Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Comparison of additive and isometric log-ratio transformations combined with machine learning and regression kriging models for mapping soil particle size fractions
Geoderma ( IF 6.1 ) Pub Date : 2020-04-01 , DOI: 10.1016/j.geoderma.2020.114214 Zong Wang , Wenjiao Shi , Wei Zhou , Xiaoyan Li , Tianxiang Yue
Geoderma ( IF 6.1 ) Pub Date : 2020-04-01 , DOI: 10.1016/j.geoderma.2020.114214 Zong Wang , Wenjiao Shi , Wei Zhou , Xiaoyan Li , Tianxiang Yue
Abstract Digital soil mapping approaches relating to the soil particle size fractions (psf) face the challenge around how to establish the statistical or geostatistical models from large sets of environmental variables, especially in a situation with sparse soil profile data. Recently, many machine learning (ML) models have sprung up with advantages over statistical models. However, few studies focused on the comprehensive comparative analyses between ML and geostatistical models in the soil psf mapping. And the exploration of optimal combination of data transformation and model simulation was even less. Therefore, two transformed methods such as additive log-ratio (ALR) and isometric log-ratio (ILR) transformations combine with two ML models such as boosted regression tree (BRT), random forest (RF) and a classic geostatistical model of regression kriging (RK) were implemented to map soil psf in the Heihe River basin, China. A total of 640 samples and thirteen scorpan factors were collected and used for the comprehensive comparative analysis. Results showed that the scorpan factors such as temperature, precipitation, elevation, soil type, soil organic carbon, vegetation types and normalized difference vegetation index had important impacts on the soil psf mapping. ILR transformation was better than ALR transformation with advantage of improving stability of data distributions and ML models could also improve the mapping performance in comparison with RK models for better handling candidate factors. For these ML models, the RF models had better accuracy performance than the BRT models. In contrast, ILR transformation combined with RF model (ILR_RF) had the best performance, with the lowest root mean square error values (sand, 15.35%; silt, 14.20%; and clay, 6.66%), Aitchison distance value (0.86), standardized residual sum of squares value (0.60), and the highest concordance correlation coefficient value (0.73) and coefficient of determination value (56.69%) for clay content. In addition, ILR_RF had a relatively higher right ratio of soil texture type (68.44%) and better predict performance for most soil texture types. The predicted maps generated from ILR_RF presented more reasonable and smoother transitions. In the future, more ML models should be explored and more variables related to soil psf should be introduced into the models to improve the predictive performance.
更新日期：2020-04-01
京公网安备 11010802027423号