Next Article in Journal
Modeling Short-Term Drought for SPEI in Mainland China Using the XGBoost Model
Previous Article in Journal
Modelling of Nanoparticle Number Emissions from Road Transport—An Urban Scale Emission Inventory
Previous Article in Special Issue
Assessing the Effects of Wheat Planting on Groundwater Under Climate Change: A Quantitative Adaptive Sliding Window Detection Strategy
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Assessing Climate Change Impacts on Cropland and Greenhouse Gas Emissions Using Remote Sensing and Machine Learning

1
Department of Architecture and Urban Planning, Zonguldak Vocational School, Zonguldak Bülent Ecevit University, 67600 Zonguldak, Turkey
2
Department of Geomatics Engineering, Engineering Faculty, Ondokuz Mayıs University, 55200 Samsun, Turkey
*
Author to whom correspondence should be addressed.
Atmosphere 2025, 16(4), 418; https://doi.org/10.3390/atmos16040418 (registering DOI)
Submission received: 24 February 2025 / Revised: 31 March 2025 / Accepted: 31 March 2025 / Published: 3 April 2025
(This article belongs to the Special Issue Observation of Climate Change and Cropland with Satellite Data)

Abstract

:
This study investigated the impact of grassland and cropland expansion on carbon (C) and nitrous oxide (N2O) emissions using remote sensing data and machine learning models. The research focused on agricultural land-use changes in South Sumatra from 1992 to 2018, utilizing Landsat satellite imagery and Google Earth Engine (GEE) for spatial and temporal analysis. Machine learning algorithms, including gradient boosting trees (GBT), random forest (RF), support vector machines (SVM), and classification and regression trees (CART), were employed to estimate greenhouse gas emissions based on multiple environmental parameters. These parameters include enhanced vegetation index (EVI), land surface temperature (LST), normalized difference vegetation index (NDVI), albedo, elevation, humidity, population density, precipitation, soil moisture, and wind speed. The results revealed a strong correlation between agricultural expansion and increased C and N2O emissions, with RF and GBT models demonstrating superior predictive accuracy. Specifically, GBT and RF achieved the highest R2 value (0.71, 0.59) and the lowest error metrics in modeling emissions, whereas SVM performed poorly across all cases. The study highlights the effectiveness of machine learning in quantifying emission dynamics and underscores the necessity of sustainable land management strategies to mitigate greenhouse gas emissions. By integrating remote sensing and data-driven methodologies, this research contributes to climate change mitigation policies and precision agriculture strategies aimed at balancing food security and environmental sustainability.

1. Introduction

Grasslands play a crucial role in maintaining global ecological balance by acting as carbon sinks and supporting biodiversity. Their ability to sequester carbon and regulate greenhouse gas (GHG) fluxes makes them essential for climate mitigation efforts. However, human activities, particularly agricultural expansion and land-use modifications, have significantly altered these ecosystems, affecting their capacity to store carbon and mitigate climate change impacts. Understanding these transformations and their implications for carbon sequestration and GHG emissions is imperative for developing effective environmental policies.
Grasslands are vital ecosystems that contribute significantly to global carbon storage and biodiversity conservation. They provide essential ecosystem services, including climate regulation, soil stabilization, and habitat provision [1,2]. However, these functions are increasingly threatened by agricultural intensification and land-use changes, which alter carbon sequestration dynamics and greenhouse gas (GHG) emissions [3,4]. Understanding these impacts is crucial for developing effective climate mitigation strategies.
Recent advancements in remote sensing and machine learning have revolutionized the assessment of land-use changes and their environmental consequences. These technologies offer high-precision tools for monitoring carbon sequestration processes and estimating emissions of key greenhouse gases, such as carbon (C) and nitrous oxide (N2O) [5,6]. Despite their potential, comprehensive evaluation of different machine learning methodologies in estimating emissions from grasslands remains an underexplored area of research [7].
Agriculture, while being the cornerstone of global food security, is also a major contributor to GHG emissions. The expansion of agricultural activities intensifies carbon and nitrogen fluxes, influencing climate change dynamics [8,9]. South Sumatra, with its extensive agricultural landscapes, represents a critical case study for understanding the relationship between land-use changes and greenhouse gas fluxes. However, existing research on this region remains limited [10,11]. This study aimed to bridge this gap by evaluating the predictive accuracy of various machine learning models in estimating C and N2O emissions using remote sensing data [12]. The findings provide data-driven insights to support sustainable land management and inform climate policies.
By refining the methodological approach and integrating state-of-the-art remote sensing techniques with advanced machine learning models, this research sought to enhance the accuracy of greenhouse gas emission estimations. This study contributes to a deeper understanding of land-use impacts on carbon cycling and provide a scientific basis for effective mitigation strategies [13].
Comprehending the dynamics of the terrestrial C cycle is fundamental for assessing greenhouse gas emissions and formulating climate change mitigation and adaptation policies [14,15,16,17]. Land use and land-use change (LU/LUC) constitute the second-largest anthropogenic source of atmospheric C emissions, following fossil fuel combustion [18]. The long-term conversion of forests and grasslands into croplands is known to deplete soil organic carbon (SOC) reserves, releasing carbon into the atmosphere. However, SOC levels can be replenished through effective carbon sequestration strategies [19,20,21,22]. The amount of SOC is closely associated with land management practices, soil characteristics, and climatic conditions [23,24,25,26,27].
Currently, agricultural soils suffer from a significant depletion of SOC due to a persistent imbalance between carbon inputs and outputs. SOC is essential for sustaining agricultural productivity, as it regulates various biological, chemical, and physical soil functions while also serving as a major carbon sink. Enhancing SOC levels in agroecosystems could simultaneously support multiple sustainable development goals, including climate change adaptation, mitigation, and food security [28,29].
Croplands are recognized as the largest anthropogenic source of atmospheric N2O emissions. However, the quantification of N2O emissions from croplands remains highly uncertain due to their pronounced spatial temporal variability and the complex interplay of biotic and abiotic factors governing soil N2O production [30,31,32,33,34,35]. Upland agricultural fields, although generally well-drained and oxic, can experience transient anoxic conditions due to irrigation, precipitation, or snowmelt, which can stimulate denitrification and lead to increased N2O emissions. These emissions predominantly result from enhanced microbial activity during such anaerobic conditions. While paddy fields typically emit lower levels of N2O than upland fields and, under specific conditions, can even act as weak atmospheric N2O sinks, emissions can still occur during midsummer drainage, intermittent irrigation events, and the final drainage period as well as following rainfall in fallow seasons [36,37,38,39,40,41,42,43,44].
Remote sensing technologies, particularly satellite imagery and cloud-based geospatial platforms such as Google Earth Engine (GEE), have transformed large-scale environmental monitoring efforts. These tools facilitate the processing of extensive datasets, allowing for the systematic assessment of land-use dynamics and their long-term environmental implications. This study utilized Landsat satellite data and GEE to analyze agricultural land changes in South Sumatra from 1992 to 2018, evaluating their influence on C and N2O emissions, which are intricately linked to agricultural activities in the region.
The primary objectives of this study are the following:
  • To analyze annual trends in agricultural land-use changes in South Sumatra, offering insights into the extent and rate of agricultural expansion;
  • To investigate the relationship between land-use changes and C and N2O emissions, identifying patterns and variations over time;
  • To support the development of sustainable agricultural policies through data-driven insights that balance food production with climate change mitigation.
This research is particularly significant, as it provides a quantitative understanding of how agricultural expansion influences greenhouse gas emissions. By integrating remote sensing data with machine learning techniques, this study presents an innovative methodology for estimating emissions and evaluating environmental sustainability in agricultural landscapes. The findings have the potential to inform both local and global policy frameworks aimed at promoting sustainable land-use practices while mitigating the adverse effects of climate change.
Remote sensing technology has become an essential tool for monitoring land-use changes and their environmental impacts, particularly in assessing greenhouse gas (GHG) emissions resulting from agricultural expansion. Various machine learning (ML) approaches have been employed to analyze remote sensing data, allowing for the efficient processing of large datasets and the extraction of meaningful patterns related to land-use and environmental changes [45,46]. However, despite the growing use of ML techniques in remote sensing applications, challenges remain in optimizing model accuracy, addressing spatial heterogeneity, and integrating diverse environmental parameters to improve predictive performance.
Among the widely used ML methods, random forest (RF) is an ensemble learning technique effective for both classification and regression tasks. RF consists of multiple decision trees trained on randomly selected feature subsets, utilizing a bootstrapping (bagging) technique to reduce variance and enhance model generalization [45,46]. Unlike the classical classification and regression trees (CART) model, which incorporates pruning to minimize overfitting, RF avoids pruning and instead relies on aggregating predictions from multiple trees to improve robustness. Although RF has demonstrated high accuracy in land-cover classification and emission estimations, it still faces limitations in handling imbalanced datasets and capturing complex spatial dependencies in heterogeneous landscapes [47].
Gradient boosting trees (GBT) is another powerful ML method used for classification and regression problems. In this approach, a series of weak learners (typically decision trees) are trained sequentially, with each new model correcting the errors of the previous ones. This sequential optimization enhances predictive accuracy by minimizing residual errors, making GBT particularly effective for detecting subtle patterns in remote sensing data [48]. However, GBT models are computationally expensive and prone to overfitting if hyperparameters are not carefully tuned, presenting a challenge in large-scale environmental studies.
Support vector machines (SVM) is a robust technique for both classification and regression, constructing an optimal decision boundary (hyperplane) to separate data points. By mapping input data into a higher-dimensional space, SVM can handle non-linearly separable data using kernel functions such as linear, radial basis function (RBF), and sigmoid [49,50,51]. While SVM has been successfully applied in land-use classification and environmental monitoring, its performance is highly dependent on kernel selection and computational efficiency, which can limit its scalability for large-scale remote sensing applications.
Despite the effectiveness of machine learning (ML) techniques, their comparative performance in estimating greenhouse gas (GHG) emissions from land-use changes remains underexplored, particularly in tropical agricultural regions. This study addressed this gap by integrating remote sensing data with advanced ML models to quantify carbon (C) and nitrous oxide (N2O) emissions resulting from grassland and cropland expansion in South Sumatra.
Utilizing Landsat imagery and Google Earth Engine (GEE), we conducted a spatial and temporal analysis of land-use changes and their impact on GHG emissions. Key environmental parameters, including enhanced vegetation index (EVI), land surface temperature (LST), normalized difference vegetation index (NDVI), albedo, elevation, humidity, population density, precipitation, soil moisture, and wind speed, are incorporated into the analysis.
This research evaluates the performance of gradient boosting trees (GBT), random forest (RF), support vector machines (SVM), and classification and regression trees (CART) in emission prediction. By offering a comparative assessment of these models, we provide a data-driven framework for improving remote sensing applications in environmental monitoring. The findings contribute to sustainable land management strategies and inform climate change mitigation policies at both local and global scales.

2. Materials and Methods

2.1. Study Area

This study was conducted in the South Sumatra region of Indonesia, a major agricultural hub characterized by extensive landscapes dominated by rice, oil palm, rubber, and other commercial crops. The region experiences a tropical climate with consistently high temperatures and humidity, significantly influencing natural ecosystems and the carbon and nitrogen cycles. Agricultural expansion in South Sumatra has profound implications for land-use change, soil properties, and greenhouse gas (GHG) emissions. Given its importance in national and global agricultural production, balancing greenhouse gas emission reduction with sustainable agricultural development is a critical challenge for the region. This study aims to contribute to this balance by leveraging remote sensing data and machine learning techniques to provide a more accurate estimation of emissions. The insights gained can support policymakers and stakeholders in implementing targeted mitigation strategies while maintaining agricultural productivity. All grassland and cropland areas in the region were utilized for this study to ensure comprehensive spatial coverage of different land-use types and their associated emission dynamics. A detailed representation of the study area is provided in Figure 1.

2.2. Data Sources

To analyze land-use changes and greenhouse gas emissions, this study utilized multi-source remote sensing datasets accessed and processed through the GEE platform. The primary data sources included the following:
  • Landsat 8 OLI: used to derive vegetation indices, including the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI);
  • Soil Grids & Harmonized World Soil Database (HWSD): used to obtain soil properties and organic carbon content;
  • DROSA-A and DROSE-A: used for modeling carbon emissions from agricultural areas and emission data;
  • GLDAS-2.1: Global Land Data Assimilation System: used for soil moisture data;
  • FLDAS: Famine Early Warning Systems Network (FEWS NET) Land Data Assimilation System: used for wind speed and humidity data;
  • MODIS: used for albedo data.

2.3. Data Processing and Analysis

In this study, the FAO DROSE-A dataset was used to select the Sumatera Selatan region, and the analysis was conducted for the 1992–2018 period. The dataset was filtered based on the study area and time range, and visualization parameters were defined to map the spatial distribution of grassland N2O emissions. To enhance model accuracy, additional environmental variables were integrated, including soil moisture from the SMAP dataset, albedo from MODIS, wind speed and humidity from FLDAS, as well as precipitation, land surface temperature (LST), population density, and elevation. Vegetation indices such as NDVI and EVI were calculated using Landsat data, and all datasets were combined to create a comprehensive dataset. A total of 10,000 pixels were randomly sampled to generate training and test datasets. The modeling process utilized the CART algorithm as the primary model, with additional comparisons made using SVM, random forest, and GBT. Model accuracy was evaluated through Pearson correlation, R2, MAE, and RMSE metrics, while visual comparisons of actual and predicted values were conducted. The predicted and actual values were exported in TIFF format, and the results, along with visualizations, were compiled into a detailed report. The workflow diagram is presented in Figure 2.
The technical roadmap of this study follows a comprehensive methodology integrating remote sensing data, machine learning models, and environmental parameters to assess the impact of land-use changes on carbon and nitrous oxide emissions. Landsat satellite imagery from 1992 to 2018 was processed using Google Earth Engine (GEE) to analyze land-use transitions in South Sumatra. A range of environmental parameters, including EVI, LST, NDVI, albedo, elevation, humidity, population density, precipitation, soil moisture, and wind speed, were selected to evaluate their relationship with greenhouse gas emissions. Machine learning algorithms—gradient boosting trees (GBT), random forest (RF), support vector machines (SVM), and classification and regression trees (CART)—were applied to model emissions based on these parameters. The models were validated using performance metrics such as R2, mean squared error (MSE), and root mean squared error (RMSE), ensuring the reliability of the predictions. Finally, the study applied its findings to inform sustainable land management strategies and climate change mitigation policies, providing valuable insights for both local and global efforts to reduce emissions from agricultural land-use changes.
To delineate agricultural areas, remote sensing-based land cover classification was performed using machine learning algorithms, including CART, RF, GBT, and SVM. High-accuracy land cover reference data from 2009 to 2018 were used for model training, and classification accuracy was evaluated using the kappa coefficient (κ) and overall accuracy metrics.
The study period was limited to 1992–2018 due to the availability of high-quality Landsat satellite imagery and remote sensing data up to 2018. The most recent Landsat data at the time of this study were available through 2018, and this period provided a comprehensive dataset for analyzing land-use changes and estimating cropland carbon (C) and nitrous oxide (N2O) emissions. While extending the study period to more recent years would provide additional insights, the data limitations at the time of analysis constrained the temporal scope of the study. Future research may benefit from incorporating more recent data as new satellite imagery becomes available. The emissions were extracted by codes. Annual emission estimates were obtained by averaging the values within the study region to derive regional mean emissions.
In addition to emission factors, several environmental parameters were integrated into the analysis to improve prediction accuracy. These included precipitation, soil moisture, LST, population density, elevation, albedo, wind speed, and humidity. NDVI and EVI were derived from remote sensing datasets to assess vegetation cover and productivity, providing key indicators for land-use classification and emissions estimation. NDVI was computed using the ratio of the difference between near-infrared (NIR) and red bands, while EVI incorporated atmospheric correction factors and was calculated using blue, red, and NIR bands.

2.4. Machine Learning Models

Four machine learning algorithms were employed to predict agricultural C and N2O emissions. The dataset for N2O emissions consisted of 900 samples, with 80% used for training and 20% for testing. For C emissions, a dataset of 1200 samples was utilized, following the same 80–20% split between training and testing. This division ensured that the models were trained on a sufficiently large dataset while maintaining an independent test set for evaluation. The selection of training and testing samples was randomized to prevent bias and improve the generalization capability of the models. Four machine learning algorithms were employed to predict agricultural C and N2O emissions:
  • Classification and regression trees (CART): a decision tree-based model that applies a series of hierarchical decision rules for classification and regression tasks;
  • Random forest (RF): an ensemble learning method that constructs multiple decision trees and averages their predictions to enhance accuracy;
  • Gradient boosting trees (GBT): a boosting technique that sequentially improves model performance by correcting errors in weak learners;
  • Support vector machines (SVM): a model that optimizes the decision boundary (hyperplane) for classification and regression, particularly useful for complex datasets.
The models were trained using historical land-use and environmental datasets, and their predictive performances were evaluated based on correlation coefficient (R), R2, MAE, and RMSE metrics.
The selection of these four models was based on their widespread application in environmental and land-use studies as well as their ability to capture both linear and nonlinear relationships within datasets. Ensemble methods like RF and GBT are known for their robustness and high accuracy, particularly when dealing with heterogeneous environmental data. CART was included as a baseline decision tree method, while SVM was selected due to its capability to handle complex relationships despite its lower performance in this study.
To further refine model accuracy, hyperparameter tuning was conducted using a grid search approach combined with cross-validation. Key hyperparameters were optimized as follows:
  • RF: the number of trees, maximum depth, and minimum samples per split were fine-tuned to balance accuracy and computational efficiency;
  • GBT: learning rate, number of boosting stages, and maximum tree depth were adjusted to prevent overfitting while enhancing predictive power;
  • CART: tree depth and pruning techniques were optimized to reduce model complexity while maintaining interpretability;
  • SVM: the choice of kernel function (linear, polynomial, and radial basis function), regularization parameter (C), and gamma values were explored to improve performance.
Despite the effectiveness of these models, deep learning techniques such as convolutional neural networks (CNNs) and long short-term memory (LSTM) networks were not explored in this study. While deep learning approaches have shown promise in environmental modeling, they typically require significantly larger datasets and higher computational resources. Future research should investigate the applicability of deep learning models for agricultural emission predictions, particularly in integrating spatial and temporal dependencies more effectively.
To determine the most effective model for predicting agricultural greenhouse gas emissions, the performance of each machine learning approach was compared. The results revealed the following trends:
  • RF outperformed other models, demonstrating superior accuracy in capturing emission patterns;
  • GBT exhibited moderate accuracy, indicating its potential for emission estimation;
  • CART provided lower performance, suggesting limitations in handling complex relationships between environmental variables;
  • SVM performed poorly, indicating that it may not be suitable for this dataset due to the nonlinear nature of emission processes.
By leveraging remote sensing datasets and machine learning techniques, this study provides a data-driven approach for estimating emissions from agricultural lands, offering valuable insights for land-use management and climate change mitigation strategies. The use of multiple ML models further provides a comparative assessment, identifying the most effective approach for emissions estimation. Ensemble learning techniques such as random forest (RF) and gradient boosting trees (GBT) demonstrated superior performance in modeling emission patterns, highlighting their potential for large-scale environmental monitoring. These findings underscore the value of ML in greenhouse gas emission research, particularly for complex systems where traditional models may be insufficient.

2.5. Statistical Analysis

The collected data were subjected to statistical analysis to evaluate trends and relationships between environmental variables and emissions. Model performance was assessed using widely recognized statistical metrics [52,53]:
  • Correlation coefficient (R): a statistical measure of the strength of a linear relationship between variables;
  • Coefficient of determination (R2): evaluates the proportion of variance explained by the model;
  • Mean absolute error (MAE): measures the average magnitude of errors between predicted and observed values;
  • Root mean squared error (RMSE): assesses model prediction errors by emphasizing larger deviations.

3. Results

3.1. Changes in Cropland and Grassland Areas

Between 1992 and 2018, there was a general increase in cropland areas in South Sumatra. Cropland areas expanded from approximately 50.46 million hectares in 1992 to 52.03 million hectares in 2018. However, there were short-term fluctuations in some years, such as between 1999 and 2000 and 2008 and 2009. In contrast, grassland areas exhibited a relatively stable trend, showing no significant expansion like cropland areas.
Cropland area changes were as follows:
  • 1992–2000: No significant changes in cropland areas;
  • 2000–2010: A declining trend in cropland areas;
  • 2010–2018: Cropland areas began to increase again.
Grassland area changes were as follows:
Grassland areas remained more stable, showing a more balanced land-use pattern, which might reflect more sustainable land management practices compared to cropland areas.

3.2. Carbon Emissions from Cropland and Grassland

Both cropland and grassland contribute to carbon emissions, but cropland areas show a more noticeable increase in emissions. For croplands, carbon emissions gradually increased until 2016, followed by a drop in 2017, likely due to data anomalies or changes in agricultural practices. In contrast, grassland carbon emissions remained stable over the period, with only minor fluctuations. This suggests that grassland areas are subject to less intensive land management compared to croplands, leading to more stable carbon emissions.
Cropland carbon emissions were as follows:
  • 1992–2016: A gradual increase in carbon emissions;
  • 2017: A noticeable anomaly in emissions;
  • 2018: Emissions resumed an increasing trend.
Grassland carbon emissions were as follows:
  • 1992–2016: Stable carbon emissions with minimal fluctuations;
  • 2017–2018: Minor fluctuations, but emissions remained consistent overall.

3.3. N2O Emissions from Cropland and Grassland

Both cropland and grassland areas showed an increase in N2O emissions, but cropland areas exhibited a more significant rise. As cropland areas expanded, a steady increase in N2O emissions was observed, particularly from 1992 to 2016. Grassland N2O emissions, however, remained relatively stable with minor increases in some years.
Cropland N2O emissions were as follows:
  • 1992–2016: N2O emissions showed a steady upward trend;
  • 2017: An anomaly was detected in the data;
  • 2018: Emissions returned to a normal trend.
Grassland N2O emissions were as follows:
  • 1992–2016: N2O emissions remained stable with slight fluctuations;
  • 2017–2018: Aside from the anomaly in 2016, emissions generally remained steady.
The expansion of cropland areas has contributed to a parallel increase in carbon and N2O emissions. In comparison, grassland emissions remained more stable over time, with only small fluctuations observed. This could be attributed to the fact that grassland areas are less impacted by intensive agricultural practices compared to cropland areas, which often undergo more significant land-use changes, fertilizer application, and other agricultural interventions. The differences in emissions trends highlight the importance of sustainable land management practices. Cropland areas, with their increasing emissions, require more efficient and environmentally friendly agricultural practices to reduce both C and N2O emissions. Meanwhile, the stable emissions from grassland areas suggest that maintaining these lands with minimal disturbance could play a role in mitigating the overall emissions from the region. In conclusion, balancing agricultural expansion with sustainable practices and preserving grassland areas will be key to mitigating emissions and ensuring a more sustainable land use strategy for South Sumatra.

3.4. Performance of Machine Learning Models for Cropland and Grassland Emissions

The performance of various machine learning models was evaluated for both cropland C and N2O emissions.
Table 1 presents the performance of four machine learning models used for carbon emission estimation in croplands. The GBT model demonstrated the highest predictive accuracy with an R2 of 0.7106, indicating a strong correlation between environmental variables and carbon emissions. Random forest (RF) followed with an R2 of 0.5927, also showing high predictive power. CART (R2 = 0.4752) exhibited moderate performance, while SVM performed poorly (R2 = 0.2406, MAE = 29.4234, RMSE = 29.4289), suggesting that SVM struggles to capture the complex relationships governing carbon emissions in croplands. These results highlight the effectiveness of ensemble learning methods (GBT and RF) in estimating cropland carbon emissions.
Table 2 compares the predictive performance of machine learning models for estimating nitrous oxide (N2O) emissions in croplands. The GBT model achieved the highest predictive accuracy (R2 = 0.7106), followed by RF (R2 = 0.5963). This finding suggests that ensemble methods effectively model the relationship between environmental parameters and N2O emissions. CART exhibited moderate performance, while SVM performed the worst (R2 = 0.2406, MAE = 29.7635, RMSE = 29.7701), indicating its inability to handle the complex interactions influencing N2O emissions. Overall, ensemble learning techniques demonstrated superior prediction capability in cropland N2O estimation.
Table 3 presents the performance of models in estimating carbon emissions in grasslands. GBT (R2 = 0.3923) and RF (R2 = 0.3875) showed similar performance, though with significantly lower accuracy compared to cropland estimations. This suggests that carbon emissions in grasslands are governed by more complex interactions, making them harder to predict. CART displayed moderate predictive capability, while SVM performed poorly (R2 = 0.1308, MAE = 29.7212, RMSE = 29.7274), confirming its inadequacy in modeling carbon emissions in grasslands. These results highlight the need for more refined modeling approaches to improve grassland carbon emission predictions.
Table 4 summarizes the performance of machine learning models in estimating N2O emissions in grasslands. RF demonstrated the highest predictive power (R2 = 0.3904), followed by GBT (R2 = 0.2669). CART exhibited moderate performance, while SVM performed the worst (R2 = 0.0036, MAE = 29.7329, RMSE = 29.7386), indicating that it is unsuitable for predicting N2O emissions in grasslands. The results suggest that RF is the most reliable model for grassland N2O estimation, but overall, predicting N2O emissions in grasslands remains more challenging than in croplands.
The results are as follows.
Cropland C emissions results are as follows:
  • GBT and RF are the best-performing models, with RF slightly outperforming GBT. The higher accuracy of RF indicates that it may be more suitable for predicting C emissions from cropland compared to GBT. Both models exhibit low error metrics, making them reliable choices for C emission predictions;
  • CART shows a relatively lower correlation compared to GBT and RF, indicating it struggles more to capture the underlying trends in C emissions. Its performance in terms of accuracy is subpar;
  • SVM performs the worst, with very low correlation and high error values. This suggests that SVM is not effective for predicting C emissions from cropland in this case.
Cropland N2O emissions results are as follows:
  • GBT leads again with the best accuracy and minimal error, showing its strong capability in predicting N2O emissions with precision;
  • RF follows closely with nearly identical error metrics to GBT, reinforcing the fact that both models are highly effective in predicting N2O emissions from cropland;
  • CART has a lower correlation and higher error metrics, suggesting it is not as efficient at capturing the relationships in the data as GBT and RF;
  • SVM again shows the poorest performance with extremely high error values, making it unsuitable for this task.
Grassland C emissions results are as follows:
  • RF model performs the best for predicting grassland C emissions. It demonstrates a good ability to capture the relationships in the data, making it the most reliable model for this task;
  • GBT also performs well but with slightly lower accuracy compared to RF. While GBT still provides reasonable predictions, it captures less of the variation in the emissions data compared to RF;
  • CART and SVM show weaker performances. CART captures some of the trends but struggles compared to RF and GBT. On the other hand, SVM performs poorly, failing to effectively capture the patterns in grassland C emissions.
Grassland N2O emissions results are as follows:
  • For grassland N2O emissions, RF again outperforms other models, showing the best predictive capability. It is the most effective model for capturing the trends and patterns in the data;
  • GBT also provides good results, though slightly less accurate than RF. It remains a strong alternative for predicting N2O emissions in grassland areas;
  • CART demonstrates moderate performance but does not capture the underlying relationships as effectively as RF or GBT;
  • SVM again shows the weakest performance, indicating that it is not well suited for this task, with poor accuracy and large errors.
In summary, RF and GBT are the most robust models for predicting both C and N2O emissions from cropland and grassland areas, while CART and SVM demonstrate weaker performance, especially in terms of accuracy and error minimization.
The findings indicate that GBT and RF models consistently outperform other machine learning approaches in predicting both carbon and N2O emissions across croplands and grasslands. SVM exhibited poor predictive performance in all cases, likely due to its inability to capture complex environmental interactions. The inclusion of key environmental variables such as rainfall, elevation, population density, land surface temperature (LST), soil moisture, albedo, wind speed, humidity, mean NDVI, mean EVI, mean SAVI, and mean NDWI significantly contributed to model performance. However, the lower predictive accuracy observed in grasslands suggests that carbon and N2O emissions in these areas are influenced by more intricate and heterogeneous processes.
Despite the strong predictive performance of GBT and RF models in cropland carbon (C) and nitrous oxide (N2O) emissions estimation (Table 1 and Table 2), the results for grasslands (Table 3 and Table 4) exhibited relatively lower R2 values, particularly for SVM and CART models. The lower R2 values, in some cases reaching as low as 0.1 or below, indicate the increased complexity of emission dynamics in grassland ecosystems compared to croplands. This discrepancy may arise due to the heterogeneous nature of grasslands, which exhibit spatially and temporally variable biophysical characteristics, such as soil organic matter content, vegetation density, and microclimatic factors, making it more challenging for machine learning models to establish strong predictive relationships.
Furthermore, the SVM model consistently underperformed across all datasets, with exceptionally high MAE and RMSE values, suggesting its inadequacy in capturing the complex, nonlinear relationships governing greenhouse gas emissions in both croplands and grasslands. The results highlight the necessity of employing ensemble learning techniques (GBT and RF) for improved predictive accuracy, particularly in cropland environments where structured agricultural practices create more predictable emission patterns. Future research should explore hybrid modeling approaches or integrate additional environmental covariates to enhance the predictive performance of machine learning models, particularly in highly heterogeneous landscapes such as grasslands.
These results highlight the potential of ensemble-based machine learning approaches in estimating greenhouse gas emissions from agricultural and grassland ecosystems. Future research should focus on refining these models by incorporating additional environmental parameters and advanced feature selection techniques to further enhance prediction accuracy.
Figure 3 illustrates the spatial distribution of carbon (C) emissions in croplands using four machine learning models: gradient boosting trees (GBT), random forest (RF), classification and regression trees (CART), and support vector machine (SVM). The reference carbon emission map (Figure 3a) provides a baseline for model evaluation. Among the models, GBT (Figure 3b) demonstrates the highest level of spatial detail and precision, effectively capturing localized variations. RF (Figure 3c) follows a similar trend but produces a slightly smoother distribution, suggesting that it generalizes the spatial patterns more than GBT. CART (Figure 3d) yields a more uniform spatial distribution, indicating reduced sensitivity to local emission variations. Meanwhile, SVM (Figure 3e) presents the least variation, suggesting that it struggles to capture fine-scale carbon emission patterns. These results indicate that GBT and RF offer superior predictive capabilities for cropland carbon estimation, whereas CART and SVM exhibit notable limitations in capturing spatial heterogeneity.
Figure 4 presents the spatial distribution of nitrous oxide (N2O) emissions in croplands based on different machine learning models. The reference N2O emission map (Figure 4a) serves as a benchmark for comparison. GBT (Figure 4b) provides a highly detailed spatial pattern, effectively identifying high-emission hotspots, demonstrating its strong sensitivity to localized variations. RF (Figure 4c) exhibits a comparable but slightly less detailed representation, indicating a tendency to generalize spatial variations. CART (Figure 4d) produces a more homogenous output, likely underestimating spatial variability in N2O emissions. In contrast, SVM (Figure 4e) displays minimal spatial differentiation, reflecting poor predictive performance in estimating N2O emissions. These findings suggest that GBT and RF offer the most reliable spatial predictions for N2O emissions, while CART and SVM struggle to capture complex spatial patterns. The spatial estimation maps indicate that ensemble models, particularly GBT and RF, outperform CART and SVM in predicting both carbon and N2O emissions in croplands. The superior performance of GBT and RF can be attributed to their ability to capture non-linear relationships and spatial heterogeneity, allowing for more precise and reliable predictions. Conversely, CART produces oversimplified results, while SVM fails to detect fine-scale emission variations, leading to weak predictive capabilities. These findings highlight the importance of utilizing advanced ensemble learning techniques for accurate greenhouse gas emission estimation in agricultural landscapes, emphasizing their role in enhancing precision agriculture and environmental monitoring efforts.
Figure 5 illustrates the spatial distribution of C emissions in grasslands using different machine learning models, including GBT, RF, CART, and SVM. The reference carbon emission map (Figure 5a) serves as a baseline for comparison. Among the models, GBT (Figure 5b) demonstrates the most detailed and spatially heterogeneous carbon distribution, indicating a strong ability to capture fine-scale variations. RF (Figure 5c) also effectively represents spatial variability but appears slightly smoother compared to GBT, suggesting a moderate generalization effect. In contrast, CART (Figure 5d) provides a more uniform estimation, likely due to its lower sensitivity to fine-scale variations, while SVM (Figure 5e) exhibits minimal spatial differentiation, indicating weak predictive capability for carbon emissions in grasslands. Overall, GBT and RF outperform CART and SVM, with GBT demonstrating the highest level of accuracy in capturing spatial patterns of carbon emissions.
Figure 6 presents the spatial distribution of N2O emissions in grasslands, estimated using the same machine learning models. The reference map (Figure 6a) provides a benchmark for evaluating model performance. Like carbon emissions, GBT (Figure 6b) captures a detailed and spatially diverse N2O distribution, indicating high predictive accuracy. RF (Figure 6c) performs similarly but appears slightly less detailed, likely due to its tendency to smooth spatial variations. CART (Figure 6d) produces a more homogeneous distribution, potentially underestimating localized variations in N2O emissions. In contrast, SVM (Figure 6e) shows minimal differentiation across the landscape, suggesting poor model performance in estimating N2O emissions. Overall, the results indicate that ensemble-based models, particularly GBT and RF, outperform CART and SVM in estimating both carbon and N2O emissions in grassland ecosystems. The GBT model consistently provides the most accurate and detailed spatial representations, while RF also performs well but with slightly less granularity. Conversely, CART and SVM demonstrate lower predictive capabilities, with CART oversimplifying spatial patterns and SVM failing to capture meaningful emission trends. These findings emphasize the advantages of using advanced ensemble learning techniques for greenhouse gas estimation in grassland ecosystems, supporting their application for high-precision environmental monitoring and sustainable land management.

4. Discussion

This study examines the changes in agricultural areas in South Sumatra between 1992 and 2018 and the effects of these changes on C and N2O emissions. The findings indicate a direct relationship between the expansion of agricultural areas and the increase in C and N2O emissions.

4.1. Agricultural Expansion and Environmental Impacts

The expansion of agricultural areas in South Sumatra over the past 26 years has been a major driver of environmental change, particularly in relation to carbon cycling and greenhouse gas emissions. This finding is consistent with other studies that have reported similar trends in tropical regions, where agricultural expansion often leads to deforestation and a reduction in soil carbon stocks [54,55]. The significant increase in palm oil plantations in Indonesia, leading to large-scale deforestation, has been widely documented [56]. Our study reinforces these findings by quantifying the direct link between land-use change and increased C and N2O emissions in South Sumatra, underscoring the broader implications of agricultural practices on climate change.

4.2. Carbon and N2O Emission Trends

The results show that carbon emissions exhibited a consistent upward trend between 1992 and 2016, followed by a sharp decline in 2017. This anomaly may have been caused by data-related errors or sudden changes in regional agricultural policies.
The consistent increase in carbon emissions aligns with trends observed in other agricultural regions undergoing similar land-use transformations, where deforestation-driven expansion has led to significant carbon losses [57]. However, the sharp decline in 2017 warrants further investigation, as similar trends have been documented in countries where policy interventions aimed at reducing deforestation resulted in temporary reductions in emissions [58]. Future research should explore the causal mechanisms behind this decline to assess its long-term implications.
N2O emissions generally remained at low levels but exhibited minor fluctuations in response to agricultural expansion. The primary source of N2O emissions in agricultural lands is fertilizer application. However, due to the relatively low reliance on nitrogen-based fertilizers in South Sumatra’s agricultural production, N2O emissions have remained lower compared to regions with more intensive farming practices [59]. These findings are consistent with previous studies indicating that fertilizer use is a major determinant of N2O emissions in agricultural landscapes. Future studies should incorporate more detailed fertilizer application data to enhance the accuracy of emission estimates and better understand the impact of land management practices on greenhouse gas emissions.

4.3. Policy Recommendations and Sustainable Agriculture

The findings emphasize the importance of sustainable agricultural practices in South Sumatra. The expansion of agricultural areas and the associated increase in C and N2O emissions necessitate the following policy recommendations:
  • Forest conservation and sustainable land management: Reforestation and agroforestry practices should be encouraged to mitigate the negative effects of agricultural expansion on natural ecosystems. Studies have shown that such strategies can effectively enhance carbon sequestration and reduce net emissions from land-use change [60];
  • Adoption of low-carbon agricultural techniques: Sustainable farming methods such as no-till farming and composting, which help preserve soil organic matter and reduce carbon loss, should be promoted. Prior research has demonstrated that no-till farming can significantly improve soil carbon retention and decrease greenhouse gas emissions from cropland [61];
  • Smart fertilizer management: Precision farming technologies and data analytics should be utilized to optimize fertilizer application and minimize excessive nitrogen-based fertilizer use.
The policy recommendations derived from this study align with previous research on sustainable agricultural practices. Reforestation and agroforestry have been widely recognized as effective strategies to mitigate the impact of agricultural expansion on carbon emissions [60]. Similarly, the promotion of low-carbon agricultural techniques, such as no-till farming, has been recommended as a crucial method for reducing soil carbon loss and improving long-term sustainability in agricultural landscapes [61]. Our findings contribute to this growing body of research by emphasizing the specific context of South Sumatra and providing localized solutions based on the observed trends in agricultural emissions. Future policies should integrate these strategies to balance food security with environmental sustainability.

4.4. Study Limitations and Future Research

This study utilized remote sensing and big data analysis techniques to assess agricultural land-use changes and associated emissions. While these methods provide valuable insights, certain limitations must be acknowledged, and future research should address these gaps to improve the robustness of similar analyses.
One key limitation is the challenge of capturing small-scale agricultural activities using satellite imagery. Remote sensing techniques, despite their advantages in large-scale land monitoring, often struggle to detect smallholder farming practices, a well-documented issue in land-use change studies [62]. Future research should integrate high-resolution UAV (unmanned aerial vehicle) imagery and field-based validation surveys to refine land cover classification and improve the accuracy of emission estimates. A hybrid approach combining satellite observations with ground-truthing data would enhance the reliability of remote sensing-based assessments.
Another limitation concerns the absence of detailed data on fertilizer use and agricultural management practices, which introduces uncertainties in model predictions. As noted in previous studies [63], incorporating precise information on fertilizer application rates, crop rotation patterns, and land management strategies would provide a more comprehensive understanding of N2O emission dynamics. Given that fertilizer use is a primary driver of agricultural N2O emissions, access to more granular datasets would enable improved modeling and contribute to the development of targeted mitigation strategies.
Additionally, while this study highlights the relationship between agricultural expansion and greenhouse gas emissions, future research should integrate localized datasets and advanced machine learning techniques to further refine predictive accuracy. The inclusion of ground-based measurements from local monitoring stations would significantly enhance model calibration, reducing uncertainties in emission estimations. Moreover, the application of deep learning approaches for land cover classification and emission modeling could provide more robust and automated solutions for large-scale environmental assessments.
By addressing these limitations, future studies can enhance the applicability of remote sensing techniques in agricultural and climate research. A more comprehensive and interdisciplinary approach, combining remote sensing, field data, and advanced AI models, will be crucial in developing scalable and effective strategies for land-use planning and climate change mitigation.

4.5. Performance of Machine Learning Models

This study evaluated the performance of machine learning models in predicting agricultural-based C and N2O emissions. The results showed that the random forest (RF) and gradient boosting trees (GBT) models provided the highest accuracy and outperformed other models. This can be attributed to the ability of RF and GBT models to effectively handle large datasets and complex variable relationships. Particularly, when modeling variables influenced by numerous environmental factors, such as carbon emissions, these models excel due to their ability to learn complex patterns in the data [64].
In contrast, the support vector machine (SVM) model performed poorly and had low accuracy rates for both emission types. The main reason for SVM’s failure is the high variability and nonlinear relationships commonly found in such environmental datasets. Kernel-based methods like SVM may not provide the expected performance in large-scale, heterogeneous datasets due to their sensitivity to parameters. This highlights that, in model selection, not only theoretical accuracy rates but also the compatibility of the model with the dataset’s structure are critical factors [65].
Another key point is the clear observation of the impact of environmental factors (humidity, precipitation, and soil moisture content) on C and N2O emissions. Integrating these variables into the model could further improve prediction accuracy. Previous studies have also shown that incorporating environmental variables into emission prediction models increases model accuracy. For instance, when assessing the impact of agricultural activities on emissions, it is necessary to adopt a more comprehensive modeling approach that considers not only land expansion but also soil management and climate variables [66].
In conclusion, RF and GBT models demonstrated the best performance in emission prediction, proving once again that they are powerful tools for large-scale data analysis. RF’s capacity to handle high-dimensional data and its resistance to overfitting made it stand out. On the other hand, the low accuracy of SVM once again demonstrated that not all models are equally effective in complex environmental problems like agricultural emission prediction.
These findings have significant implications not only theoretically but also practically. In future studies, the use of high-resolution remote sensing data and advanced machine learning techniques could help model the impact of agricultural activities on carbon sequestration capacity more accurately. Moreover, integrating such models into sustainable agricultural policies could lead to the development of more efficient strategies for reducing carbon emissions [66].

5. Conclusions

This study evaluated the changes in agricultural areas in South Sumatra between 1992 and 2018 and the impacts of these changes on C and N2O emissions. The analysis, conducted using remote sensing data and GEE, revealed that agricultural expansion has a direct impact on greenhouse gas emissions.
The main findings of the study are as follows:
  • Expansion of agricultural areas: Between 1992 and 2018, agricultural areas in South Sumatra expanded by approximately 1.57 million hectares. A small decrease was observed in the early 2000s, but after 2010, the expansion trend continued;
  • Carbon emissions: Carbon emissions have generally shown a continuous increase, with a sharp decline in 2017. This decline may be attributed to data-related errors or changes in regional agricultural policies. Emissions increased again in 2018;
  • N2O emissions: Nitrous oxide emissions showed a slight upward trend in parallel with the expansion of agricultural areas. This increase is primarily due to changes in fertilizer use and agricultural practices.
These findings support the initial hypothesis that agricultural expansion is a key driver of increasing greenhouse gas emissions. The results confirm that the conversion of natural landscapes into cropland and grassland has contributed to the observed rise in C and N2O emissions, emphasizing the significant role of land-use changes in regional emission dynamics.
It has been determined that sustainable agricultural policies play a critical role in reducing carbon and N2O emissions. Preventing deforestation, implementing agricultural techniques that enhance carbon sequestration, and optimizing fertilizer management emerge as key strategies for emission reduction. Additionally, integrating remote sensing with machine learning models provides a robust approach for monitoring and mitigating the environmental impact of land-use changes.
This study provides valuable insights into the relationship between agricultural expansion and greenhouse gas emissions. However, more detailed analyses supported by regional emission factors and field data are needed. Future research should focus on improving emission prediction accuracy by integrating local measurement data and utilizing advanced modeling techniques.
To balance agriculture and environmental sustainability, collaboration between policymakers, scientists, and farmers is essential. Additionally, the study highlights the significant role of environmental parameters in predicting C and N2O emissions. In comparison of machine learning models, RF and GBT models were found to have superior prediction accuracy, while the SVM model performed poorly. These results suggest that ensemble methods are effective in modeling the relationship between environmental variables and greenhouse gas emissions.
The findings of this study emphasize the need for integrated approaches that combine remote sensing, machine learning, and environmental analysis to assess the impacts of agricultural expansion on carbon and nitrous oxide emissions. This approach offers significant advantages in terms of accuracy and scalability, making it a valuable tool for policy making and sustainable land management. By relating the results to the existing literature, we provide a more comprehensive understanding of the environmental implications of agricultural practices and suggest practical recommendations for mitigating emissions in regions experiencing rapid land-use changes.
Future research should focus on improving prediction accuracy by integrating local emission factors, enhancing model parameters, and utilizing additional environmental variables. Sustainable land use and precision agriculture strategies are critical to reducing emissions and enhancing carbon sequestration. Furthermore, the findings of this study could be applied in the development of policy frameworks aimed at reducing greenhouse gas emissions from agriculture, supporting climate change mitigation efforts at both local and global scales.

Author Contributions

Conceptualization, A.U. and N.U.; data curation, A.U.; formal analysis, N.U.; methodology, A.U. and N.U.; supervision, N.U.; visualization, A.U.; writing—original draft, N.U.; writing—review and editing, A.U. and N.U. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The datasets are available from the corresponding author on reasonable request.

Acknowledgments

We would like to thank the anonymous reviewers and editors for their valuable comments and suggestions regarding this article.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
N2ONitrous oxide
CCarbon
LU/LUCLand use and land-use change
SOCSoil organic carbon
GEEGoogle Earth Engine
RFRandom forest
CARTClassification and regression trees
GBTGradient boosting trees
SVMSupport vector machines
RBFRadial basis function
NDVINormalized difference vegetation index
EVIEnhanced vegetation index
HWSDHarmonized World Soil Database
GLDASGlobal Land Data Assimilation System
FLDASFamine Early Warning Systems Network Land Data Assimilation System
LSTLand surface temperature
RCorrelation coefficient
R2Coefficient of determination
MAEMean absolute error
RMSERoot mean squared error

References

  1. Lehmann, J.; Kleber, M. The contentious nature of soil organic matter. Nature 2015, 528, 60. [Google Scholar] [CrossRef] [PubMed]
  2. Arnell, N.W.; Lowe, J.A.; Brown, S.; Gosling, S.N.; Gottschalk, P.; Hinkel, J.; Lloyd-Hughes, B.; Nicholls, R.J.; Osborne, T.M.; Rose, G.A.; et al. A global assessment of the effects of climate policy on the impacts of climate change. Nat. Clim. Change 2013, 3, 512–519. [Google Scholar] [CrossRef]
  3. Buchholz, T.; Friedland, A.J.; Hornig, C.E.; Keeton, W.S.; Zanchi, G.; Nunery, J. Mineral soil carbon fluxes in forests and implications for carbon balance assessments. GCB Bioenergy 2014, 6, 305–311. [Google Scholar] [CrossRef]
  4. Rogers, B.M.; Mackey, B.; Shestakova, T.A.; Keith, H.; Young, V.; Kormos, C.F.; Moomaw, W.R. Using ecosystem integrity to maximize climate mitigation and minimize risk in international forest policy. Front. For. Glob. Change 2022, 5, 929281. [Google Scholar] [CrossRef]
  5. Bernstein, L.; Bosch, P.; Canziani, O.; Chen, Z.; Christ, R.; Davidson, O.; Hare, W.; Huq, S.; Karoly, D.; Kattsov, V.; et al. Climate Change 2007 Synthesis Report. Contribution of Working Groups I, II and III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change; IPCC: Geneva, Switzerland, 2007. [Google Scholar]
  6. Dong, H.; Zeng, Q.; Sheng, Y.; Chen, C.; Yu, G.; Kappler, A. Coupled iron cycling and organic matter transformation across redox interfaces. Nat. Rev. Earth Environ. 2023, 4, 659–673. [Google Scholar] [CrossRef]
  7. Guo, L.B.; Gifford, R.M. Soil carbon stocks and land use change: A meta-analysis. Glob. Change Biol. 2002, 8, 345–360. [Google Scholar] [CrossRef]
  8. Poeplau, C.; Don, A.; Vesterdal, L.; Leifeld, J.; Van Wesemael, B.; Schumacher, J.; Gensior, A. Temporal dynamics of soil organic carbon after land-use change in the temperate zone—Carbon response functions as a model approach. Glob. Change Biol. 2011, 17, 2415–2427. [Google Scholar] [CrossRef]
  9. Popp, A.; Humpenöder, F.; Weindl, I.; Bodirsky, B.L.; Bonsch, M.; Lotze-Campen, H.; Müller, C.; Biewald, A.; Rolinski, S.; Stevanovic, M.; et al. Land-use protection for climate change mitigation. Nat. Clim. Change 2014, 4, 1095–1098. [Google Scholar] [CrossRef]
  10. Ameray, A.; Bergeron, Y.; Valeria, O.; Montoro Girona, M.; Cavard, X. Forest carbon management: A review of silvicultural practices and management strategies across boreal, temperate and tropical forests. Curr. For. Rep. 2021, 7, 245–266. [Google Scholar] [CrossRef]
  11. Doetterl, S.; Stevens, A.; Six, J.; Merckx, R.; Van Oost, K.; Casanova Pinto, M.; Casanova-Katny, A.; Muñoz, C.; Boudin, M.; Venegas, E.Z.; et al. Soil carbon storage controlled by interactions between geochemistry and climate. Nat. Geosci. 2015, 8, 780–783. [Google Scholar] [CrossRef]
  12. Lal, R. Soil carbon sequestration impacts on global climate change and food security. Science 2004, 304, 1623–1627. [Google Scholar] [CrossRef] [PubMed]
  13. Wu, H.; Cui, H.; Fu, C.; Li, R.; Qi, F.; Liu, Z.; Qiao, M. Unveiling the crucial role of soil microorganisms in carbon cycling: A review. Sci. Total Environ. 2024, 909, 168627. [Google Scholar] [PubMed]
  14. Sanderman, J.; Hengl, T.; Fiske, G.J. Soil carbon debt of 12,000 years of human land use. Proc. Natl. Acad. Sci. USA 2017, 114, 9575–9580. [Google Scholar] [CrossRef] [PubMed]
  15. Reichstein, M.; Bahn, M.; Ciais, P.; Frank, D.; Mahecha, M.D.; Seneviratne, S.I.; Zscheischler, J.; Beer, C.; Buchmann, N.; Frank, D.C.; et al. Climate extremes and the carbon cycle. Nature 2013, 500, 287–295. [Google Scholar]
  16. Koven, C.D.; Hugelius, G.; Lawrence, D.M.; Wieder, W.R. Higher climatological temperature sensitivity of soil carbon in cold than warm climates. Nat. Clim. Change 2017, 7, 817–822. [Google Scholar]
  17. Beillouin, D.; Corbeels, M.; Demenois, J.; Berre, D.; Boyer, A.; Fallot, A.; Cardinael, R. A global meta-analysis of soil organic carbon in the Anthropocene. Nat. Commun. 2023, 14, 3700. [Google Scholar] [CrossRef]
  18. Rumpel, C.; Amiraslani, F.; Chenu, C.; Garcia Cardenas, M.; Kaonga, M.; Koutika, L.-S.; Ladha, J.; Madari, B.; Shirato, Y.; Smith, P.; et al. The 4p1000 initiative: Opportunities, limitations and challenges for implementing soil organic carbon sequestration as a sustainable development strategy. Ambio 2019, 49, 350–360. [Google Scholar] [CrossRef]
  19. Butterbach-Bahl, K.; Kiese, R. Biofuel production on the margins: An analysis shows that fuel made from wild, herbaceous vegetation grown on land currently unsuitable for cultivating field crops could contribute substantially to the United States′ targets for biofuel production. Nature 2013, 492, 483–486. [Google Scholar] [CrossRef]
  20. Paustian, K.; Lehmann, J.; Ogle, S.; Reay, D.; Robertson, G.P.; Smith, P. Climate-smart soils. Nature 2016, 532, 49–57. [Google Scholar] [CrossRef]
  21. Davidson, E.A.; Kanter, D. Inventories and scenarios of nitrous oxide emissions. Environ. Res. Lett. 2014, 9, 105012. [Google Scholar] [CrossRef]
  22. Tian, H.Q.; Yang, J.; Lu, C.Q. The Global N2O Model Intercomparison Project. Bull. Am. Meteorol. Soc. 2018, 99, 1231–1252. [Google Scholar] [CrossRef]
  23. Butterbach-Bahl, K.; Baggs, E.M.; Dannenmann, M.; Kiese, R.; Zechmeister-Boltenstern, S. Nitrous oxide emissions from soils: How well do we understand the processes and their controls? Philos. Trans. R. Soc. B Biol. Sci. 2013, 368, 20130122. [Google Scholar] [CrossRef] [PubMed]
  24. Pihlatie, M.; Syväsalo, E.; Simojoki, A.; Esala, M.; Regina, K. Contribution of nitrification and denitrification to N2O production in peat, clay and loamy sand soils under different soil moisture conditions. Nutr. Cycl. Agroecosyst. 2004, 70, 135–141. [Google Scholar] [CrossRef]
  25. Bateman, E.J.; Baggs, E.M. Contributions of nitrification and denitrification to N2O emissions from soils at different water-filled pore space. Biol. Fertil. Soils 2005, 41, 379–388. [Google Scholar] [CrossRef]
  26. Yu, W.; Hayat, K.; Ma, J.; Fan, X.; Yang, Y.; Zhang, Z.; Lin, H. Effect of antibiotic perturbation on nitrous oxide emissions: An in-depth analysis. Crit. Rev. Environ. Sci. Technol. 2024, 54, 1612–1632. [Google Scholar] [CrossRef]
  27. Zhang, G.; Bai, J.; Zhai, Y.; Jia, J.; Zhao, Q.; Wang, W.; Hu, X. Microbial diversity and functions in saline soils: A review from a biogeochemical perspective. J. Adv. Res. 2024, 59, 129–140. [Google Scholar] [CrossRef]
  28. Giles, M.; Morley, N.; Baggs, E.M.; Daniell, T.J. Soil nitrate reducing processes–Drivers, mechanisms for spatial variation, and significance for nitrous oxide production. Front. Microbiol. 2012, 3, 407. [Google Scholar] [CrossRef] [PubMed]
  29. Philippot, L.; Hallin, S.; Borjesson, G.; Baggs, E.M. Biochemical cycling in the rhizosphere having an impact on global change. Plant Soil 2009, 321, 61–81. [Google Scholar] [CrossRef]
  30. Nishimura, S.; Sawamoto, T.; Akiyama, H.; Sudo, S.; Yagi, K. Methane and nitrous oxide emissions from a paddy field with Japanese conventional water management and fertilizer application. Glob. Biogeochem. Cycl. 2004, 18, GB2017. [Google Scholar] [CrossRef]
  31. Cai, Z.C.; Xing, G.X.; Yan, X.Y.; Xu, H.; Tsuruta, H.; Yagi, K.; Minami, K. Methane and nitrous oxide emissions from rice paddy fields as affected by nitrogen fertilisers and water management. Plant Soil 1997, 196, 7–14. [Google Scholar] [CrossRef]
  32. Kaiser, E.-A.; Kohrs, K.; Kucke, M.; Schnug, E.; Heinemeyer, O.; Munch, J.C. Nitrous oxide release from arable soil: Importance of N-fertilization, crops and temporal variation. Soil Biol. Biochem. 1998, 30, 1553–1563. [Google Scholar] [CrossRef]
  33. Nishimura, S.; Akiyama, H.; Sudo, S.; Fumoto, T.; Cheng, W.G.; Yagi, K. Combined emission of CH₄ and N₂O from a paddy field was reduced by preceding upland crop cultivation. Soil Sci. Plant Nutr. 2011, 57, 167–178. [Google Scholar] [CrossRef]
  34. Liu, D.; Song, C.; Xin, Z.; Fang, C.; Liu, Z.; Xu, Y. Agricultural management strategies for balancing yield increase, carbon sequestration, and emission reduction after straw return for three major grain crops in China: A meta-analysis. J. Environ. Manag. 2023, 340, 117965. [Google Scholar]
  35. Qian, H.; Zhu, X.; Huang, S.; Linquist, B.; Kuzyakov, Y.; Wassmann, R.; Jiang, Y. Greenhouse gas emissions and mitigation in rice agriculture. Nat. Rev. Earth Environ. 2023, 4, 716–732. [Google Scholar]
  36. Toma, Y.; Kimura, S.D.; Hirose, Y.; Kusa, K.; Hatano, R. Variation in the emission factor of N₂O derived from chemical nitrogen fertilizer and organic matter: A case study of onion fields in Mikasa, Hokkaido, Japan. Soil Sci. Plant Nutr. 2007, 53, 692–703. [Google Scholar] [CrossRef]
  37. Akiyama, H.; Yagi, K.; Yan, X.Y. Direct N₂O emissions from rice paddy fields: Summary of available data. Glob. Biogeochem. Cycl. 2005, 19, GB1005. [Google Scholar] [CrossRef]
  38. Kusa, K.; Hu, R.G.; Sawamoto, T.; Hatano, R. Three years of nitrous oxide and nitric oxide emissions from silandic andosols cultivated with maize in Hokkaido, Japan. Soil Sci. Plant Nutr. 2006, 52, 103–113. [Google Scholar] [CrossRef]
  39. Toma, Y.; Hatano, R. Effect of crop residue C:N ratio on N₂O emissions from gray lowland soil in Mikasa, Hokkaido, Japan. Soil Sci. Plant Nutr. 2007, 53, 198–205. [Google Scholar] [CrossRef]
  40. Jia, J.X.; Sun, L.Y.; Kong, X.W.; Yan, X.Y.; Xiong, Z.Q. Annual N₂O and CH₄ emissions from intensively managed vegetable fields in Nanjing, China. Soil Sci. Plant Nutr. 2012, 58, 91–103. [Google Scholar] [CrossRef]
  41. Shimizu, M.; Hatano, R.; Arita, T.; Kouda, Y.; Mori, A.; Matsuura, S.; Niimi, M.; Jin, T.; Desyatkin, A.R.; Kawamura, O.; et al. The effect of fertilizer and manure application on CH₄ and N₂O emissions from managed grasslands in Japan. Soil Sci. Plant Nutr. 2013, 59, 69–86. [Google Scholar] [CrossRef]
  42. Yano, M.; Toyoda, S.; Tokida, T.; Hayashi, K.; Hasegawa, T.; Makabe, A.; Koba, K.; Yoshida, N. Isotopomer analysis of production, consumption and soil-to-atmosphere emission processes of N2O at the beginning of paddy field irrigation. Soil Biol. Biochem. 2014, 70, 66–78. [Google Scholar] [CrossRef]
  43. Arunrat, N.; Sereenonchai, S.; Wang, C. Carbon footprint and predicting the impact of climate change on carbon sequestration ecosystem services of organic rice farming and conventional rice farming: A case study in Phichit province, Thailand. J. Environ. Manag. 2021, 289, 112458. [Google Scholar] [CrossRef] [PubMed]
  44. Nadkarni, S.B.; Vijay, G.S.; Kamath, R.C. Comparative study of random forest and gradient boosting algorithms to predict airfoil self-noise. Eng. Proc. 2023, 59, 24. [Google Scholar] [CrossRef]
  45. Imandoust, S.; Bolandraftar, M. Forecasting the direction of stock market index movement using three data mining techniques: The case of Tehran Stock Exchange. Int. J. Eng. Res. Appl. 2014, 4, 106–117. [Google Scholar]
  46. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  47. Dudek, G. Short-Term Load Forecasting Using Random Forests. In Intelligent Systems’2014; Springer International Publishing: Cham, Switzerland, 2015; pp. 821–828. [Google Scholar] [CrossRef]
  48. Friedman, J.H. Greedy function approximation: A gradient boosting machine. Ann. Stat. 2001, 29, 1189–1232. [Google Scholar]
  49. Wang, J.; Wu, X.; Zhang, C. Support vector machines based on K-means clustering for real-time business intelligence systems. Int. J. Bus. Intell. Data Min. 2005, 1, 54–64. [Google Scholar] [CrossRef]
  50. Boser, B.E.; Guyon, I.M.; Vapnik, V.N. A training algorithm for optimal margin classifiers. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory, Pittsburgh, PA, USA, 27–29 July 1992; pp. 144–152. [Google Scholar] [CrossRef]
  51. Zhai, Y.; Hsu, A.; Halgamuge, S.K. Combining news and technical indicators in daily stock price trends prediction. In Advances in Neural Networks-ISNN 2007; Springer: Berlin/Heidelberg, Germany, 2007; pp. 1087–1096. [Google Scholar] [CrossRef]
  52. Hastie, T.; Tibshirani, R.; Friedman, J.H.; Friedman, J.H. The Elements of Statistical Learning: Data Mining, Inference, and Prediction; Springer: New York, NY, USA, 2009; Volume 2, pp. 1–758. [Google Scholar] [CrossRef]
  53. Dekking, F.M. A Modern Introduction to Probability and Statistics: Understanding Why and How; Springer Science & Business Media: Berlin, Germany, 2005. [Google Scholar] [CrossRef]
  54. Fargione, J.; Hill, J.; Tilman, D.; Polasky, S.; Hawthorne, P. Land clearing and the biofuel carbon debt. Science 2008, 319, 1235–1238. [Google Scholar] [CrossRef]
  55. Strassburg, B.B.; Kelly, A.; Balmford, A.; Davies, R.G.; Gibbs, H.K.; Lovett, A.; Miles, L.; Orme, C.D.L.; Price, J.; Turner, R.K.; et al. Global convergence of carbon storage and biodiversity in terrestrial ecosystems. Conserv. Lett. 2010, 3, 98–105. [Google Scholar] [CrossRef]
  56. Koh, L.P.; Ghazoul, J. Biofuels, biodiversity, and people: Understanding the conflicts and finding opportunities. Biol. Conserv. 2008, 141, 2450–2460. [Google Scholar] [CrossRef]
  57. Le Quéré, C.; Andrew, R.M.; Friedlingstein, P.; Sitch, S.; Pongratz, J.; Manning, A.C.; Korsbakken, J.I.; Peters, G.P.; Canadell, J.G.; Jackson, R.B.; et al. Global carbon budget 2017. Earth Syst. Sci. Data 2018, 10, 405–448. [Google Scholar] [CrossRef]
  58. Teixeira, A.M.; de Oliveira Arinelli, L.; de Medeiros, J.L.; Araújo, O.D.Q.F. Recovery of thermodynamic hydrate inhibitors methanol, ethanol and MEG with supersonic separators in offshore natural gas processing. J. Environ. Manag. 2019, 52, 166–186. [Google Scholar] [CrossRef]
  59. Velthof, G.L.; Nelemans, J.A.; Oenema, O.; Kuikman, P.J. Gaseous nitrogen and carbon losses from pig manure derived from different diets. J. Environ. Qual. 2005, 34, 698–706. [Google Scholar] [CrossRef]
  60. Berenguer, E.; Ferreira, J.; Gardner, T.A.; Aragão, L.E.O.C.; De Camargo, P.B.; Cerri, C.E.; Durigan, M.; De Oliveira, R.C.; Vieira, I.C.G.; Barlow, J. A large-scale field assessment of carbon stocks in human-modified tropical forests. Glob. Change Biol. 2014, 20, 3713–3726. [Google Scholar] [CrossRef]
  61. Lal, R. Soil carbon management and climate change. Carbon Manag. 2014, 4, 439–462. [Google Scholar] [CrossRef]
  62. Pettorelli, N.; Laurance, W.F.; O’Brien, T.G.; Wegmann, M.; Nagendra, H.; Turner, W. Satellite remote sensing for applied ecologists: Opportunities and challenges. J. Appl. Ecol. 2014, 51, 839–848. [Google Scholar] [CrossRef]
  63. Shah, S.A.R.; Zhang, Q.; Abbas, J.; Balsalobre-Lorente, D.; Pilař, L. Technology, urbanization, and natural gas supply matter for carbon neutrality: New evidence of environmental sustainability under the prism of COP26. Resour. Policy 2023, 82, 103465. [Google Scholar] [CrossRef]
  64. Cutler, D.R.; Edwards Jr, T.C.; Beard, K.H.; Cutler, A.; Hess, K.T.; Gibson, J.; Lawler, J.J. Random forests for classification in ecology. Ecology 2007, 88, 2783–2792. [Google Scholar] [CrossRef]
  65. Sorum, N.G.; Pal, D. Identification of the best machine learning model for the prediction of driver injury severity. Int. J. Inj. Control. Saf. Promot. 2024, 31, 360–375. [Google Scholar] [CrossRef]
  66. He, H.; Ma, Y. Class imbalance learning methods for support vector machines. In Imbalanced Learning: Foundations, Algorithms, and Applications; Wiley: Hoboken, NJ, USA, 2013; pp. 83–99. [Google Scholar] [CrossRef]
Figure 1. South Sumatra study area.
Figure 1. South Sumatra study area.
Atmosphere 16 00418 g001
Figure 2. Flowchart.
Figure 2. Flowchart.
Atmosphere 16 00418 g002
Figure 3. Carbon estimation maps in cropland: (a) cropland carbon estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Figure 3. Carbon estimation maps in cropland: (a) cropland carbon estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Atmosphere 16 00418 g003aAtmosphere 16 00418 g003b
Figure 4. N2O estimation maps in cropland: (a) cropland N2O estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Figure 4. N2O estimation maps in cropland: (a) cropland N2O estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Atmosphere 16 00418 g004
Figure 5. Carbon estimation maps in grassland: (a) grassland carbon estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Figure 5. Carbon estimation maps in grassland: (a) grassland carbon estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Atmosphere 16 00418 g005aAtmosphere 16 00418 g005b
Figure 6. N2O estimation maps in grassland: (a) grassland N2O estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Figure 6. N2O estimation maps in grassland: (a) grassland N2O estimation, (b) GBT, (c) RF, (d) CART, and (e) SVM.
Atmosphere 16 00418 g006aAtmosphere 16 00418 g006b
Table 1. Carbon estimation model results in cropland.
Table 1. Carbon estimation model results in cropland.
ModelRR2MAERMSE
GBT0.84300.71060.06320.0793
CART0.68930.47520.08340.1154
SVM0.49060.240629.423429.4289
RF0.76980.59270.06800.0894
Table 2. N2O estimation model results in cropland.
Table 2. N2O estimation model results in cropland.
ModelRR2MAERMSE
GBT0.84300.71060.000070.00009
CART0.68930.47520.00010.0001
SVM0.49060.240629.763529.7701
RF0.77220.59630.000080.0001
Table 3. Carbon estimation model results in grassland.
Table 3. Carbon estimation model results in grassland.
ModelRR2MAERMSE
GBT0.62640.39230.00940.0132
RF0.62250.38750.00990.0125
SVM0.36160.130829.721229.7274
CART0.42200.17810.01190.01791
Table 4. N2O estimation model results in grassland.
Table 4. N2O estimation model results in grassland.
ModelRR2MAERMSE
GBT0.51660.26690.000050.00006
RF0.62480.39040.000050.00006
SVM0.06070.003629.732929.7386
CART0.35980.12950.000060.00008
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Uyar, N.; Uyar, A. Assessing Climate Change Impacts on Cropland and Greenhouse Gas Emissions Using Remote Sensing and Machine Learning. Atmosphere 2025, 16, 418. https://doi.org/10.3390/atmos16040418

AMA Style

Uyar N, Uyar A. Assessing Climate Change Impacts on Cropland and Greenhouse Gas Emissions Using Remote Sensing and Machine Learning. Atmosphere. 2025; 16(4):418. https://doi.org/10.3390/atmos16040418

Chicago/Turabian Style

Uyar, Nehir, and Azize Uyar. 2025. "Assessing Climate Change Impacts on Cropland and Greenhouse Gas Emissions Using Remote Sensing and Machine Learning" Atmosphere 16, no. 4: 418. https://doi.org/10.3390/atmos16040418

APA Style

Uyar, N., & Uyar, A. (2025). Assessing Climate Change Impacts on Cropland and Greenhouse Gas Emissions Using Remote Sensing and Machine Learning. Atmosphere, 16(4), 418. https://doi.org/10.3390/atmos16040418

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop