Mapping Soil Organic Carbon Research Articles

Improved Surface Soil Organic Carbon Mapping of SoilGrids250m Using Sentinel-2 Spectral Images in the Qinghai–Tibetan Plateau

Soil organic carbon (SOC) is a critical indicator for the global carbon cycle and the overall carbon pool balance. Obtaining soil maps of surface SOC is fundamental to evaluating soil quality, regulating climate change, and global carbon cycle modeling. However, efficient approaches for obtaining accurate SOC information remain challenging, especially in remote or inaccessible regions of the Qinghai–Tibet Plateau (QTP), which is influenced by complex terrains, climate change, and human activities. This study employed field measurements, SoilGrids250m (SOC_250m, a spatial resolution of 250 m × 250 m), and Sentinel-2 images with different machine learning methods to map SOC content in the QTP. Four machine learning methods including partial least squares regression (PLSR), support vector machines (SVM), random forest (RF), and artificial neural network (ANN) were used to construct spatial prediction models based on 396 field-collected sampling points and various covariates from remote sensing images. Our results revealed that the RF model outperformed the PLSR, SVM, and ANN models, with a higher determination coefficient (R2 of 0.82 is from the training datasets) and the ratio of performance to deviation (RPD = 2.54). The selected covariates according to the variable importance in projection (VIP) were: SOC_250m, B2, B11, Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Vegetation Index (NDVI), B5, and Soil-Adjusted Total Vegetation Index (SATVI). The predicted SOC map showed an overall decrease in SOC content ranging from 69.30 g·kg−1 in the southeast to 1.47 g·kg−1 in the northwest. Our prediction showed spatial heterogeneity of SOC content, indicating that Sentinel-2 images were acceptable for characterizing the variability of SOC. The findings provide a scientific basis for carbon neutrality in the QTP and a reference for the digital mapping of SOC in the alpine region.

Comparison of Random Forest and Kriging Models for Soil Organic Carbon Mapping in the Himalayan Region of Kashmir

The knowledge about the spatial distribution of soil organic carbon stock (SOCS) helps in sustainable land-use management and ecosystem functioning. No such study has been attempted in the complex topography and land use of Himalayas, which is associated with great spatial heterogeneity and uncertainties. Therefore, in this study digital soil mapping (DSM) was used to predict and evaluate the spatial distribution of SOCS using advanced geostatistical methods and a machine learning algorithm in the Himalayan region of Jammu and Kashmir, India. Eighty-three soil samples were collected across different land uses. Auxiliary variables (spectral indices and topographic parameters) derived from satellite data were used as predictors. Geostatistical methods—ordinary kriging (OK) and regression kriging (RK)—and a machine learning method—random forest (RF)—were used for assessing the spatial distribution and variability of SOCS with inter-comparison of models for their prediction performance. The best fit model validation criteria used were coefficient of determination (R2) and root mean square error (RMSE) with resulting maps validated by cross-validation. The SOCS concentration varied from 1.12 Mg/ha to 70.60 Mg/ha. The semivariogram analysis of OK and RK indicated moderate spatial dependence. RF (RMSE = 8.21) performed better than OK (RMSE = 15.60) and RK (RMSE = 17.73) while OK performed better than RK. Therefore, it may be concluded that RF provides better estimation and spatial variability of SOCS; however, further selection and choice of auxiliary variables and higher soil sampling density could improve the accuracy of RK prediction.

Geospatial modeling and mapping of soil organic carbon and texture from spectroradiometric data in Nile basin

Rapid soil properties monitoring and modeling are crucial for sustainable soil management interventions. The analysis of soil properties in the laboratory using standard procedures is time-consuming and requires the use of chemicals. The research aimed to develop a regression model based on the visible-near-infrared-short-wave infrared (Vis-NIR-SWIR, 0.35–2.5 μm) spectral reflectance of soils in Aba Gerima (425 ha), a highland of the Ethiopian Nile Basin. Soil samples were obtained at a 0–20 cm depth from 101 plots in the study area and analyzed for soil organic carbon (SOC) and texture using standard laboratory procedures. Soil reflectance was measured using a spectroradiometer on the air-dried, ground, and sieved soils. ArcGIS software 10.5 was used to conduct the semi-variogram and geostatistical methods—Ordinary Kriging (OK) and Least Distance Weighting (IDW). Strong degrees of spatial dependence are depicted through semi-variogram models. This outcome illustrates a strong spatial dependence and continuity between nearby soil samples. The prediction accuracy of the model based on the values of coefficient of determination (R2) and root mean square error (RMSE) was explained. For clay content, the R2 and RMSE obtained in the Vis-NIR-SWIR were excellent (0.91) and 2.93. An R2 of 0.85 with an RMSE of 1.96 was obtained in predicting sand. The SOC model established indicated an R2 of 0.94 and an RMSEC of 0.01. The generated maps showed that the soil characteristics' spatial variability was adequate to forecast the values of soil fertility indicators in the research area and nearby regions. The soil maps may assist farmers and policymakers in identifying soil nutrient status, improving soil management strategies, rising crop productivity, and guaranteeing ecological integrity. Moreover, further study is required to corroborate our findings and improve the prediction performance of Vis-NIR-SWIR spectroscopy to help better land management interventions.

Digital Mapping of Soil Organic Carbon Based on Machine Learning and Regression Kriging.

In the last two decades, machine learning (ML) methods have been widely used in digital soil mapping (DSM), but the regression kriging (RK) model which combines the advantages of the ML and kriging methods has rarely been used in DSM. In addition, due to the limitation of a single-model structure, many ML methods have poor prediction accuracy in undulating terrain areas. In this study, we collected the SOC content of 115 soil samples in a hilly farming area with continuous undulating terrain. According to the theory of soil-forming factors in pedogenesis, we selected 10 topographic indices, 7 vegetation indices, and 2 soil indices as environmental covariates, and according to the law of geographical similarity, we used ML and RK methods to mine the relationship between SOC and environmental covariates to predict the SOC content. Four ensemble models-random forest (RF), Cubist, stochastic gradient boosting (SGB), and Bayesian regularized neural networks (BRNNs)-were used to fit the trend of SOC content, and the simple kriging (SK) method was used to interpolate the residuals of the ensemble models, and then the SOC and residual were superimposed to obtain the RK prediction result. Moreover, the 115 samples were divided into calibration and validation sets at a ratio of 80%, and the tenfold cross-validation method was used to fit the optimal parameters of the model. From the results of four ensemble models: RF performed best in the calibration set (R2c = 0.834) but poorly in the validation set (R2v = 0.362); Cubist had good accuracy and stability in both the calibration and validation sets (R2c = 0.693 and R2v = 0.445); SGB performed poorly (R2c = 0.430 and R2v = 0.336); and BRNN had the lowest accuracy (R2c = 0.323 and R2v = 0.282). The results showed that the R2 of the four RK models in the validation set were 0.718, 0.674, 0.724, and 0.625, respectively. Compared with the ensemble models without superimposed residuals, the prediction accuracy was improved by 0.356, 0.229, 0.388, and 0.343, respectively. In conclusion, Cubist has high prediction accuracy and generalization ability in areas with complex topography, and the RK model can make full use of trends and spatial structural factors that are not easy to mine by ML models, which can effectively improve the prediction accuracy. This provides a reference for soil survey and digital mapping in complex terrain areas.

Predicting spatiotemporal changes of surface soil organic carbon in China

AbstractEvaluating the spatiotemporal dynamics of soil organic carbon (SOC) is essential for better understanding the role of SOC in global C cycling. In this study, we report on the application of a two‐compartment model for mapping and examining SOC changes in China. We develop a SOC change prediction model combined with an environmental stratification approach and map SOC changes for the top 20‐cm soil depth in China at a grid resolution of 1 km. We perform a model comparison to quantify the effects of environmental stratification on model performance. A mean SOC increase of 0.63 kg m−2 was observed over the past 40 yr. Compared to the model without environmental stratification (the concordance correlation coefficient [CCC]: 0.34–0.40; RMSE: 2.16–2.33 kg m−2), we found that the models incorporating spatial or temporal stratification provide better performance. Notably, the model with spatiotemporal stratification achieved the better performance compared with results without stratification, with the prediction accuracy improved up to 57% in terms of CCC. Our results suggest that both the spatial‐ and temporal‐scale effects of environmental heterogeneity on SOC processes need to be accounted for in the prediction of SOC changes. Furthermore, the results indicate that the developed approach can be used to map SOC changes over large regions. This study also emphasizes that to achieve accurate predictions, it is necessary to choose the proper environmental stratification method to refine the study unit.

Mapping Soil Organic Carbon Content in Patagonian Forests Based on Climate, Topography and Vegetation Metrics from Satellite Imagery

Soil organic carbon (SOC) content supports several ecosystem services. Quantifying SOC requires: (i) accurate C estimates of forest components, and (ii) soil estimates. However, SOC is difficult to measure, so predictive models are needed. Our objective was to model SOC stocks within 30 cm depth in Patagonian forests based on climatic, topographic and vegetation productivity measures from satellite images, including Dynamic Habitat Indices and Land Surface Temperature derived from Landsat-8. We used data from 1320 stands of different forest types in Patagonia, and random forest regression to map SOC. The model captured SOC variability well (R² = 0.60, RMSE = 22.1%), considering the huge latitudinal extension (36.4° to 55.1° SL) and the great diversity of forest types. Mean SOC was 134.4 ton C ha−1 ± 25.2, totaling 404.2 million ton C across Patagonia. Overall, SOC values were highest in valleys of the Andes mountains and in southern Tierra del Fuego, ranging from 53.5 to 277.8 ton C ha−1 for the whole Patagonia region. Soil organic carbon is a metric relevant to many applications, connecting major issues such as forest management, conservation, and livestock production, and having spatially explicit estimates of SOC enables managers to fulfil the international agreements that Argentina has joined.

Exploring the Impacts of Data Source, Model Types and Spatial Scales on the Soil Organic Carbon Prediction: A Case Study in the Red Soil Hilly Region of Southern China

Rapid and accurate mapping of soil organic carbon (SOC) is of great significance to understanding the spatial patterns of soil fertility and conducting soil carbon cycle research. Previous studies have dedicated considerable efforts to the spatial prediction of SOC content, but few have systematically quantified the effects of environmental covariates selection, the spatial scales and the model types on SOC prediction accuracy. Here, we spatially predicted SOC content through digital soil mapping (DSM) based on 186 topsoil (0–20 cm) samples in a typical hilly red soil region of southern China. Specifically, we first determined an optimal covariate set from different combinations of multiple environmental variables, including multi-sensor remote sensing images (Sentinel-1 and Sentinel-2), climate variables and DEM derivatives. Furthermore, we evaluated the impacts of spatial resolution (10 m, 30 m, 90 m, 250 m and 1000 m) of covariates and the model types (three linear and three non-linear machine learning techniques) on the SOC prediction. The results of the performance analysis showed that a combination of Sentinel-1/2-derived variables, climate and topographic predictors generated the best predictive performance. Among all variables, remote sensing covariates, especially Sentinel-2-derived predictors, were identified as the most important explanatory variables controlling the variability of SOC content. Moreover, the prediction accuracy declined significantly with the increased spatial scales and achieved the highest using the XGBoost model at 10 m resolution. Notably, non-linear machine learners yielded superior predictive capability in contrast with linear models in predicting SOC. Overall, our findings revealed that the optimal combination of predictor variables, spatial resolution and modeling techniques could considerably improve the prediction accuracy of the SOC content. Particularly, freely accessible Sentinel series satellites showed great potential in high-resolution digital mapping of soil properties.

Impact of Atmospheric Correction Methods Parametrization on Soil Organic Carbon Estimation Based on Hyperion Hyperspectral Data

Visible Near infrared and Shortwave Infrared (VNIR/SWIR, 400–2500 nm) remote sensing data is becoming a tool for topsoil properties mapping, bringing spatial information for environmental modeling and land use management. These topsoil properties estimates are based on regression models, linking a key topsoil property to VNIR/SWIR reflectance data. Therefore, the regression model’s performances depend on the quality of both topsoil property analysis (measured on laboratory over-ground soil samples) and Bottom-of-Atmosphere (BOA) VNIR/SWIR reflectance which are retrieved from Top-Of-Atmosphere radiance using atmospheric correction (AC) methods. This paper examines the sensitivity of soil organic carbon (SOC) estimation to BOA images depending on two parameters used in AC methods: aerosol optical depth (AOD) in the FLAASH (Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes) method and water vapor (WV) in the ATCOR (ATmospheric CORrection) method. This work was based on Earth Observing-1 Hyperion Hyperspectral data acquired over a cultivated area in Australia in 2006. Hyperion radiance data were converted to BOA reflectance using seven values of AOD (from 0.2 to 1.4) and six values of WV (from 0.4 to 5 cm), in FLAASH and ATCOR, respectively. Then a Partial Least Squares regression (PLSR) model was built from each Hyperion BOA data to estimate SOC over bare soil pixels. This study demonstrated that the PLSR models were insensitive to the AOD variation used in the FLAASH method, with R2cv and RMSEcv of 0.79 and 0.4%, respectively. The PLSR models were slightly sensitive to the WV variation used in the ATCOR method, with R2cv ranging from 0.72 to 0.79 and RMSEcv ranging from 0.41 to 0.47. Regardless of the AOD values, the PLSR model based on the best parametrization of the ATCOR model provided similar SOC prediction accuracy to PLSR models using the FLAASH method. Variation in AOD using the FLAASH method did not impact the identification of bare soil pixels coverage which corresponded to 82.35% of the study area, while a variation in WV using the ATCOR method provided a variation of bare soil pixels coverage from 75.04 to 84.04%. Therefore, this work recommends (1) the use of the FLAASH AC method to provide BOA reflectance values from Earth Observing-1 Hyperion Hyperspectral data before SOC mapping or (2) a careful selection of the WV parameter when using ATCOR.

Soil total and organic carbon mapping and uncertainty analysis using machine learning techniques

• We conducted spatial prediction and uncertainty analysis of SOC and STC. • RFRK gave best predictions for SOC and STC. • RFRK produced higher local accuracy than RF. • Land use types, NDVI and elevation were key factors for SOC and STC predictions. Soil carbon is the largest terrestrial carbon pool. Reliable mapping of soil organic carbon (SOC) and soil total carbon content (STC) is essential for agricultural ecosystem management and carbon accounting under global warming conditions. This study was conducted at a fine scale and aimed to perform spatial distribution prediction and uncertainty mapping of SOC and STC and quantify the contribution of environmental variables affecting the variability of SOC and STC. Three machine learning models, namely, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Random Forest plus residuals Kriging (RFRK), were developed using 4345 agricultural topsoil samples and 16 environmental covariates. Mean absolute error (MAE), root mean square error (RMSE), Nash–Sutcliffe model efficiency coefficient (NSE), and Lin’s concordance correlation coefficient (LCCC) were used to evaluate the prediction global accuracy. Accuracy plot was used to quantify the uncertainty (i.e. local accuracy) of SOC and STC predictions. RFRK performed best with MAE, RMSE, NSE, and LCCC of 20.76%, 27.61%, 0.39, and 0.56 for SOC and 27.20%, 38.59%, 0.35, and 0.53 for STC, respectively. RF outperformed XGBoost in terms of NSE (0.33 vs 0.29 for SOC and 0.36 vs 0.32 for STC). Accuracy plots showed that RFRK produced higher local accuracy than RF both in quantifying the prediction uncertainty of SOC and STC. XGBoost performed excellently in the uncertainty estimation of SOC. Land use types, mean annual Normalized Difference Vegetation Index, and elevation were the top three important indicators in determining the spatial variability of SOC and STC. These results could provide inspiration and support for monitoring soil carbon in complex terrain areas.

Retraction: Accuracy Assessment of Kriging, artificial neural network, and a hybrid approach integrating spatial and terrain data in estimating and mapping of soil organic carbon.

Retraction: Accuracy Assessment of Kriging, artificial neural network, and a hybrid approach integrating spatial and terrain data in estimating and mapping of soil organic carbon.

Towards spatially continuous mapping of soil organic carbon in croplands using multitemporal Sentinel-2 remote sensing

Intensified human activities can augment soil organic carbon (SOC) losses from the world’s croplands, making SOC a highly dynamic parameter both in space and time. Sentinel-2 spectral imagery is well placed to capture the spatiotemporal variability of SOC, but its capability has only been demonstrated for agricultural regions mostly located in Europe. Furthermore, most studies so far only used single-date images that resulted in spatially non-continuous SOC maps, hampering their ability to aid multiscale SOC assessments. Here, we aim to achieve spatially continuous mapping of SOC in croplands, by creating multitemporal bare soil composites that maximize cropland coverage in two regions of varying agroecosystems and landscape structure in the Northeast China Chernozem region and the Belgian Loam Belt. Bare soil pixels were extracted via spectral index thresholding that excluded contaminated pixels from external perturbance. Multitemporal soil composites were then obtained by averaging over multiple single-date bare soil images that were selected within pre-determined optimal time-windows, corresponding to the region-specific crop sowing periods when best possible surface conditions were expected. Results show that the optimal time-window filter ensured selective inclusion of single-date images that themselves yielded stable and robust SOC predictions across multiple years. Spectral-based models developed from multitemporal composites consistently produced better or similar prediction accuracies than single-date images for both study regions (R2: 0.52–0.62; RMSE: 0.17–0.21 g 100 g−1), while also achieved maximum cropland coverage (>82 %). Bootstrap modelling demonstrated that SOC mapping via multitemporal Sentinel-2 data was associated with small uncertainties. Investigations into the significant spectral bands that contributed to the prediction of SOC suggested that, regardless of the study regions, the physical relationship between spectral bands and SOC that predominantly exists for laboratory spectra is largely translated into Sentinel-2 platforms. This study highlights the widespread applicability of multitemporal Sentinel-2 remote sensing for effective and high-resolution SOC mapping, in order to detect localized soil degradation as well as to inform regional cropland management in diverse agroecosystems.

Spatial prediction of organic carbon in German agricultural topsoil using machine learning algorithms

Abstract. As the largest terrestrial carbon pool, soil organic carbon (SOC) has the potential to influence and mitigate climate change; thus, SOC monitoring is of high importance in the frameworks of various international treaties. Therefore, high-resolution SOC maps are required. Machine learning (ML) offers new opportunities to develop these maps due to its ability to data mine large datasets. The aim of this study was to apply three algorithms commonly used in digital soil mapping – random forest (RF), boosted regression trees (BRT), and support vector machine for regression (SVR) – on the first German agricultural soil inventory to model the agricultural topsoil (0–30 cm) SOC content and develop a two-model approach to address the high variability in SOC in German agricultural soils. Model performance is often limited by the size and quality of the soil dataset available for calibration and validation. Therefore, the impact of enlarging the training dataset was tested by including data from the European Land Use/Cover Area frame Survey for agricultural sites in Germany. Nested cross-validation was implemented for model evaluation and parameter tuning. Grid search and the differential evolution algorithm were also applied to ensure that each algorithm was appropriately tuned . The SOC content of the German agricultural soil inventory was highly variable, ranging from 4 to 480 g kg−1. However, only 4 % of all soils contained more than 87 g kg−1 SOC and were considered organic or degraded organic soils. The results showed that SVR produced the best performance, with a root-mean-square error (RMSE) of 32 g kg−1 when the algorithms were trained on the full dataset. However, the average RMSE of all algorithms decreased by 34 % when mineral and organic soils were modelled separately, with the best result from SVR presenting an RMSE of 21 g kg−1. The model performance was enhanced by up to 1 % for mineral soils and by up to 2 % for organic soils. Despite the ability of machine learning algorithms, in general, and SVR, in particular, to model SOC on a national scale, the study showed that the most important aspect for improving the model performance was to separate the modelling of mineral and organic soils.

The possibility of spatial mapping of soil organic carbon content at three depths using easy-to-obtain ancillary data in a Mediterranean area

Background: Unlike most of Europe, Andalucía in southern Spain as a Mediterranean area still lacks digital maps of soil organic carbon (SOC) content at multiple depths, which can be generated by machine learning algorithms. The wide diversity of climate, geology, hydrology, landscape, topography, vegetation, and micro-relief data as easy-to-obtain covariates has facilitated the development of digital soil mapping (DSM). The purpose of this research is to model and map the spatial distribution of SOC at three depths, in an area of approximately 10000 km2 located in Seville and Cordoba Provinces, and to use R programming to compare two machine learning techniques (cubist and random forest) for developing SOC maps at multiple depths. Methods: Environmental covariates used in this research include nine derivatives from digital elevation models (DEM), three climatic variables, and 18 remotely-sensed spectral data (band ratios calculated by Landsat-8 Operational Land Imager ‘OLI’ and Sentinel-2A Multispectral Instrument ‘MSI’ in July 2019). In total, 300 soil samples from 100 points at three depths (0-25 cm, 25-50 cm, and 50-75 cm) were taken from existing literature. Both machine learning techniques were compared taking into account their accuracy using the goodness-of-fit criteria to predict SOC. Results: The findings showed that integrating the indices derived by Landsat-8 OLI and Sentinel-2A MSI satellite data had a better result than when satellite data was used separately. Conclusions: We obtained evidence that the resolution of satellite images is a key parameter in modelling and digital mapping.

Use of the time series and multi-temporal features of Sentinel-1/2 satellite imagery to predict soil inorganic and organic carbon in a low-relief area with a semi-arid environment

ABSTRACT Accurate mapping of soil organic carbon (SOC) and inorganic carbon (SIC) contents at regional scales can be very important for sustainable agriculture and soil management. Low variation in terrain attributes (classically used for digital soil mapping) at low relief areas calls for additional spatial data to explain soil variability. The main objective of this study was to evaluate the predictive capability of Sentinel-1 (radar) and Sentinel-2 (optical) time series and statistics, summarized as multi-temporal features (MTF) to improve the spatial predictions of SOC and SIC in Ghorveh plain, located in Kurdistan Province, Western Iran. A systematic grid sampling was then employed to collect 150 soil surface samples (0–30 cm) for SOC and SIC measurements. We applied boosted regression trees (BRT) and random forest (RF) to predict SOC and SIC contents by using covariate sets compiled from radar and optical time series and topographic attributes. Model performance, evaluated by 10-fold cross-validation, showed that RF using the covariate set containing time series of Sentinel-1, Sentinel-2 and topographic attributes performed the best in predicting SOC (RMSE = 0.23, ME = 0.005, R2 = 0.29). On the other hand, for SIC, the covariate set containing MTF of Sentinel-1, Sentinel-2 and topographic attributes ranked the best with BRT (RMSE = 0.77, ME= −0.001, R2 = 0.48). The study indicates that using the time series and MTF from multiple dates of remote sensing data with topographic attributes results in improved predictions. However, model performance for SIC and SOC was moderate to poor, respectively. Therefore more substantial studies would be required to verify if the computational effort is likely justified by an increase in accuracy in general.

A CNN-LSTM Model for Soil Organic Carbon Content Prediction with Long Time Series of MODIS-Based Phenological Variables

The spatial distribution of soil organic carbon (SOC) serves as critical geographic information for assessing ecosystem services, climate change mitigation, and optimal agriculture management. Digital mapping of SOC is challenging due to the complex relationships between the soil and its environment. Except for the well-known terrain and climate environmental covariates, vegetation that interacts with soils influences SOC significantly over long periods. Although several remote-sensing-based vegetation indices have been widely adopted in digital soil mapping, variables indicating long term vegetation growth have been less used. Vegetation phenology, an indicator of vegetation growth characteristics, can be used as a potential time series environmental covariate for SOC prediction. A CNN-LSTM model was developed for SOC prediction with inputs of static and dynamic environmental variables in Xuancheng City, China. The spatially contextual features in static variables (e.g., topographic variables) were extracted by the convolutional neural network (CNN), while the temporal features in dynamic variables (e.g., vegetation phenology over a long period of time) were extracted by a long short-term memory (LSTM) network. The ten-year phenological variables derived from moderate-resolution imaging spectroradiometer (MODIS) observations were adopted as predictors with historical temporal changes in vegetation in addition to the commonly used static variables. The random forest (RF) model was used as a reference model for comparison. Our results indicate that adding phenological variables can produce a more accurate map, as tested by the five-fold cross-validation, and demonstrate that CNN-LSTM is a potentially effective model for predicting SOC at a regional spatial scale with long-term historical vegetation phenology information as an extra input. We highlight the great potential of hybrid deep learning models, which can simultaneously extract spatial and temporal features from different types of environmental variables, for future applications in digital soil mapping.

High-resolution soil organic carbon mapping at the field scale in Southern Belgium (Wallonia)

• A Gradient Boosting machine method predicted SOC contents for agricultural fields in Wallonia with 77% accuracy. • We constructed a covariate to quantify the carbon input to soils integrating main crops and cover crops. • The final map shows patterns of SOC content at field scale with distinctive differences between adjacent fields. • Different environmental covariates shape the distribution of SOC at different spatial scales. Accurate soil organic carbon content estimation is critical as a proxy for carbon sequestration, and as one of the indicators for soil health. Here, we collected 497 soil samples during 2015 and 2019, as well as five environmental covariates (organic carbon (OC) input from the crops, normalized difference vegetation index (NDVI), elevation, clay content and precipitation) at a resolution of 30 m. We then aggregated these to represent agricultural fields and compiled a soil organic carbon (SOC) content map for the agricultural soils of Wallonia using Gradient Boosting Machine. We calculated OC input from both main crops and cover crops for each individual field. As the cover crops do not occur in the agricultural census, we identified cover crops based on long time-series of NDVI values obtained from the Google Earth Engine platform. The quality of the SOC predictions was assessed by validation data and we obtained an R 2 of 0.77. The Empirical Mode Decomposition indicated that OC input and NDVI were the dominant factors at field scale, whereas the remaining covariates determined the distribution of SOC at the scale of the entire Walloon region. The SOC map showed an overall northwest to southeast trend i.e. an increase in SOC contents up to the Ourthe river followed by a decrease further to the South. The map shows both regional trends in SOC and effects of differences in land use and/or management (including crop rotation and frequency of cover crops) between individual fields. The field-scale map can be used as a benchmark and reference to farmers and agencies in maintaining SOC contents at an appropriate level and optimizing decisions for sustainable land use.

Proportional allocation with soil depth improved mapping soil organic carbon stocks

Soil organic carbon (SOC) is vital to the assessment of land quality, management of farmland and ecological environment, and carbon cycle. A more accurate spatial prediction of multilayer soil organic carbon density (SOCD) can contribute to a better interpretation of the changes in multilayer SOC stocks and carbon dynamics. However, previous mapping techniques still have limitations, such as ignoring the relationship of profile depths, not further taking advantage of vertical distribution and surface categorical information. In addition, it is unclear whether it is better to model each depth interval of SOC separately or to model the total layer and then allocate it. Here, we propose two new methods based on the proportional allocation of soil depth for multilayer mapping: vertical log-ratio method (VLR) of SOCD by applying the percentage of SOCD data and isometric log-ratio (ILR) transformation, and vertical distribution method (VD) of SOCD by considering different land-use types. We compared five methods, including the two new methods, the exponential and equal-area spline functions, and independent modeling without depth information. We combined these five methods with the generalized linear model (GLM) and random forest (RF) to produce predictions of the Sanjiang Plain, northeastern China. The results demonstrated that SOCD did not always decrease with increasing soil depth, and classification of SOCD vertical distribution features needs to be considered by combining with soil depths. For accuracy assessment, the exponential mode with both GLM and RF over-calculated the predicted values and performed poorly, indicating that the blind use of depth information increased the prediction error. The spline function prediction was scarcely better than that of independent modeling. The proportional allocation methods performed better than other separate modeling methods for accuracy and interpretability with GLM or RF, especially for the middle and surface layers. The GLM generated more aggregated predictions than the RF, losing the distribution pattern of the original data. Therefore, we recommend RF combined with proportional allocation methods for spatial SOCD prediction in large-scale study areas. We calculated the SOC stocks in the Sanjiang Plain using our new methods, which were more reasonable compared with those of previous studies and had the advantages of in-depth information, environmental variable selection, and model optimization. Our findings provide not only other perspectives for SOCD mapping, with more fully integrated depth information and more accurate assessment of multilayer SOC stocks, but also provide guidance for the evaluation of land quality, farmland, and ecological environmental management.

Three-dimensional mapping of soil organic carbon using soil and environmental covariates in an arid and semi-arid region of Iran

This study focus on modeling and mapping soil organic carbon (SOC) at high spatial resolution and at four standard depths in an arid and semi-arid region of Iran. The SOC data includes 850 soil samples collected from 278 observation profiles. In parallel, a wide range of environmental covariates (n=62) were obtained from multiple sources. Six individual machine learning (ML) algorithms were compared to modeling and predicting SOC. Two scenarios were investigated. The first one accounts for soil and environmental covariates (S1) while the second one only accounts for environmental covariates (S2). Our results show that accounting for soil variables in the prediction (S1) leads to a twofold increase of R2 for all ML algorithms, while random forest (RF) outperformed the other ML approaches at all depths. Whenever possible, using additionally the soil variables that are at hand in a study area is thus beneficial for improving SOC predictions.

National soil organic carbon map of agricultural lands in Nepal

Reliable and accurate soil organic carbon (SOC) maps are needed to monitor and improve SOC status in croplands and for agro-environmental applications. Topsoil (0–20 cm) SOC content from agricultural lands was predicted and mapped with quantified uncertainty across Nepal using state-of-the-art soil mapping techniques. Altogether 25,312 SOC observations were used to build and evaluate prediction models derived from four machine learning algorithms, namely Random Forest (RF), Cubist, Extreme Gradient Boosting (XGB) and Support Vector Machines. Twenty two environmental variables were selected as SOC predictors based on their correlation with measured SOC contents and non-collinearity with other predictors. The predictive performance of these models was compared using calibration (80% observations) and validation (20% observations) datasets. The performance of the models was also compared against a global SOC dataset compiled by International Soil Reference and Information Centre (ISRIC). The best model among the four algorithms was used to map and quantify the spatial distribution of SOC contents, and the model uncertainty was assessed with the Quantile Regression Forests technique with standard deviation representing prediction uncertainty. The RF model performed the best among all tested models, closely followed by the Cubist, and then the XGB model. The predictive performance of all of these models was better than the global SOC prediction from ISRIC. This study provides a baseline map for the topsoil SOC contents from the croplands in Nepal, and also provides a reference for similar SOC mapping studies.

Remote estimates of soil organic carbon using multi-temporal synthetic images and the probability hybrid model

Soil organic carbon (SOC) plays a key role in soil function, ecosystem services, and the global carbon cycle. Digital SOC mapping is essential for agricultural production management. Digital SOC mapping based on multi-source remote sensing data has been integrated well into prediction models and methodological approaches on different mapping scales. However, the mixed use of synthetic images and the application of the hybrid model considering the soil classification probability are rare. Here, we propose a probability hybrid model to estimate and map SOC content and distribution. Multi-temporal synthetic images were used to build the probability hybrid model. One hundred forty topsoil samples were collected in Suihua, a city in a typical black soil region in China. Cloud-free Sentinel-2 images were acquired from the bare soil periods between 2018 and 2021. Minimum, maximum, mean, and median synthetic images were calculated using single images from the same period. The random forest and support vector machine models were used for discriminating soil class and calculating soil classification probability, and then random forest regression model was applied to SOC mapping. Soil class mapping and classification probability were performed for Phaeozems, Chernozems, and Cambisols of the World Reference Base for Soil Resources (WRB) based on soil texture, climatic factors, and the normalized differential vegetation index (NDVI). Based on soil class mapping results, global models using single temporal images and multi-temporal images, the hybrid model using multi-temporal images for the three soil classes as well as a probability hybrid model using multi-temporal images were built and compared their SOC predictions. Our study showed that (1) Phaeozems, Chernozems and Cambisols could be classified accurately, and the overall validation accuracy of random forest model was 91.67%. (2) The correlations of SOC with bands and band indices improved using multi-temporal images, and the use of mixed synthetic images was better than the use of only one synthetic image. (3) The hybrid model performed far better than the global models, and the probability hybrid model led to the highest prediction accuracy, a validation R2 of 0.77 and an RMSE of 2.30 g kg−1. (4) The probability hybrid model was more accurate than the original hybrid model for digital SOC mapping, and the SOC distribution in boundary regions was smoother and more continuous. Our results suggest that the probability hybrid model has a large potential for SOC prediction and mapping.