Soil Organic Carbon Variability in High-Andean Ecosystems: Bringing Together Machine Learning and Proximal Soil Sensing
Abstract
The Andes represent the largest and highest mountain range in the tropics and is considered an important reserve of biodiversity, water provision and soil organic carbon (SOC) stocks. Nevertheless, limited attention has been given to estimate these stocks due to the lack of recent soil data, the poor accessibility and the wide range of coexistent ecosystems. In addition, conventional methods to determine SOC are usually time consuming and expensive to use in large-scale studies, hindering the possibility to have an accurate SOC assessment in the region. Proximal soil sensing techniques, such as visible near infrared (VNIR) and mid infrared (MIR) spectroscopy, have proven to be useful as an alternative to conventional methods for characterizing SOC but have not been tested in Andean soils. The aim of this study was to evaluate the potential of using VNIR and MIR spectroscopy to predict SOC content in the Central Andean region, using multivariate methods. Three study areas were selected across the Peruvian Central Andes. A total of 400 topsoil samples (0-30 cm) were collected and analyzed for SOC. The VNIR and MIR reflectance of the soil samples was measured in the laboratory. Three modeling approaches: Partial least squares regression (PLSR), random forest (RF) and support vector machine (SVM) were used to predict SOC from VNIR and MIR spectra in the study areas. The data was preprocessed in order to minimize the noise and optimize the accuracy of predictions. The models, for each study area, were assessed using 10-fold cross validation. Independent validation was implemented in the whole dataset (400 observations) by splitting it into calibration (70 %) and validation (30%) sets. Overall, the results indicate potential for both VNIR and MIR spectra to predict SOC content in the Andean soils. SOC content predictions from MIR spectra outperformed those from VNIR spectra. The evaluation of model performance shows that RF and SVM provide more accurate SOC predictions compared to PLSR. These findings suggest that integrating VNIR and MIR spectroscopy with machine learning algorithms constitutes a promising approach for assessing SOC content in high-Andean ecosystems.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2017
- Bibcode:
- 2017AGUFM.B51G1886G
- Keywords:
-
- 0439 Ecosystems;
- structure and dynamics;
- BIOGEOSCIENCES;
- 0476 Plant ecology;
- BIOGEOSCIENCES;
- 1615 Biogeochemical cycles;
- processes;
- and modeling;
- GLOBAL CHANGE;
- 1813 Eco-hydrology;
- HYDROLOGY