科研之友
微信
新浪微博
Facebook
分享链接
摘要
Nonparametric techniques are of interest for soil and environmental sciences because they enable to effectively predict soil data from basic soil properties without the need of a priori selected equations. We applied two nonparametric techniques, k-nearest neighbor (k-NN) and boosted regression tree (BRT), on data of an existing soil survey database to predict topsoil bulk density (BD) of a tropical mountain forest (Nyungwe) in Rwanda. Soil particle size distribution, organic carbon (OC) content, pH, and cation exchange capacity (CEC) were used as input data and soil depth (topsoil or subsoil), land use (forest or nonforest) and soil horizon notation were tested as possible grouping and limiting factors for model training. The k-NN and BRT techniques showed a comparable performance and predicted BD for an independent data set equally well as the Adams-Minasny-Hartemink and Adams-Rawls-Brakensiek pedotransfer PTF) but significantly better than the Adams-De Vos-et al. PTF developed for tropical nonforest soils, nontropical (United States) nonforest soils and soils in the tropics, and nontropical (Belgian) forest soils, respectively. Adding particle size distribution, pH, and cation exchange capacity (CEC) as input variables or grouping samples by different limiting factors did not enhance the predictive capacity significantly compared to a model that used OC content as the sole input. Thus, it appears that robust soil OC data is essential for successfully predicting soil BD in African tropical forests, which in turn is an essential parameter for soil fertility assessments and drives many biogeochemical models. Despite this, OC levels still remain largely unknown for such areas. High throughput analyses based on infrared (IR) spectroscopy might help in collecting OC data for data poor areas.
- 出版日期2012-7
