基于光谱分类的土壤盐分含量预测

Prediction on Soil Salt Content Based on Spectral Classifi cation

基于相似土壤组分和光谱特征,利用土壤光谱反射率数据和曲线特征来进行土壤光谱分类,同时充分挖掘有效信息是光谱分析的重要应用方向之一。借助模糊k-均值聚类方法将土壤光谱数据分成四个类别(分类前先将原始光谱进行范围归一化处理),比较分析了不同类型土壤在光谱分类前后的高光谱特征,然后利用Kennard-Stone法将各类别样本划分为建模集和预测集,将预处理后的建模集光谱数据作为输入量,采用偏最小二乘回归法(PLSR)方法建立全局和各自类别的盐分预测模型。结果表明:光谱分类建模较按土壤系统分类建模和全局建模的精度有明显提高,其预测模型总体的预测决定系数RP2、预测均方根误差RMSEP、相对分析误差RPD和RPIQ(样本观测值三四分位数Q3与一四分位数Q1之差与RMSEP的比值)四个指标分别从0.664、1.219、1.733和1.461提高至0.818、1.132、2.356和2.422,其中RPD提高幅度达23.13%,四个类别所建模型RPD均大于2.0,可以对土壤含盐量进行较为精确的定量研究。研究结果为利用大样本光谱数据建立大尺度区域的盐分等土壤属性预测模型提供一种新的思路和方法。

【Objective】Soil salinization is an issue related to the strategy of sustainable development and improvement of environment quality in arid regions. Soil salinization and irrigation-triggered soil secondary salinization are problems that severely restrict further development of the agriculture in Xinjiang and also the major factors that affect stability of the oasis ecological environment. Therefore,it is of profound significance to sustainable development of the agriculture in Xinjiang and in China as well to manage and ameliorate salinized arable lands and readjust and optimize planting structure in the region,moreover,the information of soil salinization is the fundamental basis for formulation and implementation of these policies. How to acquire the dynamic information of large areas of salinized soil quickly,accurately and inexpensively and thereby formulate rational policies and strategies for management and amelioration of salinized soils are new themes of the study of agricultural science. 【Method】Hyperspectral remote sensing is a relatively ideal means of monitoring soil. A series of tests and experiments of monitoring soil salinization with near-ground hyperspectra were performed,thus providing a theoretical basis and technical support for application of the technique of hyperspectral remote sensing to monitoring soil salinization. Based on similar soil components generating similar spectral characteristics,soil spectra is to be classified and at the same time effective information to be fully mined by making use of the data and curve characteristics of soil spectral reflectance,which is one of the importance fields for application of spectral analysis. To that end a total of three hundred and thirtynine topsoil samples（0~20cm）were collected in Wensu,Baicheng,Awati,Xinhe and Hetian counties of Xinjiang,covering anthropogenic-alluvial soil,saline soil,fluvo-aquic soil,paddy soil. The soil samples were then air dried and ground to pass a 2 nm sieve for determination of salt content with the residue drying method. Hyperspectral reflectivities of the samples were measured separately using the American Field Spec Pro FR spectrograph（American Analytical Spectral Devices Company）,and spectral curves obtained of the samples were pre-treated by erasing the edge bands,i.e. 350~399 and 2 401~2 500 nm on the two ends of a curve and keeping the reflection spectral data of the band 400~2 400 nm for successive research. Clustering of fuzzy k-means is a commonly used unsupervised clustering method,of which the basic idea is to sort a data set into k categories so as to minimize the iteration of target functions and the advantage is the capability of yielding an analysis index for setting an optimal number of categories. 【Result】By the means of clustering fuzzy k-means,3 indices,i.e. fuzziness performance index（FPI）,modified partition entropy（MPE）and clustering independence index（S）（all being the lower the best）were obtained,and the spectral data of the soil samples were sorted into four categories（before sorting,the original spectral range was normalized as pretreatment）. Hyperspectral characteristics of the soils different in soil type before and after the sorting were compared and analyzed,and then the soil samples of each category were divided into modeling dataset and prediction dataset at a ratio of 2∶1 using the Kennard-Stone algorithm. The preprocessed spectral data of the modeling dataset were used as input,and then a global and category-based salinity prediction models were built up with the PLSR method. The prediction dataset was used to validate the models and evaluate the models in terms of accuracy and stability. Results demonstrate that after the spectral data were classified with the fuzzy k-means clustering method,the Kennard-Stone algorithm was used to further classify the data for modeling,which was significantly higher than the global and category-based models in precision. The overall prediction model determination coefficient（RP2）,root mean squares error（RMSEP）,relative percent deviation（RPD）and ratio of performance to IQ（RPIQ）of the prediction model was increased from 0.664,1.219,1.733 and 1.461 to 0.818,1.132,2.356 and 2.422,respectively,among which RPD was increased by up to 23.13% and the PRDs of the all the four category-based models were higher than 2.0. 【Conclusion】The findings demonstrate that it is feasible to predict soil salt contents quantitatively and quite accurately,and that the study has opened up a new way of thought a new method for using large sample spectral data to establish soil attribute prediction models including that for sail salt content,over a large scale area.