基于高光谱与协同克里金的土壤耕作层含水率反演

为研究土壤耕作层(0~40 cm)含水率的空间分布情况,利用EO-1的Hyperion传感器高光谱数据,对研究区域(106°20′~109°19′E,40°19′~41°18′N)的表层(0~10 cm)含水率进行定量反演,并利用表层含水率反演结果作为协同克里金插值的协变量,同时利用103个采样点实测的耕作层含水率作为主变量,进行协同克里金插值。结果表明:通过特征指数法提取水分反演的敏感波段集中在1 295~2 224 nm波长区间,特征指数法模型校正的相关系数r0.5但模型验证的精度较低(r0.2);通过偏最小二乘法建模,模型校正的r0.8,模型验证的r0.5,效果较好;运用协同克里金插值时,将反演的表层含水率作为协变量,可以弥补主变量耕作层含水率样本点少,变异函数欠稳定的缺点,同时,所提取理论模型的块金值(C0)与基台值(C0+C)的比值均25%,随机因素比例小,模型稳健。此外,协同克里金插值方法与利用表层与耕作层含水率线性拟合进行预测相比,能够有效提高预测精度,决定系数r2和Nash效率系数(nash-sutcliffe modelling efficiency,NSE)分别提高72.6%和89.9%,因此,将高光谱反演与协同克里金方法相结合,可以综合两者优势,节约采样成本,提高预测效率。

Modelling soil water content in plow layer using co-kriging interpolation based on hyperspectral data

Abstract: Understanding the distribution of soil water content of plow layer (0-40 cm) is important for agriculture water management for plant growth. In our study, Hyperion data (EO-1, USGS) was firstly used to inverse topsoil water content and then the measured water content values of 0-40 cm depth were used to calculate the average water content of plow layer. Both data can be used together to obtain a distribution of regional plow layer water content. By use of such method, a study was carried out by taking soil samples in a 64 hm2 area located in Hetao Irrigation District, Inner Mongolia, China in late April. The soil samples were arranged in gird and the grids size were 20, 50, and 100 meters respectively. There were 136 different sampling points and 103 of them had soil samples (0-40 cm depth with 10 cm increment). The time of Hyperion data was April 11, 2013 and it was pre-processed by EVNI 5.0 software. Then derivative filter (1st) was used to remove the scattering and other disturbance. Both raw and filtered images were used to inverse the water content of topsoil using the flag index method and partial least square regression (PLS). After that, the water content of topsoil obtained by Hyperion data were used as the co - variable and the average water contents of 0-40 cm depth were used as the main variable in the co - kriging method to map the water content of plow layer (0-40 cm depth) in the study area. The results indicated that sensitive wavelength bands for topsoil water content were ranged from 1295 nm to 2224 nm when using the flag index method, and the accuracy of prediction models of the flag index method was poor (r<0.2 in the validation process). However, prediction models established by PLS method can yield higher accuracy compared to the flag index method (r>0.5 in both calibration and validation process. The co-kriging interpolation had a consideration of water content of both topsoil (0-10 cm) and plow layer (0-40 cm), and the C0/(C0+C) values in the models were all <25%, demonstrating small random variations in these interpolation models. In addition, compared with the method of linear fitting using topsoil water content and plow layer water content, the co-kriging method can increase 72.6% of r2 and 89.9% of NSE. Therefore, combined hyperspectral inversion with the co-kriging interpolation can be used to predict soil water content of plow layer effectively.