📄 earth_science

Prediction of Soil Organic Carbon Using Machine Learning Algorithm in Long-Term Ecological Monitoring Plots of Southern Dry Mixed Deciduous Forests, Telangana, India

This study evaluates five machine learning algorithms to predict soil organic carbon in Telangana's dry deciduous forests, identifying random forest as the most effective model and soil moisture as the most critical predictive variable.

Original authors: Ruby Patel, Bharati Patel, Sravanthi Palle, Jaison M

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruby Patel, Bharati Patel, Sravanthi Palle, Jaison M

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The soil beneath our feet is far more than just dirt; it is a living, breathing reservoir that holds more carbon than all the world's forests and the atmosphere combined. This hidden carbon, stored within organic matter like decaying leaves and roots, plays a critical role in regulating the planet's climate. When this organic matter breaks down, it releases carbon back into the air, but when it is preserved, it acts as a powerful tool against climate change. Understanding how much carbon is stored in the ground, and what factors help keep it there, is essential for managing forests sustainably. However, measuring this carbon is difficult because soil conditions vary wildly from one spot to another, influenced by everything from rainfall to the specific nutrients available to plants. Scientists have long sought better ways to predict these hidden reserves without having to dig up and test every single square meter of a forest.

In the southern dry mixed deciduous forests of Telangana, India, a team of researchers set out to solve this puzzle by testing whether modern computer programs could learn to predict soil carbon levels more accurately than traditional methods. They focused on a specific ten-hectare plot of forest, a permanent monitoring site where they had already established a detailed grid for observation. To build their models, the team collected hundreds of soil samples from various depths across this plot. They analyzed each sample for key characteristics: how acidic or neutral the soil was, how much moisture it held, and the levels of essential plant nutrients like nitrogen, phosphorus, and potassium. With this rich dataset in hand, they fed the information into five different types of machine learning algorithms. These are computer programs designed to find patterns in data, ranging from simple statistical tools to complex systems that mimic the way the human brain processes information. The goal was to see which program could best guess the amount of soil organic carbon based solely on the other soil properties.

The results revealed a clear hierarchy in how well these digital tools performed. The most successful approach was a method called Random Forest, a technique that builds many small decision trees and combines their answers to reach a conclusion. This model outperformed all others, accurately predicting the carbon content with a high degree of reliability. It was followed closely by another advanced method known as Extreme Gradient Boosting and a system modeled after neural networks. In contrast, a simpler traditional method called Multiple Linear Regression, which assumes a straight-line relationship between variables, performed moderately well but could not capture the full complexity of the soil. The least effective tool was the Support Vector Machine, a sophisticated algorithm that struggled significantly with this specific dataset, producing predictions that were far less accurate and much more uncertain than the others.

A deeper look into why these models worked or failed highlighted a crucial factor: soil moisture. Across the most successful models, the amount of water in the soil emerged as the single most important clue for predicting how much carbon was stored. While the availability of nutrients like nitrogen and phosphorus also played a strong role, the moisture content was the dominant driver. The researchers found that the soil in this dry deciduous forest was generally near neutral in acidity, with moderate levels of nitrogen and potassium, but relatively low levels of phosphorus. The carbon content itself varied significantly across the plot, ranging from very low to moderately high levels, reflecting the complex and uneven nature of the forest floor. The study confirmed that while nutrients are vital, the physical state of the soil, particularly its ability to hold water, is the primary key to understanding carbon storage in these dry environments.

This work demonstrates that for forest managers and scientists in similar dry regions, the best way to estimate soil carbon is not through simple averages or older statistical tricks, but through advanced computer models that can handle complex, non-linear relationships. The study suggests that by focusing on soil moisture and nutrient levels, and by using robust machine learning tools like Random Forest, it is possible to map soil health with greater precision. This precision is vital for understanding how these forests function as carbon sinks and for developing strategies to protect them. While some of the more complex algorithms struggled, the success of the tree-based models offers a practical path forward, proving that the right digital tools can unlock the secrets hidden in the earth beneath our feet.

Technical Summary: Prediction of Soil Organic Carbon Using Machine Learning in Telangana's Dry Deciduous Forests

Problem Statement and Context
Soil organic carbon (SOC) is a critical component of the global carbon cycle, with forest soils storing significantly more carbon than terrestrial vegetation and the atmosphere. However, predicting SOC variability in complex forest ecosystems remains challenging due to the non-linear interactions between SOC and environmental drivers such as soil physical properties, moisture, and nutrient status. While traditional statistical methods like Multiple Linear Regression (MLR) have been widely used, they often fail to capture the complex, non-linear relationships inherent in heterogeneous forest soils. Furthermore, there is a notable research gap regarding the use of specific soil nutrients (Nitrogen, Phosphorus, Potassium) and soil moisture as covariates for SOC prediction in dry deciduous forests, particularly within the Indian context. This study addresses the need to identify the most effective machine learning (ML) algorithm and the most influential soil variables for predicting SOC in the southern dry mixed deciduous forests of Telangana, India.

Methodology
The study was conducted in a 10-hectare permanent long-term ecological monitoring plot established in the southern dry mixed deciduous forest of Narsapur, Telangana. The plot was systematically divided into subplots and quadrats to ensure representative sampling.

Data Collection: A total of 450 soil samples were collected from five depth intervals (0–10 cm to 50–100 cm) across nine sampling points per hectare.
Laboratory Analysis: Samples were analyzed for six key parameters: Soil pH, Soil Organic Carbon (SOC), Available Nitrogen (N), Available Phosphorus (P), Available Potassium (K), and Soil Moisture.
Modeling Approach: Five machine learning algorithms were employed to model SOC variability using the measured soil properties as predictors:
1. Multiple Linear Regression (MLR)
2. Random Forest (RF)
3. Neural Networks (NN)
4. Extreme Gradient Boosting (XGBoost)
5. Support Vector Machines (SVM)
Evaluation: The dataset was split into training (60%), validation (20%), and testing (20%) sets. Model performance was assessed using cross-validation statistics, specifically the Coefficient of Determination ( $R^2$ ), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and uncertainty analysis. Feature importance was also evaluated to determine the influence of each soil variable.

Key Results

Soil Characteristics: The study area exhibited near-neutral soil pH (mean 7.03) and moderate fertility for Nitrogen and Potassium, while Available Phosphorus was comparatively low. SOC content ranged from 0.17% to 2.45% with a mean of 0.78%, showing high spatial variability (Coefficient of Variation = 51.06%).
Correlations: SOC showed a very weak negative correlation with pH but moderate positive correlations with soil moisture and strong positive correlations with available nutrients (N, P, and K).
Model Performance:
- Random Forest (RF) emerged as the best-performing model, achieving the highest $R^2$ (0.597) and the lowest error metrics (RMSE = 0.257, MAE = 0.135).
- XGBoost and Neural Networks (NN) followed with moderate performance ( $R^2$ of 0.586 and 0.571, respectively).
- Multiple Linear Regression (MLR) showed moderate performance ( $R^2$ = 0.433).
- Support Vector Machine (SVM) exhibited the poorest predictive capability, with the lowest $R^2$ (0.275) and the highest uncertainty.
Feature Importance: Across the ensemble models (RF and XGBoost), soil moisture was identified as the most influential predictor for SOC content. In the NN model, available phosphorus was the top predictor, while SVM highlighted available nitrogen. On average, soil moisture was the most significant variable across all algorithms.

Significance and Claims
The paper asserts that soil moisture and available nutrients (N, P, K) are critical covariates for accurately predicting SOC in dry deciduous forest soils. The study demonstrates that ensemble machine learning methods, particularly Random Forest, outperform traditional linear regression and other algorithms like SVM in capturing the non-linear dynamics of soil carbon in these ecosystems. The authors conclude that while RF provides the most robust predictions with the lowest uncertainty, the SVM model requires further optimization (such as hyperparameter tuning) to be effective for this specific application. The findings support the integration of soil physical and chemical properties into predictive models to better understand carbon sequestration potential and inform sustainable forest land management strategies in semi-arid regions.

More like this