Historical surveys and current soil maps can identify soil differences and establish carbon monitoring baselines. We compared 330 samples from 1980 to 1983 with 376 independent samples from 2025 in Pingle County, China. Random forest models used 11 common covariates and nested spatial cross-validation. Land-cover maps supported comparisons of predicted soil organic matter (SOM). Cropland and forest sample means were 4.70 and 4.97 g/kg lower in 2025. Within the common valid area, predicted means were 26.43 and 23.54 g/kg. The mean predicted difference was −2.89 g/kg, with a conditional block-resampling range of −5.61 to −0.13 g/kg. Mean differences were −2.31, −3.47, and −4.81 g/kg for forest–forest, cropland–cropland, and forest–cropland combinations. With water bodies and built-up land excluded, a 20 cm depth and 0.58 conversion factor gave a 2025 soil organic carbon stock of 6.180 Tg C. Mean density was 3.35 kg/m², and the joint-model resampling range was 5.805–6.497 Tg C. Forest and cropland contributed 74.01% and 25.97%. Unpaired sampling and unverified analytical comparability prevent direct inference of temporal change or conversion effects. Conditional ranges exclude potential analytical bias between periods. These estimates provide a spatial basis for identifying priorities for soil carbon protection and soil quality resampling at the county scale. They can support sustainable land management and future soil carbon monitoring in karst landscapes.