National-scale 3D mapping of soil organic carbon in a Japanese forest considering microtopography and tephra deposition

Geographic information on soil organic carbon (SOC) is important for climate change research and understanding the global carbon cycle. At present, this information is limited for forest soils in mountainous regions that have complex topographies and are affected by tephra deposition. This study predicted the SOC concentration (OC), bulk density (BD), apparent rock coarse fraction (CF), and SOC stock (OCS) at 0-5, 5-15, and 15-30 cm in the soil horizon for the forested areas (ca. 230 000 km(2)) of Japan. This was carried out with a 10 m grid resolution using a digital soil mapping (DSM) approach. To determine the appropriate spatial prediction model, we evaluated the accuracy of a two dimensional approach for ordinary kriging (OK) and regression kriging (RK), and a three dimensional approach for random forest model (RF). The RF was based on topography, climate and vegetation factors (RFenv), distance and direction from 152 volcanos (RFenv+vol), distance from survey points (RFenv+sp); these factors were also combined (RFall) as explanatory attributes. The results demonstrated an improvement in the average root mean square errors (RMSEs) of RFenv+vol, RFenv+sp, and RFall by approximately 12%, 14%, 5%, and 21% for OC, BD, CF, and OCS using 10-fold cross-validation on a site-bysite basis, respectively, compared with OK. Based on the relatively small difference in improvement among RFenv+vol, RFenv+sp, and RFall (<0.5%), and the considerably high computational cost of RFenv+sp and RFall, RFenv+vol was considered as the appropriate model for the area. The R-2 of RFenv+vol was calculated at 0.59, 0.44, 0.30, and 0.38 in OC, BD, CF, and OCS, respectively, using 10-fold cross validation on a site-by-site basis. The predicted map by RFenv+vol accurately reproduced the large spatial variation in OCS at small watershed, regional, and national scales; this have been derived from the effect of microtopography, tephra deposition, and the gradient of temperature with latitude, respectively. A larger OCS was predicted to accumulate on the ridges, highland slopes, near volcanos, and higher latitudes than at other areas. The mean OCS was estimated to be 7.6 kg.m(-2) for a soil depth of 0-30 cm in Japanese forests. The results also suggest that the fine-scale three dimensional random forest approach using the distances from volcanoes showed promise for regional-scale OCS prediction. This approach was considered particularly effective in hills and mountainous areas that are strongly exposed to tephra deposition, this includes the Pacific Rim regions and other volcanic countries.

Automated summary

Geographic information on soil organic carbon (SOC) is crucial for climate change research and understanding the global carbon cycle. This study focused on predicting SOC concentration in forested areas of Japan, using digital soil mapping (DSM) with a 10 m grid resolution. The random forest model (RF) was found to be the most appropriate for the area in comparison to ordinary kriging (OK) and regression kriging (RK), with RFenv+vol considered the best model due to its higher accuracy in predicting SOC stock (OCS).