Testing and comparing machine learning algorithms for predicting soil organic carbon stocks in Brazilian mangrove soils

Mangrove soils represent the second largest global reservoir of soil organic carbon (SOC) stocks, making accurate quantification of these stocks essential for supporting conservation strategies and providing inputs to global bioclimatic models. Although machine learning (ML) algorithms are widely applied to predict SOC stock in other ecosystems, comparative evaluations of their performance in mangrove soils remain limited. This study evaluated the performance of four ML algorithms in predicting SOC stocks in the 0-30 cm layer of mangrove soils along the Brazilian coast, using the largest available legacy soil database, comprising 245 samples collected from 65 sampling sites. Predictor variables included total organic carbon, bulk density, sampling depth, and oceanographic variables (seawater pH, temperature, and salinity). The algorithms tested were Random Forest (RF), Support Vector Machine (SVM), Kernel-weighted k-Nearest Neighbors (kknn), and Linear Model (LM). Model performance was assessed using the coefficient of determination (R²), Lin’s concordance correlation coefficient (CCC), mean absolute error (MAE), and root mean square error (RMSE). RF and SVM demonstrated superior predictive performance, both achieving R² = 0.59 and CCC = 0.74. The RF model yielded MAE = 4.20 and RMSE = 6.40, while SVM achieved MAE = 3.84 and RMSE = 6.00. In contrast, kknn and LM showed weaker performance, with lower R² (0.46 and 0.07, respectively) and CCC (0.68 and 0.25, respectively), and higher prediction errors (MAE = 4.59 and 8.30; RMSE = 7.69 and 15.39, respectively). These findings support the application of RF and SVM for robust SOC prediction in mangrove ecosystems.