is based on the local scale or the small regional soil samples, regardless of the diversity of soil genesis and mineralogical backgrounds in different place and the differences in soil spectral characteristics, even the same group of soil also significantly relates to the development from a particular parent material. Although these researches successfully predicted soil total nitrogen by DRS, we are not sure if these prediction models can be use to predict other regional soil samples. The universality and stability of these models haven’t been well validated. Therefore, construction of national soil spectral libraries has been sugested to validate and facilitate the wider use of DRS.In this study,1661 soil samples were collected from 13 provinces in China, including Zhejiang, Tibet, Xinjiang, Sichuan, Henan, Heilongjiang, Hainan and so on. The samples represent 17 soil groups of the Genetic Soil Classification of China. Based on the national scale soil spectral libraries, our research focus on the prediction ability of the Vis-NIR DRS and modelling strategies, examine the effects of subsetting by soil properties such as soil total nitrogen content and soil groups and the robustness of the model, especially examine the effects of subsetting by soil spectral characteristics and the application of the soil libraries, how to find proper model which have the universality and stability and why. The main results are as follows:(1) Spectral library modeling strategy based on soil propertiesBased on the national soil spectral libraries, this study examined the feasibility of using vis-NIR to build model and search the influence factors of total nitrogen in soil spectral modeling accuracy by using global linear model PLSR and non-linear model SVM. A comparison of the model accuration between PLSR and SVM has been studied.The results show that the soil total nitrogen-visible near infrared spectral library has potential applications,but global modelling strategy can not be used in soil library, other strategies need to be explored. The soil library is divided into different groups by soil total nitrogen content and soil groups separately using PLSR and SVM model. The results show that two kinds of modeling strategies do not work well in the diversity and heterogeneity of soil spectra in a large scale,the accuracy of all the models are low; the strategy of subsetting by soil total nitrogen content indicates that model accuracy is closely related to the range of total nitrogen content, which can not be predicted if the content below a certain value; the strategy of subsetting by soil groups indicates that model accuracy is greatly influenced by the diversity of the regional areas and parent material. The accuracy of the homogeneity model is significantly higher than different soil groups.(2) Spectral library modeling strategy based on soil spectral characteristicsThis study further explores the application of spectral library based on soil spectral characteristics and achieves success. According to the research results above, the prediction of soil total nitrogen use global partial least squares regression (PLSR), locally weighted regression (LWR) and fuzzy K-means clustering combined with PLSR (FKMC-PLSR). Another 104 paddy soil samples collected from Zhejiang Province were used to validate the established models. Results showed that when predicting soil total nitrogen from a large dataset, global PLSR underestimated high values of TN and overestimated low values of TN, generating an overall low prediction accuracy (Rp12=0.64, RPDP1=1.4). By contrast, LWR (RP22=0.76, RPDP2=2.1) and FKMC-PLSR (RP32=0.82, RPDP3=2.4) performed better than global PLSR. It was suggested that the results can provide useful information for establishing robust and universal models for soil TN prediction using large soil spectral libraries.