Improving the Spatial Prediction of Soil Organic Carbon Content in Two Contrasting Climatic Regions by Stacking Machine Learning Models and Rescanning Covariate Space

Abstract

Understanding the spatial distribution of soil organic carbon (SOC) content over different climatic regions will enhance our knowledge of carbon gains and losses due to climatic change. However, little is known about the SOC content in the contrasting arid and sub-humid regions of Iran, whose complex SOC–landscape relationships pose a challenge to spatial analysis. Machine learning (ML) models with a digital soil mapping framework can solve such complex relationships. Current research focusses on ensemble ML models to increase the accuracy of prediction. The usual ensemble method is boosting or weighted averaging. This study proposes a novel ensemble technique: the stacking of multiple ML models through a meta-learning model. In addition, we tested the ensemble through rescanning the covariate space to maximize the prediction accuracy. We first applied six state-of-the-art ML models (i.e., Cubist, random forests (RF), extreme gradient boosting (XGBoost), classical artificial neural network models (ANN), neural network ensemble based on model averaging (AvNNet), and deep learning neural networks (DNN)) to predict and map the spatial distribution of SOC content at six soil depth intervals for both regions. In addition, the stacking of multiple ML models through a meta-learning model with/without rescanning the covariate space were tested and applied to maximize the prediction accuracy. Out of six ML models, the DNN resulted in the best modeling accuracies, followed by RF, XGBoost, AvNNet, ANN, and Cubist. Importantly, the stacking of models indicated a significant improvement in the prediction of SOC content, especially when combined with rescanning the covariate space. For instance, the RMSE values for SOC content prediction of the upper 0–5 cm of the soil profiles of the arid site and the sub-humid site by the proposed stacking approaches were 17% and 9% respectively, less than that obtained by the DNN models—the best individual model. This indicates that rescanning the original covariate space by a meta-learning model can extract more information and improve the SOC content prediction accuracy. Overall, our results suggest that the stacking of diverse sets of models could be used to more accurately estimate the spatial distribution of SOC content in different climatic regions.

Prediction of Soil Organic Carbon under Different Land Use Types Using Sentinel-1/-2 Data in a Small Watershed

“…The influence of land use type on SOC prediction and the importance of optical and radar data under different land use types were analyzed. The overall prediction result of the model is better than some existing researches based on satellite remote sensing data [16,21,[23][24][25]33], and is similar to the result of Taghizadeh-Mehrjardi et al [17]. Among the three land use types, the prediction results of orchard (R 2 = 0.86 and MSE = 0.004%) are better than dry land and paddy field.…”

“…Six common performance indicators including: R 2 , mean absolute error (MAE), mean squared error (MSE), percent root mean square error (%RMSE), ratio of performance to interquartile range (RPIQ), and corrected akaike information criterion (AICc) were used to assess model performance under different land use types. R 2 varies between 0 and 1, which indicates the closeness between the observed value and the fitting regression line or the variance ratio explained by independent predictors [17]. The model has great fitting degree and stability when R 2 closes to 1.…”

“…In this study, the aspect made an important contribution to paddy field, and the other terrain variables data had a lower contribution than Sentinel-1 and Sentinel-2 data sets in orchard and dry land. Taghizadeh-Mehrjardi et al [17] also found that Sentinel-2 bands have a greater correlation with SOC content than elevation while Zhou et al [33] concluded that terrain variables are more important than Sentinel data in a study area with plains and plateaus. The difference may be because the range of altitude changes in this study was relatively small.…”

“…Therefore, it is necessary to explore the contribution of remote sensing data to SOC prediction. There have been some studies on SOC prediction based on remote sensing data that achieved good prediction results, especially optical data (e.g., Sentinel-2 [15][16][17][18][19][20][21], Landsat [22][23][24][25][26][27][28], and MODIS satellite data [29][30][31]); their bands cover from visible to short-wave infrared, providing more information. However, the application of optical data is susceptible to weather conditions, especially in the Sichuan Basin where clouds occur most frequently [32], so the available optical data are very limited.…”

“…Specifically, SAR satellite imagery (Sentinel-1), optical satellite imagery (Sentinel-2), and digital elevation model (DEM) data were used as input variables. The extreme gradient boosting (XGboost) algorithm has been used to predict SOC with good prediction results [17,22,29,30]. Therefore, we used XGboost to build prediction models of SOC content under different land use types and analyze the importance of feature variables.…”