Soil Organic Carbon Prediction Research Articles

Monitoring Land Management Practices Using Vis–NIR Spectroscopy Provides Insights into Predicting Soil Organic Carbon and Limestone Levels in Agricultural Plots

This study aimed to establish sound relationships between soil properties of agricultural land in central Spain and their spectral attributes to contribute to deriving an indicator for sustainable farm management. Sixteen plots, managed under various conditions, eight with traditional tillage and eight with other alternative managements, were selected to gather soil samples representing three predominant soil orders (Leptosols, Cambisols, and Luvisols). Soil sampling was conducted from depths ranging from 0 to 30 cm (0–50; 5–10, 10–20, and 20–30 cm), ensuring a broad spectrum of sample variability across different times and locations. The reflectance of soil 64 soil samples was measured within the range of 400 to 900 nm, and the corresponding concentrations of soil organic carbon and majoritarian minerals, calcium carbonate, quartz, phyllosilicates, K-Feldspar, and plagioclase were determined for each sample. Partial least squares analysis was employed to construct prediction models using a calibration dataset comprising 66% of randomly selected samples. The remaining 33% of samples were utilized for model validation. The prediction models for the measured soil chemical properties yielded R2 values ranging from 0.14 to 0.79. Only SOC and limestone provided accurate prediction models. These findings hold promise for developing a soil health indicator tailored for site-specific crop management. However, the complex composition of soil organic carbon and calcium carbonate in certain soils underscores the importance of careful interpretation and validation of remote sensing data, as well as the need for advanced modeling approaches that can account for the interactions between multiple soil constituents.

Synergistic Use of Multi-Temporal Radar and Optical Remote Sensing for Soil Organic Carbon Prediction

Exploring soil organic carbon (SOC) mapping is crucial for addressing critical challenges in environmental sustainability and food security. This study evaluates the suitability of the synergistic use of multi-temporal and high-resolution radar and optical remote sensing data for SOC prediction in the Kaffrine region of Senegal, covering over 1.1 million hectares. For this purpose, various scenarios were developed: Scenario 1 (Sentinel-1 data), Scenario 2 (Sentinel-2 data), Scenario 3 (Sentinel-1 and Sentinel-2 combination), Scenario 4 (topographic features), and Scenario 5 (Sentinel-1 and -2 with topographic features). The findings from comparing three different algorithms (Random Forest (RF), XGBoost, and Support Vector Regression (SVR)) with 671 soil samples for training and 281 samples for model evaluation highlight that RF outperformed the other models across different scenarios. Moreover, using Sentinel-2 data alone yielded better results than using only Sentinel-1 data. However, combining Sentinel-1 and Sentinel-2 data (Scenario 3) further improved the performance by 6% to 11%. Including topographic features (Scenario 5) achieved the highest accuracy, reaching an R2 of 0.7, an RMSE of 0.012%, and an RPIQ of 5.754 for the RF model. Applying the RF and XGBoost models under Scenario 5 for SOC mapping showed that both models tended to predict low SOC values across the study area, which is consistent with the predominantly low SOC content observed in most of the training data. This limitation constrains the ability of ML models to capture the full range of SOC variability, particularly for less frequent, slightly higher SOC values.

Mapping sub-surface distribution of soil organic carbon stocks in South Africa's arid and semi-arid landscapes: Implications for land management and climate change mitigation

Soil organic carbon (SOC) stocks are critical for land management strategies and climate change mitigation. However, understanding SOC distribution in South Africa's arid and semi-arid regions remains a challenge due to data limitations, and the complex spatial and sub-surface variability in SOC stocks driven by desertification and land degradation. Thus, to support soil and land-use management practices as well as advance climate change mitigation efforts, there is an urgent need to provide more precise SOC stock estimates within South Africa's arid and semi-arid regions. Hence, this study adopted remote-sensing approaches to determine the spatial sub-surface distribution of SOC stocks and the influence of environmental co-variates at four soil depths (i.e., 0-30 cm, 30-60 cm, 60-100 cm, and 100-200 cm). Using two regression-based algorithms, i.e., Extreme Gradient Boosting (XGBoost) and Random Forest (RF), the study found the former (RMSE values ranging from 7.12 t/ha to 29.55 t/ha) to be a superior predictor of SOC in comparison to the latter (RMSE values ranging from 7.36 t/ha to 31.10 t/ha). Nonetheless, both models achieved satisfactory accuracy (R2 ≥ 0.52) for regional-scale SOC predictions at the studied soil depths. Thereafter, using a variable importance analysis, the study demonstrated the influence of climatic variables like rainfall and temperature on SOC stocks at different depths. Furthermore, the study revealed significant spatial variability in SOC stocks, and an increase in SOC stocks with soil depth. Overall, these findings enhance the understanding of SOC dynamics in South Africa's arid and semi-arid landscapes and emphasizes the importance of considering site specific topo-climatic characteristics for sustainable land management and climate change mitigation. Furthermore, the study offers valuable insights into sub-surface SOC distribution, crucial for informing carbon sequestration strategies, guiding land management practices, and informing environmental policies within arid and semi-arid environments.

Coupled Vis-NIR spectroscopy with chemometrics strategy for soil organic carbon prediction in the Agro-pastoral Transitional zone of northwest China

Rapidly and accurately grasp the change of soil organic carbon content in farmland, which is of great significance in guiding the timely and effective mastery of farmland soil fertility and improvement of soil physical properties. In this study, an ASD FieldSpec 4 spectrometer was used to collect spectral reflectance data on 128 agricultural soil samples taken from Jingbian County, Yulin City, Shaanxi Province, China. Firstly, descriptive statistics of the SOC in the study area were performed, and secondly, after 10 spectral transformations were performed, the correlation analysis and the Boruta algorithm were used to extract the characteristic wavebands of soil organic carbon, respectively, in order to reduce the redundancy of the data. Finally, by comparing the accuracies of different strategies, we constructed a spectral prediction model of soil organic carbon in farmland of the Northwest Agricultural and Animal Husbandry Intertwined Zone that integrates the optimal preprocessing, feature selection strategy and modelling method. The results indicate that: 1) The mean SOC content of the farmland in the study area was low and at the nutrient deficient level, with the standard errors and coefficients of variation for the modelling and validation sets were 1.596 g kg−1, 1.457 g kg−1, 54 % and 52 %, respectively; 2) The shape and trend of spectral special curves with different SOC contents show consistency, and the SOC content is negatively correlated with spectral reflectance; 3) CA selects more feature bands, but the feature bands are more homogeneous, while the Boruta algorithm can effectively remove irrelevant variables and improve the SOC feature selection effect; 4) The SOC prediction model based on Boruta-FD-RF can be better for soil organic carbon estimation, with R2 of 0.899 and 0.748 for the training set and validation set, respectively, RMSE of 1.432 g kg−1 and 1.967 g kg−1, and RPD of 2.557 and 1.647, respectively. The results show that the SOC model established by integrating optimal spectral pre-processing, feature selection strategy and chemometrics strategy has obvious improvement in prediction accuracy and stability, and this study provides an important reference for the fast and accurate estimation of SOC content in farmland of Agro-pastoral Transitional zone in northwest China.

Subsetting reduces the error of MIR spectroscopy models for soil organic carbon prediction in the U.S. Great Plains

High demand for soil organic carbon data to support soil health and climate change mitigation efforts must be met with rapid, accurate, and inexpensive measurement methods. Mid-infrared spectroscopy shows promise as an alternative to conventional soil carbon analysis; however, its practicality depends on the construction and efficient use of soil spectral libraries. Subsetting, a calibration optimization technique, has potential to reduce model prediction errors. Nevertheless, the effectiveness of different subsetting criteria has yet to be well explored. This study assessed whether several subsetting criteria would result in calibration models with reduced error in the prediction of soil organic carbon content compared to models constructed from a full spectral dataset. A mid-infrared spectral library composed of Nebraska and Kansas soil samples was subset by (i) a nested wetland criterion, (ii) soilscapes, (iii) presence or absence of carbonates, and (iv) a combination of soilscape and carbonates. Partial least squares regression was used to construct all calibration models. Predictive performance of the subset models was compared to that of their corresponding full set model using several statistical metrics. Subsetting by wetlands reduced model error by 22 and 56%. Subsetting by soilscape yielded a 13 to 55% reduction in model error, while presence or absence of carbonates reduced model error by 21 and 46%. Five of the eight combination soilscape and carbonate subset models reduced model error by 14 to 51%. Overall, subsetting by soilscape or carbonate presence proved effective in improving model performance, with combination subsets proving beneficial under specific calibration set conditions.

National-scale mapping of soil-thickness probability in hilly and mountainous areas of Japan using legacy and modern soil survey

Soil thickness is a crucial parameter for local, regional and globalscale assessments, including those related to carbon cycles, water retention capacity, and landslide risk in the context of biogeochemical, soil physical, and slope stability models. This study had two primary objectives: 1) to predict maps depicting the probability of the A-horizon thickness exceeding 15 cm (AT15) and the A and B-horizon thickness exceeding 75 cm (soil thickness, ST75), across the expansive mountainous regions of Japan. 2) To evaluate the enhancement of model performance through the preselection of the maximum value and implementation of enhanced modeling using gridded modern soil survey data to utilize geographically clustered data. Initially, the study modeled the SOC, bulk density, and coarse fraction proportion using data from a modern soil survey of 0–30 cm. Subsequently, we employed a random forest model for AT15 and ST75 using a dataset of geographically clustered legacy soil profiles as training data, and topographic, vegetation, climatic, and volcanic attributes, and the predicted soil organic carbon and related properties as explanatory variables. We reduced the number of legacy soil profiles in the densely sampled areas by selecting the maximum soil thickness before modeling. The study assessed the final model performance at the national scale, with the accuracy and F1 scores reaching 0.691 and 0.754 for AT15 and 0.600 and 0.511 for ST75, respectively. Maximum value selection proved effective for modeling ST75 but not AT15. Although an enhanced model incorporating data from legacy and modern surveys did not improve accuracy at the national level, it did improve recall in sparsely sampled areas, for ST75. The resulting spatial pattern of the predicted ST75 effectively captured variations in soil thickness associated with topography, geology, and tephra deposition. However, ST75, representing pedological soil thickness, tended to be strongly underestimated compared to geotechnical soil thickness, including the C-horizon, especially in areas affected by tephra deposition. This study represents a significant contribution as it delivers Japan's first national-scale soil thickness map and an approach for utilizing clustered legacy data. It is expected to find practical applications in estimating water storage capacity for forest management and other relevant purposes.

On the impact of soil texture on local scale organic carbon quantification: From airborne to spaceborne sensing domains

Soil organic carbon (SOC) distribution and interaction with light is influenced by soil texture parameters (clay, silt and sand), which makes SOC prediction complicated, especially in samples with considerable pedological variability. Hence, understanding the relationship between SOC and soil texture is important within the context of SOC prediction using remote sensing data. The main objective of this study was to find the impact of soil texture on the performance of local SOC prediction models that were developed on Sentinel-2 (S2) multispectral and CASI/SASI (CS) hyperspectral airborne data as the main predictor variables. One approach to that objective was to lowering the texture variance by stratification of the samples. Therefore, soil samples collected from four agricultural sites in the Czech Republic were segregated based on the i) site-based and ii) texture-based stratification strategies. Random forest (RF) models were then developed on all stratified classes with and without considering the soil texture parameters as predictor variables and results were compared with those obtained by the RF models developed on the non-stratified (NS) samples. Both stratification strategies provided more homogeneous classes, which enhanced the accuracy of SOC prediction, compared to using the NS samples. In addition, the texture-based RF models yielded higher accuracy predictions than the site-based ones. Except sand, adding texture parameters to the main predictors improved accuracy of the models, so that the highest prediction performance was obtained by a texture-based model developed on clay added CS data. Overall, texture-based stratification could significantly enhance the SOC prediction, when the texture parameters were added to the S2 and CS data as the main predictor variables.

Prediction of soil organic carbon using machine learning techniques and geospatial data for sustainable agriculture

Sustainable agriculture revolves around soil organic carbon (SOC), which is essential for numerous soil functions and ecological attributes. Farmers are interested in conserving and adding additional soil organic carbon to certain fields in order to improve soil health and productivity. The relationship between soil and environment that has been discovered and standardized throughout time has enhanced the progress of digital soil-mapping techniques; therefore, a variety of machine learning techniques are used to predict soil properties. Studies are thriving at how effectively each machine learning method maps and predicts SOC, especially at high spatial resolutions. To predict SOC of soil at a 30 m resolution, four machine learning models—Random Forest, Support Vector Machine, Adaptive Boosting, and k-Nearest Neighbour were used. For model evaluation, two error metrics, namely R2 and RMSE have been used. The findings demonstrated that the calibration and validation sets’ descriptive statistics sufficiently resembled the entire set of data. The range of the calculated SOC content was 0.06 to 1.76 %. According to the findings of the study, Random Forest showed good results for both cases, i.e. evaluation using cross validation and without cross validation. Using cross validation, RF confirmed highest R2 as 0.5278 and lowest RMSE as 0.1683 for calibration dataset while without cross validation it showed R2 as 0.8612 and lowest RMSE as 0.0912 for calibration dataset. The generated soil maps will help farmers adopt precise knowledge for decisions that will increase farm productivity and provide food security through the sustainable use of nutrients and the agricultural environment.

Prediction of soil organic carbon in black soil based on a synergistic scheme from hyperspectral data: Combining fractional-order derivatives and three-dimensional spectral indices

Monitoring soil organic carbon (SOC) content is crucial for climate change mitigation and sustaining ecological balance. Despite the unparalleled advantages of hyperspectral data in capturing nuanced variations in soil properties through its high spectral resolution, effectively extracting useful features from numerous bands via spectral processing techniques remains a formidable challenge. This study proposes an integrated approach combining fractional-order derivative (FOD) technique and optimal band combination algorithm using ZY1-02D satellite hyperspectral data to estimate SOC in Northeast China’s Black soil region. Three modeling strategies were compared: (1) FOD-transformed reflectance (FOD spectra), (2) FOD spectra with traditional 2D spectral indices (FOD + 2D SI), and (3) FOD spectra with new 3D spectral indices (FOD + 3D SI). These strategies were implemented using the random forest model with the aim of the optimal SOC prediction. Results showed that the application of FOD technique for spectral transformation effectively addressed the challenges posed by overlapping peaks and baseline drift inherent in the original spectral reflectance. Additionally, FOD transformation enhanced subtle soil spectral features and yielded more pronounced spectral variations with increasing fractional order, as compared to the original spectral data and conventional integer-order derivatives (i.e., first and second-order derivatives). However, as the FOD order continued to increase beyond 1.4, the spectral curve exhibited amplified noise and distortion, thereby adversely impacting subsequent model development. The 3D spectral indices correlate more robustly with SOC than 2D indices. The model that combines 0.6-order FOD and 3D spectral indices achieved the best accuracy (R2 = 0.66, RMSE = 2.99 g/kg and MAE = 2.42 g/kg), significantly outperforming the models built by 0.6-order FOD spectra (R2 = 0.48, RMSE = 3.65 g kg−1, and MAE = 2.93 g kg−1) and 0.8-order FOD + 2D SI modeling strategy (R2 = 0.55, RMSE = 3.54 g kg−1, and MAE = 2.85 g kg−1). These findings indicated that FOD and 3D spectral indices exhibit superior synergistic performance in SOC prediction, demonstrating their feasibility and providing valuable insights for large-scale soil property prediction and mapping using satellite hyperspectral data.

Estimation of soil organic carbon by combining hyperspectral and radar remote sensing to reduce coupling effects of soil surface moisture and roughness

Soil organic carbon (SOC) is important in the global carbon cycle. Accurate estimation of SOC content in cultivated land is a prerequisite for evaluating the carbon sequestration potential and quality of soils. However, existing SOC prediction studies based on hyperspectral remote sensing neglect the spectral response of the physical properties of surface soil, leading to inadequate model generalization. With the exponential growth of remote sensing data, the development of pixel-level soil spectral correction methods based on multi-source remote sensing data has become an interesting and challenging topic. This method aims to minimize the effect of soil physical properties on spectra, thus addressing the poor spatiotemporal transferability of SOC prediction models due to uncertain variations in surface soil physical properties. In this study, a soil spectral correction strategy is constructed using satellite hyperspectral image (HSI) and synthetic aperture radar (SAR) images through multi-order polynomial regression and convolutional neural networks. This strategy considers soil physical variables such as soil moisture (SM) content and root mean square height (RMSH) of soil surface roughness. The soil spectral correction model and SOC content prediction model were established using 80 soil samples collected from Site 1. Afterward, the performance and transferability of both models were verified using the remaining 25 samples from Site 1 and 50 samples from Site 2. The results showed that: 1) The effect of SM and RMSH on the soil pixel spectrum can be significantly reduced after correcting HSI using soil spectral correction strategy. The correlation coefficients between the corrected pixel spectrum and the ground-based spectrum increase by over 60 % compared with those between the original spectrum and the ground-based spectrum. 2) Soil spectral correction improves the prediction accuracy and mapping capability of HSI for SOC content, with the highest RP2 of 0.743 and RMSEP of 3.455 g/kg at Site 1. 3) Compared with the original HSI-based SOC prediction model, the soil spectral correction strategy based on multi-order polynomial and convolutional neural network reduced the RMSEP of SOC prediction results at Site 2 by 5.082 g/kg and 5.454 g/kg, and the RP2 increased by 0.390 and 0.409, respectively. 4) When predicting SOC content from raw HIS, SM and RMSH contribute to more than 60 % of the bias, with SM having a larger bias than RMSH. The findings of this study emphasize the influence of soil physical properties on SOC prediction and contribute to the existing research on SOC mapping using HSI and SAR data.

Coupling of microbial-explicit model and machine learning improves the prediction and turnover process simulation of soil organic carbon

Coupling of microbial-explicit model and machine learning improves the prediction and turnover process simulation of soil organic carbon

Evaluation of RothC model for predicting soil organic carbon stock in north-west Ethiopia

Assessing soil organic carbon (SOC) is vital for water retention, soil health, nutrient cycling, greenhouse gas emissions, and pollutant reduction and thereby contributes to sustainable agricultural production and food security. Thus, using long-term climate, soil, and land management inputs, the Rothamsted Carbon (RothC) model was applied to assess the current and future SOC stocks in the Anjeni watershed using long term climate, soil and land management data. RothC was calibrated with long-term SOC, land management, and climatic data from the Anjeni watershed in north-west Ethiopia. The correlation coefficient between simulated and observed SOC in 1997 and 2021 were 0.77 and 0.86, respectively, suggesting that the model could characterize the SOC of the Anjeni watershed. Then, the RothC was used to estimate SOC in the watershed for 30 years, from 2022 to 2052, under three slope gradients and four land use type and carbon storage scenarios (business as usual (BAU), low, medium and high carbon inputs). The result indicated that in the lower slope gradient, the current SOC simulation is less than all future scenarios considered under all land use types. Grass/fallow land showed higher current and projected SOC than cultivated land and plantation forest. Moreover, grass/fallow land with a gentle slope gradient had higher SOC than the watershed's middle and high-elevation parts. Overall, the model projected an increase of SOC under different future scenarios that could be due to climate and land use cover changes, the long-term soil-water conservation camping works and better soil and land managements in the watershed. This future assists for water retention, soil health, nutrient cycling, soil aeration, and greenhouse gas emission reduction, which in turn could enhance agricultural productivity, food security, and sustainable development.

Comparing laser-induced breakdown spectroscopy and visible near-infrared spectroscopy for predicting soil properties: A pan-European study

Soil organic carbon (SOC), texture, clay/organic carbon (OC) ratio, and extractable phosphorus are among the key soil descriptors representing the physical, chemical, and biological properties. However, analyzing these soil properties using conventional methods of analysis is time-consuming and often involves the use of hazardous chemicals. Therefore, scaling these methods to analyze more samples covering large geographical distributions within a reasonable turn-around time has become a challenge. In this study, two rapid techniques, namely laser-induced breakdown spectroscopy (LIBS) and visible and near-infrared spectroscopy (vis-NIRS), were compared for the prediction of SOC, texture, extractable phosphorus, and clay/OC ratio. A total of 896 samples collected from 23 European countries from the Land Use/Cover Area frame Survey (LUCAS) database were used to represent a wide distribution in terms of soil texture and geographical coverage. The data was manually split into calibration and validation samples while ensuring a representative geographical distribution in the calibration and validation set for the selected samples. Partial least square regression (PLSR) models were developed for both LIBS and vis-NIRS and compared based on prediction accuracy. LIBS outperformed vis-NIRS in predicting SOC, sand content, and clay/OC ratio, with root mean square error of prediction (RMSEP) values of 1.79 %, 12 %, and 7.69, respectively. Conversely, vis-NIRS yielded slightly higher RMSEP values for SOC, sand content, and clay/OC ratio predictions at 1.88 %, 15 %, and 8.13, respectively. Both LIBS and vis-NIRS performed similarly in predicting clay content, with an RMSEP of 6 %. However, there was a low predictive ability for silt and phosphorus concentration, as shown by the weak correlations (R2 < 0.5) between the predicted and reference soil properties. There was no significant difference between LIBS and vis-NIRS predictions for clay, silt, and SOC content (p > 0.05). However, there was a significant difference between the predicted sand content (p = 0.011) and clay/OC ratio (p < 0.001). The overall better performance of LIBS highlights its potential for application in a wide range of soils, especially for soils covering large geographical distributions and varying soil properties. Further, both techniques present opportunities for rapid determination of soil properties important for evaluating soil health on continental scales.

Using local ensemble models and Landsat bare soil composites for large-scale soil organic carbon maps in cropland

National soil organic carbon (SOC) maps are essential to improve greenhouse gas accounting and support climate-smart agriculture. Large-scale SOC models based on wall-to-wall soil information from remote sensing remain a challenge due to the high diversity of natural soil conditions and the difficulty of accounting for the spatial location of the soil samples. In this study, we tested if the implementation of local ensemble models (LEM) can be used to improve the SOC predictions from Landsat-based soil reflectance composites (SRC) for Germany. For this, we divided the research area into 30 times 30 km tiles and calculated local generalized linear models (GLM) based on random, nearby observations. Based on the GLMs, local SOC maps were predicted and aggregated using a moving window approach. The local variable importance was analyzed to identify spatial dependencies in the correlation between the SRC and SOC. For the final SOC map, a Random Forest (RF) model was trained using the aggregated local SOC predictions, the SRC, and a full set of training samples from the agricultural soil inventory. The results show that the LEM was able to improve the accuracy (R2 = 0.68; RMSE = 5.6 g kg−1), compared to the maps based on a single, global model (R2 = 0.52; RMSE = 6.8 g kg−1). The local variable importance of the spectral bands showed clear spatial patterns throughout the research area. Differences can be explained by the local soil conditions, influencing the correlation between SOC and the spectral properties. Compared to the widely adopted integration of distance covariates such as geographical coordinates, the LEM was able the reduce the spatial autocorrelation to a greater extent and to improve the prediction accuracy, especially for underrepresented SOC values. The LEM presents a new method to integrate spatial information and increase the interpretability of DSM models.

Monitoring soil organic carbon under coastal restoration using time series Sentinel‐1

AbstractSalt marsh soils are key reservoirs of organic carbon. Although the potential for using Sentinel‐1 data to estimate organic carbon in marsh soils has been recognized, there is still a need to find suitable proxies and effective algorithms that enhance predictions. In this study, we assessed the effectiveness of Sentinel‐1 data for predicting soil organic carbon (SOC) stocks in salt marshes located on the coast of eastern‐central China. We defined the relations between SOC stocks and remotely sensed data using a knowledge‐based approach (i.e., structural equation modeling (SEM)), and the results were compared with those obtained using a common data‐driven approach (i.e., random forest (RF)). Predictive models were developed using (1) refined images associated with phenological stages (phenological variables), (2) refined images that were classified as statistically important using Monte Carlo feature selection (MCFS‐based variables), and (3) images from the combination of these refined datasets. The predictions were validated using a leave‐one‐out cross‐validation approach. The results showed that SEM was more accurate than RF when predicting SOC stocks using the same type of predictor variables. Models using phenological variables alone yielded the least accurate predictions. Adding phenological variables to the structural equation model with MCFS‐based variables increased the prediction accuracy by 36% in terms of the R2. The results of the case study suggest that images related to phenological stages are good predictors for mapping SOC stocks in marshes. This study highlights the superiority of SEM over RF for developing effective remote sensing‐based models to quantify and map SOC stocks in salt marshes.

Graph convolutional networks for spatial interpolation of correlated data

Several deep learning methods for spatial data have been developed that report good performance in a big data setting. These methods typically require the choice of an appropriate kernel and some tuning of hyperparameters, which are contributing reasons for poor performance on smaller data sets.In this paper, we propose a mathematical construction of a graph-based neural network for spatial prediction that substantially generalizes the KCN model in [Appleby, Liu and Liu (2020). Kriging convolutional networks. In Proc. AAAI Conf. AI 34, pp. 3187–3194]. In particular, our model, referred to as SPONGE, allows for integrated learning of the convolutional kernel, admits higher order neighborhood structures and can make use of the distance between locations in the neighborhood and between labels of neighboring nodes. All of this yields higher flexibility in capturing spatial correlations.We investigate in simulation studies including small, medium and (reasonably) large data sets in what situations and to what extent SPONGE comes close to or (if the conditions for optimality are violated) even beats universal Kriging, whose predictions incur a high computational cost if n is large. Furthermore we study the improvement for general SPONGE in comparison with the usual KCN.Finally, we compare various graph-based neural network models on larger real world data sets and apply our method to the prediction of soil organic carbon in the southern part of Malawi.

Predicting soil organic carbon content using simulated insitu spectra and moisture correction algorithms in southern Xinjiang, China

Soil organic carbon (SOC) is an essential component of the terrestrial carbon pool, and changes in SOC are critical to ecosystem stability and agricultural production. In-situ field spectroscopy, such as visible and near-infrared (vis-NIR) spectroscopy, has proven to be an ideal tool to achieve rapid and efficient detection of SOC. However, soil moisture interference is one of the main challenges for the measurement of SOC using in-situ spectroscopy. Great progress has been made in using external parameter orthogonalization (EPO), direct standardization (DS) and piecewise direct standardization (PDS) to remove the influence of moisture on using vis-NIR spectroscopy to predict SOC. However, most current studies were conducted based on a certain moisture, which makes it difficult to provide a comprehensive guide for SOC detection using in situ spectroscopy. In this study, 135 surface (0–20 cm) soil samples were collected from the Aksu region of Xinjiang, northwestern China. The soil samples were wetted to a saturated moisture content state after manual removal of salts. The spectra data of 11 different moisture levels during natural air-drying of the wet soil samples were measured to analyze and compare the performance of three algorithms, EPO, DS and PDS, in removing the effect of moisture and improving the accuracy of SOC prediction. The results show that PDS has a relatively poor ability to remove the effect of moisture among the three moisture correction algorithms. When the moisture content of the soil exceeds 48%, EPO, DS and PDS cannot effectively remove the disturbance of moisture. The EPO algorithm was the most effective in removing moisture effects when the soil moisture content (SMC) was 25–48%. Using EPO, the predicted R2 and RPIQ of SOC increased by >0.19 and 0.87, respectively. The DS algorithm was the best method to remove the moisture effect on soil vis-NIR spectra when the SMC was 6–25%, and the R2 and RPIQ predictions of SOC after removing the effect of moisture by DS increased by >0.09 and 0.87, respectively. However, when the moisture content is <6%, the influence of soil moisture on the prediction accuracy of SOC spectra is negligible, and SOC can be predicted with high accuracy using in-situ spectra without the introduction of moisture correction algorithms. An appropriate moisture correction algorithm should be selected to remove moisture interference according to the SMC condition.

Learning vs. understanding: When does artificial intelligence outperform process-based modeling in soil organic carbon prediction?

In recent years, machine learning (ML) algorithms have gained substantial recognition for ecological modeling across various temporal and spatial scales. However, little evaluation has been conducted for the prediction of soil organic carbon (SOC) on small data sets commonly inherent to long-term soil ecological research. In this context, the performance of ML algorithms for SOC prediction has never been tested against traditional process-based modeling approaches. Here, we compare ML algorithms, calibrated and uncalibrated process-based models as well as multiple ensembles on their performance in predicting SOC using data from five long-term experimental sites (comprising 256 independent data points) in Austria. Using all available data, the ML-based approaches using Random forest and Support vector machines with a polynomial kernel were superior to all process-based models. However, the ML algorithms performed similar or worse when the number of training samples was reduced or when a leave-one-site-out cross validation was applied. This emphasizes that the performance of ML algorithms is strongly dependent on the data-size related quality of learning information following the well-known curse of dimensionality phenomenon, while the accuracy of process-based models significantly relies on proper calibration and combination of different modeling approaches. Our study thus suggests a superiority of ML-based SOC prediction at scales where larger datasets are available, while process-based models are superior tools when targeting the exploration of underlying biophysical and biochemical mechanisms of SOC dynamics in soils. Therefore, we recommend applying ensembles of ML algorithms with process-based models to combine advantages inherent to both approaches.

Predicting soil organic carbon with different approaches and spatial resolutions for the southern Iberian Peninsula, Spain

Quantification and monitoring of soil organic carbon (SOC) stocks across local-to-global scales are needed to assess soil resource management practices and adapt environmental policies. Multiple SOC estimates are available worldwide; however, verification and validation are required to quantify the discrepancies and provide improved estimates. Here, we evaluated four different digital soil mapping (DSM) approaches (i.e., linear models, support vector machine, random forest, and quantile regression forest) to estimate SOC concentration (SOCc) and SOC stocks (SOCs) in the Region of Murcia (11,313 km2), a complex topographic and climatic area in the southern Iberian Peninsula, at three spatial resolutions (100, 250, and 1000 m). We estimated SOC spatially using a local harmonized database of 255 soil profiles for modeling and 1100 topsoils for external validation. We found that the quantile regression forest (QRF) approach had the best data-model agreement at 100 m spatial resolution, with the highest accuracy percentage (79%), external validation (correlation coefficient of 52%), and spatial interpretability of patterns, especially for SOCc. The QRF model showed a mean SOCc of 12.18 g/kg with an overall uncertainty of 10.54 g/kg and an accuracy percentage of 79%, whereas the total SOCs was 27,572 GgC with an uncertainty of 0.016 GgC. Our results showed that using local environmental covariates and local soil information to predict SOC within this region resulted in a relative improvement in the prediction accuracy of ∼40% for SOCc and ∼ 65% for SOCs compared to the SOC products derived from national and global databases. Our results showed a large discrepancy between the national and global estimates for reporting SOC locally. Consequently, local-to-regional efforts are needed to describe SOC spatial variability better to reduce uncertainty and improve the assessment of soil resources. We provide the resulting SOC maps with associated spatial uncertainty on the public Environmental Data Initiative Repository at https://portal.edirepository.org/nis/mapbrowse?packageid=edi.1238.2.

A random forest model to predict soil organic carbon storage in mangroves from Southern Colombian Pacific coast

Mangrove ecosystems serve as significant reservoirs of carbon stocks, particularly within the soil. However, these ecosystems face widespread degradation globally, underscoring the necessity for comprehensive technical insights to facilitate effective management. Enhancing the precision of soil organic carbon estimates within mangroves necessitates access to precise field data. Hence, this study aimed to assess the soil organic carbon stored in mangrove ecosystems along the southern Colombian Pacific coast. Employing a Random Forest model, constructed with data from 28 soil profiles spanning a depth of 2 m and 16 predictor variables, we evaluated the model's performance using repeated 10-fold cross-validation (with 5 random repeats). Uncertainty in carbon storage predictions was quantified through 95% confidence intervals. Our model successfully accounted for 88% of the variability in soil organic carbon storage. Depth, cation exchange capacity, pH, and sand content emerged as the most influential predictors of soil organic carbon storage variation. The study estimated a mean of 270.32 ± 52.75 t C ha−1, with a total carbon reserve of 25458088 t in the study area. Notably, this value falls below global estimates, suggesting a potential overestimation of carbon stocks in mangrove soils by existing global models.