From Rangelands to Cropland, Land-Use Change and Its Impact on Soil Organic Carbon Variables in a Peruvian Andean...

Area of Science:

• Soil Science
• Environmental Science
• Machine Learning Applications

Background:

• Andean highland soils store substantial soil organic carbon (SOC).
• Understanding SOC accumulation and persistence processes requires further investigation.
• The Central Andean Highlands of Peru, characterized by grasslands and "bofedales" (wetlands), are vital ecosystems for SOC storage.

Purpose of the Study:

• To model SOC, refractory SOC (RSOC), and SOC's ¹³C isotope composition (δ¹³C_SOC) in the Central Andean Highlands of Peru.
• To identify the most effective machine learning (ML) algorithms for predicting these SOC variables.
• To determine the key environmental predictors influencing SOC dynamics in the region.

Main Methods:

• Collected 198 soil samples (0.3 m depth) to assess SOC, RSOC, and δ¹³C_SOC.
• Employed four ML algorithms: random forest (RF), support vector machine (SVM), artificial neural networks (ANNs), and eXtreme gradient boosting (XGB).
• Utilized remote sensing data, land-use/land-cover (LULC), climate, topography, and soil physical-chemical properties as predictors.

Main Results:

• Random Forest (RF) was the best algorithm for predicting SOC and δ¹³C_SOC; ANNs excelled at modeling RSOC.
• "Bofedales" exhibited 2-3 times higher SOC and RSOC, and more depleted δ¹³C_SOC compared to other LULC types.
• Land-use/land-cover (LULC) and near-infrared vegetation indices were critical for SOC and δ¹³C_SOC prediction, while climate indices were more important for RSOC.

Conclusions:

• Machine learning, particularly RF, shows significant potential for mapping SOC and its fractions in Andean highland soils.
• "Bofedales" are critical "hotspots" for SOC storage, highlighting their ecological importance.
• Understanding the interplay between LULC, vegetation, climate, and soil properties is essential for managing SOC in these sensitive ecosystems.