Estimating soil organic carbon stocks using machine learning methods in the semi-arid rangelands of New South Wales

Soil organic carbon (SOC) is pivotal for biological, chemical and physical processes and provides vital information on changes in soil fertility and land degradation. Rangelands, accounting for about 81% of Australian land area, represent considerable carbon storage potential. Efficient modelling techniques to evaluate the potential for rangeland SOC stocks are vitally important in the assessment for the global carbon cycle and quantum abatement. This study aimed to evaluate boosted regression trees (BRT) and random forest (RF) in predicting SOC stocks from ground measured and remotely-sensed variables using two feature selection techniques to identify the dominant variables that affect SOC stocks in the rangelands. Using field-based measurement of SOC stock collected from 564 sites across the study area and 28 of GIS-based environmental variables including climate, topography, radiometry, vegetation and land fractional cover data, we employed stepwise regression (SR, linear approach) and genetic algorithm (GA, nonlinear approach) to select the most informative variables. These selected predictors were then used to train the BRT and RF models. In all, four models were evaluated; BRT using stepwise selection of predictors (SR_BRT); RF using stepwise (SR_RF); BRT using GA selection of predictors (GA_BRT) and RF using GA (GA_RF). In addition, BRT using all predictors (All_BRT) and the RF using all predictors (All_RF) were used as benchmarks to test the performance of the four models. Of the field-based data, 75% was used to train the model (“calibration dataset”) and the remaining 25% was used to validate the prediction of SOC stocks (“validation dataset”). The results indicate that the RF exhibited a better performance in predicting SOC stocks than the BRT regardless of input variables. The two models explained ~45% of the total SOC stocks. In addition, we verified that feature selection for both machine learning techniques is necessary for estimating SOC stocks, even though BRT was relatively insensitive to the input features selected by SR. The GA_RF was the most promising model with reliable predictors to predict SOC stocks, with the lowest root mean square error (RMSE) and the highest R² values (7.44 Mg C ha^-1 and 0.48, respectively), suggesting that the proposed methodology may provide a cost effective method to predict SOC stocks in the rangelands. The important variables for explaining the observed SOC stocks were rainfall, elevation, prescott index (PI), and land fractional cover (bare ground fraction).