Multiple proximal soil sensors provide more useful and complementary information than a single sensor does for improving the prediction of soil properties. In this study, two data fusion methods (i.e., sample data fusion and sensor data fusion) were performed using a Fourier transform near-infrared (NIR) spectrometer and a portable X-ray fluorescence (PXRF) analyzer for predicting the contents of soil clay, silt, and sand in three sets of soil sample data, which were collected in a medium-scale area and two small-scale areas, respectively. To verify the performance of these methods and confirm related theoretical analysis, calibration models on individual sample and sensor data sets, fused sample data sets, and fused sample and sensor data sets were constructed for three soil textural fractions. First, we applied stepwise multiple linear regression to PXRF measurements and partial least-squares regression to NIR spectral data in the three different sampling areas. Then, we used the sample data set from the medium-scale area as reference and combined it with other two sample data sets to calibrate the models. Finally, we further fused the principal components of NIR spectral data from principal component analysis with the PXRF measurements to improve the accuracy of soil texture prediction. It was found that the models based on fused data could significantly improve the accuracy of soil texture prediction compared with those based on individual sample or sensor data sets. The best model for the prediction of clay and sand contents was calibrated on fused PXRF measured data and NIR principal components from different sampling areas. Whereas geographical locations and soil types might impact the accuracy and fitness of the models, data fusion through multiple sensors and different sampling areas showed great promise as a technique for rapidly assessing soil textural fractions. Further evaluation of the data fusion methods suggested here may provide a practical technique for future applications in soil research.
1Department of Resource and Environmental Information Science, College of Resources and Environment, Huazhong Agricultural University, Wuhan, China.
2State Key Laboratory of Soil and Sustainable Agriculture, Institute of Soil Science, Chinese Academy of Sciences, Nanjing, China.
3Department of Geography, University of Connecticut, Storrs, Connecticut, USA.
Address for correspondence: Dr. Shan-qin Wang, Department of Resource and Environmental Information Science, College of Resources and Environment, Huazhong Agricultural University, Wuhan 430070, China. E-mail: email@example.com
Financial Disclosures/Conflicts of Interest: This study was partially supported by the National Natural Science Foundation of China (Grant No. 40801082), the State Key Laboratory of Soil and Sustainable Agriculture (Grant No. 812000028), the China Scholarship Council (Grant No. 2011676506), and the National High Technology Research and Development Program of China (863 program, Grant No. 2013AA102401-3).
Received June 13, 2013.
Accepted for publication December 12, 2013.