Source-linked AI summary
Predicting and Mapping of Soil Organic Carbon Using Machine Learning Algorithms in Northern Iran
Mostafa Emadi, Ruhollah Taghizadeh-Mehrjardi, Ali Cherati, Majid Danesh, Amir Mosavi, Thomas Scholten
TL;DR
Soil organic carbon is important to soil functioning and carbon cycling, but SOC prediction accuracy depends on local geography, sampling, and auxiliary variables. This study evaluates six machine-learning algorithms for province-scale SOC mapping in Mazandaran, using genetic-algorithm feature selection; DNN delivers the strongest reported prediction performance.
Problem
SOC supports soil functioning, plant growth, and terrestrial carbon sequestration, while reliable province-scale mapping in northern Iran is needed because SOC varies with human activities and natural attributes.
Method
The study compares six machine-learning algorithms using 105 environmental auxiliary variables, with genetic algorithms selecting 35 predictors before model training.
Results
DNN outperformed the other models, explaining 65% of SOC variability in 10-fold cross-validation; precipitation, NDVI, MODIS day temperature, MrVBF, and land use were the most important predictors.
Takeaways & Limitations
DNN is reported as suitable for robust SOC mapping in Mazandaran province, while selected environmental predictors identify major spatial drivers of SOC variation.
Takeaways & Limitations
The apparent DNN advantage may depend on the large training dataset because deep learning is sensitive to training-data size.
Abstract
from arXiv · showhide
Estimation of the soil organic carbon content is of utmost importance in understanding the chemical, physical, and biological functions of the soil. This study proposes machine learning algorithms of support vector machines, artificial neural networks, regression tree, random forest, extreme gradient boosting, and conventional deep neural network for advancing prediction models of SOC. Models are trained with 1879 composite surface soil samples, and 105 auxiliary data as predictors. The genetic algorithm is used as a feature selection approach to identify effective variables. The results indicate that precipitation is the most important predictor driving 15 percent of SOC spatial variability followed by the normalized difference vegetation index, day temperature index of moderate resolution imaging spectroradiometer, multiresolution valley bottom flatness and land use, respectively. Based on 10 fold cross validation, the DNN model reported as a superior algorithm with the lowest prediction error and uncertainty. In terms of accuracy, DNN yielded a mean absolute error of 59 percent, a root mean squared error of 75 percent, a coefficient of determination of 0.65, and Lins concordance correlation coefficient of 0.83. The SOC content was the highest in udic soil moisture regime class with mean values of 4 percent, followed by the aquic and xeric classes, respectively. Soils in dense forestlands had the highest SOC contents, whereas soils of younger geological age and alluvial fans had lower SOC. The proposed DNN is a promising algorithm for handling large numbers of auxiliary data at a province scale, and due to its flexible structure and the ability to extract more information from the auxiliary data surrounding the sampled observations, it had high accuracy for the prediction of the SOC baseline map and minimal uncertainty.
1. Introduction
SOC is important for soil functioning, plant growth, and carbon cycling, motivating province-scale mapping. Because SOC varies complexly with local conditions, the study emphasizes site-specific machine-learning prediction and feature selection.
- SOC supports soil aggregation, water holding capacity, ion exchangeability, nutrient availability, and plant growth.
- Carbon in the upper meter of soil exceeds the carbon stored in the biosphere and atmosphere, underscoring SOC’s global-cycle importance.
- Digital soil mapping predicts SOC spatial variation from relationships with environmental auxiliary variables derived from terrain, remote sensing, and climate data.
- Machine-learning methods often outperform linear and geostatistical approaches by capturing nonlinear SOC relationships and extracting information for unsampled locations.
- Deep learning is considered for complex SOC environments where classical machine learning may encounter problems.
- Feature selection can reduce redundant and collinear predictors, but no universal method or worldwide best algorithm exists for SOC mapping.
2.1. Study Area
The study area is Mazandaran province in northern Iran, spanning diverse climates, landscapes, vegetation, and soil regimes. Its west-to-east precipitation gradient contributes to this environmental diversity.
- Mazandaran lies on Iran’s southern Caspian coast and borders the Alborz Mountains to the south.
- The province covers 2,388,179 ha and is predominantly forested, with dense, moderate, and low-density forests covering 39%, 4%, and 2% of its area.
- Annual temperature ranges from 18 °C on the coastal plain to below 8 °C in the highlands.
- Precipitation decreases from around 1400 mm in the west to around 450 mm in the east, producing diverse soil moisture and temperature regimes.
- Mazandaran contains Mollisols, Entisols, Inceptisols, Alfisols, and Ultisols across 12 suborders.
2.2. Soil Data
The SOC dataset comprises 1879 composite surface-soil samples assembled from university research projects and governmental soil surveys. Sample coordinates were recorded by GPS, and SOC was measured after laboratory preparation.
- The dataset contains 1879 composite surface-soil samples from two main sources.
- University research projects contributed 1055 samples, collected mostly from uncultivated areas using simple random sampling.
- Remaining samples came from soil surveys conducted by governmental agricultural and research organizations.
- Samples were air-dried, passed through a 2 mm sieve, and analyzed for SOC using the Walkley and Black wet-oxidation procedure.
- Sampling-site coordinates were recorded with a global positioning system.
2.3. Auxiliary Variables
The study initially considered 105 auxiliary predictors spanning remote sensing, terrain, climate, and categorical maps. These variables represent spatial information on vegetation, topography, precipitation, and land use for SOC modeling.
- The predictor pool contained 105 variables: 60 remote-sensing, 30 terrain, 10 climatic, and five categorical variables.
- Remote-sensing predictors were derived from Landsat 8 and MODIS imagery, including eight Landsat images from 2012–2016 aligned with soil-sampling dates.
- Terrain predictors were derived from 30 m SRTM DEM data and included slope, elevation, curvature, wetness, channel, and MrVBF indices.
- Ten climatic variables came from WorldClim version 2 at 1000 m spatial resolution, while categorical predictors came from five choropleth maps.
- Precipitation declines across the province, while NDVI ranges from −0.5 to above 0.8, indicating varied environmental conditions for SOC modeling.
- Figure 2 illustrates four auxiliary-variate types: MrVBF, NDVI, precipitation, and land use.
2.4. Selection of Auxiliary Variables Using Genetic Algorithms (GA)
The study uses a genetic algorithm to select a compact, informative set of auxiliary variables for SOC modeling, reducing redundancy and multicollinearity before machine-learning prediction.
- GA selects auxiliary data that are essential for SOC modeling and may improve prediction performance.It can accommodate nonlinear relationships between SOC and auxiliary variables.
- The GA represents candidate auxiliary-variable subsets as binary individuals, where 1 indicates selection and 0 indicates exclusion.Selection, crossover, and mutation generate new candidate subsets across generations.
- RMSE from random-forest fitting provides the fitness value used to select and reproduce candidate subsets.The two fittest individuals are selected for reproduction in each generation.
- The procedure used 10-fold cross-validation and 100 iterations, with population size 50, crossover rate 0.6, and mutation rate 0.001.
2.5. Machine Learning Techniques
The study compares six machine-learning techniques for SOC prediction, including tree-based, neural-network, support-vector, and boosting approaches, with DNNs modeled as feedforward networks.
- Machine Learning Techniques: Six algorithms were evaluated: SVM, ANN, Cubist regression tree, RF, XGBoost, and DNN.Their hyperparameters are summarized in Table 1.
- Regression Tree (Cubist): Cubist combines rule-based tree models with linear regression models fitted at tree leaf nodes.The study improves Cubist models by specifying the number of model trees and nearest neighbors.
- Random Forest (RF): RF aggregates predictions from many individual trees trained on bootstrap samples.Its rule-based trees model relationships between input and dependent variables.
- Extreme Gradient Boosting (XGBoost): XGBoost extends a strong learner from weak learners through boosting and uses supplemental training strategies to improve computation and reduce over-estimation.
- Deep Neural Networks (DNN): DNN is a feedforward network with an input vector, L hidden layers, and an output vector, without looping from output back to input.Weights and biases are adjusted during learning to identify patterns in the training data.
- Deep Neural Networks (DNN): The DNN implementation used the H2O package with a rectifier activation function and included specifications for hidden layers, network size, initialization, learning rate, and dropout regularization.
2.6. Evaluation of Algorithm Performance
Algorithm performance was assessed with 10-fold cross-validation using four criteria, and the best SOC predictor was selected by combining error minimization with agreement and explained-variance maximization.
- Evaluation criteria: The evaluation criteria were R2, Lin’s CCC, MAE, and RMSE.These metrics assess explained variance, concordance, absolute error, and squared error, respectively, as defined in the study.
- Algorithm selection: The four validation criteria were averaged across the 10-fold realizations for each prediction algorithm.
- Algorithm selection: The best algorithm was defined by the lowest MAE and RMSE together with the highest R2 and CCC.
2.7. Uncertainty Assessment
The study quantifies spatial uncertainty by summarizing SOC predictions across 10-fold realizations and mapping 90% confidence-interval limits for each raster cell.
- Uncertainty calculation: For each raster cell, the study calculates the mean and standard deviation of SOC predictions across the 10-fold model realizations.
- Confidence intervals: The 90% confidence interval is defined as the mean ±1.64 SD under an assumed normal distribution for each raster cell.Upper and lower confidence limits are mapped spatially.
3. Results and Discussions
Genetic-algorithm selection reduced 105 environmental variables to 35 predictors, with climate, vegetation, terrain, and land use among the strongest SOC-related factors. DNN consistently produced the most accurate and least uncertain province-scale predictions, while SOC varied across climate, soil, landform, and land-use classes.
- 3.2. Selected Auxiliary Data: 35 predictors were selected from 105 environmental variables using genetic algorithms with 10-fold cross-validation and 100 iterations.The best external RMSE estimate occurred at generation 0.86.
- 3.2. Selected Auxiliary Data: 14.9% of SOC spatial variability was attributed to precipitation, followed by NDVI at 12.5%, MODIS day temperature at 10.6%, MrVBF at 8.7%, and land use at 8.2%.Together, precipitation and mean annual temperature explained 18.9% of SOC variation.
- 3.2. Selected Auxiliary Data: NDVI ranked second because SOC depended strongly on vegetation-cover intensity and plant residues left after harvesting.MrVBF was associated with deposition of fine, organic-enriched particles and sediment in lower valleys, while land-use effects reflected recent conversion.
- 3.3. Machine Learning Performances: DNN explained 65% of SOC variability, exceeding SVM, ANN, Cubist, RF, and XGB, which explained 53%, 55%, 57%, 58%, and 57%, respectively.DNN also had the lowest mean MAE at 0.59% and RMSE at 0.75%, whereas SVM had the highest RMSE at 0.87%.
- 3.3. Machine Learning Performances: DNN achieved a CCC of 0.83, with most predictions following the 1:1 line except for slight underestimation at large observed SOC values.Its flexible structure extracted information from many auxiliary variables, although its performance was apparently aided by the large training dataset.
- 3.4. Spatial Prediction of SOC with Uncertainty Estimates: Approximately 88% of observations fell within DNN’s defined 90% confidence interval, and the resulting map showed higher SOC west-to-east patterns associated with precipitation and NDVI.The predicted maps were proposed as baseline products for temporal monitoring, inventories, accounting, and carbon-balance studies.
- 3.5. SOC Contents in Soil Classes and Geological Eras: SOC was highest in the udic soil-moisture regime at 3.85%, followed by aquic at 2.45% and xeric at 2.10%.Mesic soils averaged 2.75%, exceeding thermic soils at 2.20% and cryic soils at 1.25%.
4. Conclusions
The study identifies DNN as a robust province-scale estimator for SOC mapping in Mazandaran, using selected environmental predictors and producing a baseline map for carbon-sequestration research.
- 35 auxiliary predictors selected from 105 environmental variables identified precipitation, NDVI, MODIS day temperature, MrVBF, and land use as the most important predictors.The variables were derived from terrain attributes, remote sensing, and climatic data, with selection performed using a genetic algorithm.
- DNN outperformed the other five machine-learning algorithms for prediction uncertainty at the province scale.The study evaluated six algorithms and reported DNN as suitable for robust SOC estimation in Mazandaran province.
- SOC was lower in Cenozoic soils and accumulated more in developed Ultisols and Mollisols, virgin forests and rangelands, and udic soil-moisture classes.The conclusion links the highest SOC patterns to more developed soils, forest and rangeland land cover, and udic moisture conditions.
- The mesic soil-temperature class had significantly higher SOC, associated with high vegetation cover and biomass and probably lower carbon-decomposition rates.
- The predicted SOC map could serve as a baseline for further studies and projects related to carbon sequestration.
Appendix A
Appendix A lists the environmental auxiliary variables initially considered for SOC prediction, spanning terrain, remote sensing, climatic, land-use, soil, geology, physiography, and erosion data.
- Table A1 presents the environmental auxiliary data initially considered for predicting SOC distribution.
- Terrain variables include plan curvature, relative slope position, slope gradient, slope length, topographic wetness index, total insolation, upslope curvature, valley depth, and vertical distance to channel networks.
- Remote-sensing variables include Landsat spectral bands, vegetation indices, brightness and canopy indices, mineral and soil indices, and MODIS temperature and vegetation measures.
- Climatic variables include annual precipitation, precipitation seasonality, wettest- and driest-month precipitation, mean annual temperature, wind speed, water-vapor pressure, evapotranspiration, and aridity.
- Additional predictors include soil, geology, land-use, physiography, and erosion-class maps from the Soil and Water Research Institute of Iran.