Geochemical predictability quantification and spatial transferability of soil organic carbon in Ganges Delta

Abstract. Deltas are among the least studied but most dynamically active reservoirs that store a significant amount of the world’s soil organic carbon (SOC). This study evaluates the complex dynamics of deltas by combining soil profiles from four geomorphological zones in the Ganges Delta of Bangladesh with 6,652 soil profiles from fourteen major deltas worldwide. These combined profiles test the limits of regional SOC models, showing where local frameworks hold and where they break. Statistical analysis shows that salinity, cation exchange capacity, and moisture correlate most strongly with SOC (r = 0.66, 0.64, and 0.56, respectively), with most saline profiles storing approximately 6.5 times more carbon than freshwater sites. Salinity also exhibits a strict scale-dependent effect of strong correlations across geomorphological zones but effectively vanishing within them. This spatial paradox introduces major implications for digital soil mapping protocols in coastal deltas. Projections from a process model show that mangrove restoration efforts on half of the most saline tracts would increase SOC levels by 26.6 % in 2050, far exceeding baseline model of business-as-usual projections that show a 5.9–11.6 % decline across climate pathways. Model simulations indicate that local management has 4.6 to 7.2 times more influence on SOC storage than the gap between RCP 4.5 and RCP 8.5. To prevent data overlap from multi-depth sampling, the cross-validation had to hold back entire locations to avoid spatial leakage. As a result gradient boosting reached an R² of 0.64, against an inflated 0.91, and deeper neural networks provided no reliable improvement at this sample size. Analyzing every combination of the fifteen deltas exposed a sharp limit: models generalize within a single delta (median R² = +0.25) but fail between deltas, yielding positive scores in only 6 % of 210 pairs (median R² = −0.55). Transfer failed in both directions, even when fourteen deltas were used to predict the fifteenth delta. Unlike statistical weights, the structural framework of the process model transferred well, showing that the form of the model is broadly applicable when parameter tuning remains local site specific. Consequently, gathering data from foreign deltas cannot substitute for local profile measurements. These findings demonstrate that global datasets can guide structural frameworks but effective coastal carbon management ultimately depends on local site-specific observation.