Soil organic carbon (SOC) and total nitrogen (TN) play major roles in soil quality and the global carbon budget. They can be measured rapidly and cost-effectively via visible and near-infrared reflectance (VNIR) spectroscopy. However, the reliability of this method is questionable because of the effects of heterogeneity. At present, only a few publications have addressed the effect of soil layers on model applicability, especially for highly heterogeneous soils in forest ecosystems. In the current work, we evaluated the performance of VNIR spectroscopy in estimating SOC and TN contents of soils collected from a mixed mountain forest in Central China. We also investigated the applicability of spectroscopic models between soil layers. We then further explored the possibility of using spiking with extra-weighting to improve model applicability. To achieve such objectives, we evaluated the applicability accuracy of the initial models (global and layered models) and modified models (spiked models with and without extra-weighting). Results showed that all the initial models successfully predicted SOC and TN. That is, for SOC, R2 P ranged from 0.79 to 0.90, ratio of performance to inter-quartile range (RPIQ) ranged from 3.07 to 3.97, and the root mean square error (RMSEP) ranged from 0.54% to 0.88%; for TN, R2 P ranged from 0.66 to 0.86, RPIQ ranged from 2.12 to 3.78, and RMSEP ranged from 0.05% to 0.08%. However, the prediction accuracies were seriously reduced when the model constructed from the top soil layer was used to predict the sub-surface soil properties, and vice versa. In terms of model applicability, our results demonstrated that spiking improved the applicability of the initial calibrations (RMSEP and absolute prediction bias were obviously reduced) and that the accuracy was further improved when the spiking subset was extra-weighted. When the extra-weighting reached a certain level, the accuracies remained stable or slightly reduced. Our results illustrated that spiking alone and spiking with extra-weighing are effective approaches to improve model applicability in the VNIR estimation of SOC and TN between different soil layers in a highly heterogeneous forest. This approach is potentially useful in rapidly quantifying and monitoring soil carbon and nitrogen pools in heterogeneous landscapes.