Search

• 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. [ABSTRACT FROM AUTHOR]

• Filter, wrapper, embedded and ensemble feature selection approaches (FS) were evaluated. • The proposed ensemble FS incorporates ten individual selectors to identify top significant variables. • Ensemble FS produced the highest predictive variable subset for mapping SOM. • XGBoost outperformed RF in mapping SOM in subtropical restored forests. Mapping Soil organic matter (SOM) over a complex forest landscape is challenging due to the difficulty in selecting the most insightful variables from high-dimensional datasets in the recent explosion of geospatial-data. Feature selection (FS) is necessary to reduce data redundancy and noise as well as to achieve more reliable SOM spatial predictions. However, it is still unclear that which is most effective among various FS methods in mapping SOM. Therefore, four types of FS approaches (i.e., filter, wrapper, embedded and ensemble) were adopted to generate optimum variable subsets from an original variable dataset of 60 candidates, respectively, for mapping SOM of restored forest land in a typical subtropical region of southern China. The most used methods for each type of FS approaches were selected in this study, including three filters (Chi-square, InfoGain and pearson correlation analysis), three wrappers (genetic algorithm, simulated annealing algorithm and support vector machine-recursive feature elimination) and three embedded methods (Boruta, random forest (RF) and extreme gradient boosting (XGBoost)), as well as an ensemble method (robust rank aggreg algorithm (RRA)). Meanwhile, the RF and XGBoost models were applied with a 10-fold cross-validation method to compare the relative advantages of the different FS methods in SOM mapping, by utilizing the correlation coefficients R2 between observed and predicted values and predicting errors of root mean square error (RMSE). The results show that the SOM prediction accuracies with optimized variable subsets generated by the different FS methods are better than those with full variables, yet the improvements of prediction performance are different among the four types of FS approaches. The ensemble method (RRA) is superior to the other three types of approaches with an average RMSE reduction of 9.16% comparing to that without using FS methods, followed by wrapper and embedded methods which obtained the average RMSE reduction by 7.81%, 7.32%, respectively, and the filter methods are the weakest in the RMSE reduction with slight decreases of 4.32%. The XGBoost model achieved a better performance in predicting SOM than the RF model regardless of input variables, and the XGBoost model combined with RRA FS method shows the greatest potential to map SOM in the restored forest land. This study provides a reference for obtaining more parsimonious and robust variable sets from the available big geo-data freely for soil mapping in other areas. [ABSTRACT FROM AUTHOR]

• Simulation accuracy and cost are two core issues in soil organic carbon mapping. • A zonal simulation approach was proposed for realizing high accuracy and low cost. • A typical subzone with an area proportion <20% of the study area was identified. • The zonal simulation achieves high accuracy with a cost decrease by more than 50%. • Geographical integrity, typicalness, and scale are key points of zonal simulation. Digital soil mapping (DSM) has been developed for decades and aims to accurately simulate the soil factors with low cost. However, it still cost considerably in coastal wetlands because of the high difficulty in field survey. A zonal simulation approach was proposed for decreasing the cost of field survey in the premise of high accuracy for soil organic carbon stock mapping in coastal wetlands, and Chongming Island, an important estuarine wetland in China, was used to demonstrate the study. A subzone with an area proportion lower than 20% was identified based on the typicalness of land surface characteristics and soil influencing factors for the whole study area. Entire and zonal simulations were conducted using abundant predictors sourced from remote sensing, partial least square regression, and 10-fold cross validation. The results revealed that root mean squared errors of entire and zonal simulations were 1.70 g/kg and 1.95 g/kg, respectively, which were in a low level compared with those in previous studies and indicated the high simulation accuracy. The zonal simulation exhibited a slightly lower accuracy than the entire simulation, while considerably decreased the cost by more than 50%. The decrease in the cost showed a much more practical significance than the increase in the accuracy in areas with high difficulty in field survey, and the geographical integrity and continuity of the whole study area, the typicalness of the selected subzone, and the spatial extent were the three key points when promoting the zonal simulation in DSM. The soil organic carbon stock in the study area was generally high in areas with good vegetation condition, low soil salinity, complex landscape configuration, and long distances from the sea. Human activities and the resulting landscape fragmentation have been the dominant factor that drives the spatial pattern of soil organic carbon stock. [ABSTRACT FROM AUTHOR]

Digital soil mapping approaches related to soil organic matter (SOM) are crucial to quantify the process of the carbon cycle in terrestrial ecosystems and thus, can better manage soil fertility. Recently, many studies have compared machine learning (ML) models with traditional statistical models in digital soil mapping. However, few studies focused on the application of hybrid models that combine ML with statistical models to map SOM content, especially in loess areas, which have a complicated geomorphologic landscape. In this study, the trend prediction used two ML models, i.e., gradient boosting modeling and random forest (RF), and a traditional stepwise multiple linear regression plus interpolated residuals generated from two classic geostatistical models, i.e., ordinary kriging and inverse distance weighting, and a high accuracy surface modeling (HASM) were implemented to map SOM content in the Dongzhi Loess Tableland area of China. A total of 145 topsoil samples and heterogeneous environmental variables were collected to develop the hybrid models. Results showed that 18 variables related to soil properties, climate variables, terrain attributes, vegetation indices, and location attributes played an important role in SOM mapping. The models that incorporate ML algorithms and interpolated residuals to predict SOM variation were found to have a better ability to handle complex environment relationships. The HASM model outperformed traditional geostatistical models in interpolating the residuals. In contrast, RF combined with HASM residuals (RF_HASM) gave the best performance, with the lowest mean absolute error (1.69 g/kg), root mean square error (2.30 g/kg), and the highest coefficient of determination (0.57) and concordance correlation coefficient (0.69) values. Moreover, the spatial distribution pattern obtained with RF_HASM yielded a spatial distribution of SOM that better fit the actual distribution pattern of the study area. In conclusion, these results suggest that RF_HASM is particularly capable of improving the mapping accuracy of SOM content at the regional scale. [ABSTRACT FROM AUTHOR]

• The feasibility of adding multi-temporal radar Sentinel-1A data to SOC prediction was explored. • Climate variables contributed the most to explaining SOC content variation. • In this ecosystem, RF and BRT models showed similar performance and produced similar SOC spatial distribution patterns. Soil organic carbon (SOC) has a large impact on soil quality and global climate change. It is therefore important to be able to predict SOC accurately to promote sustainable soil management. Although the synthetic aperture radar (SAR) has many advantages and has been widely used in soil science research, it has rarely been used in previous SOC mapping studies based on remote sensing images. The purpose of this study was to investigate the ability of multi-temporal Sentinel-1A data in SOC prediction, by comparing the predictive performance of random forest (RF) and boosted regression tree (BRT) models in the Heihe River Basin in northwestern China. A set of 162 topsoil (0–20 cm) samples were taken and 15 environmental variables were obtained including land use, topography, climate, and remote sensing images (optical and SAR data). Using a cross-validation procedure to evaluate the performance of the models, three statistical indices were calculated. Overall, both RF and BRT models effectively predicted SOC content, exhibiting similar performance and producing similar spatial distribution patterns of SOC. The results showed that the addition of multi-temporal Sentinel-1A images improved prediction accuracy, with the root mean squared error (RMSE), the mean absolute error (MAE) and the coefficient of determination (R2) improving by 9.0%, 8.3% and 13.5%, respectively. Furthermore, the combination of all environmental variables had the best prediction performance explaining 75% of SOC variation. The most important environmental variables explaining SOC variation were precipitation, elevation, and temperature. The multi-temporal Sentinel-1A data in RF and BRT models explained 9% and 7%, respectively. The results from our case study highlight the usefulness of multi-temporal Sentinel-1 data in SOC mapping. [ABSTRACT FROM AUTHOR]