Spatial Distribution Characterization And Digital Mapping Of Soil Organic Carbon At Topsoil In Mount Sejila

Soil organic carbon is an essential component of soil fertility. Soil organic carbon also plays an important role in soil and water conservation as well as landscape resilience. Moreover, soil organic carbon is known to increase the soil’s nutrient cycling capability. What’s more, as the largest C pool in the terrestrial ecosystem, soil organic carbon pool plays a key role in the organic carbon balance between different ecosystems. Low temperature and high altitude are characteristics of the Qinghai-Tibet Plateau, which make it extremely sensitive to soils and plants under climate change, so the Qinghai-Tibet Plateau is vital to Global C storage and balance. Because of the difficulty of collecting soil samples in the Qinghai-Tibet Plateau, present researches on estimation of soil organic carbon storage in this region are mainly dependent on the method of combining soil classification and GIS. However, this method bring huge uncertainty in soil organic carbon estimation due to high spatial variation in the Qinghai-Tibet Plateau.The study area was located in the Mount Sejila, southeastern of Tibet. The Scorpan function, the core of digital soil mapping, was used to predict soil organic carbon at topsoil (0-20 cm). Land cover was obtained firstly, then environmental covariates such as soil, climate and relief were calculated by scale transforming algorithms at 90 m resolution. After that, different data mining algorithms were compared to get the best prediction model between soil organic carbon at topsoil from field survey and environmental covariates. After estimating soil organic carbon at topsoil in the study area, the spatial distribution characteristics of soil organic carbon were analyzed. Finally, soil organic carbon density and soil organic carbon storage were estimated. The main research contents and results were as follows:(I) The distribution of land cover using high resolution satellite imagesThe GF-1 satellite images in the study area were pre-processed, followed by accuracy comparison of land cover classification between minimum distance classification and decision trees classification which were based on the data of band 1, band 2, band 3, band 4, NDVI, result of ISODATA and DEM. The results showed that decision trees classification performed better than minimum distance classification with total accuracy in 79.76% and Kappa coefficient in 0.71. The best GF-1 satellite images were obtained in September with lots of study area were covered by snow, thus Landsat 8 satellite images in July were used to correct snow covered regions. Among all kinds of land covers, forest occupied the largest proportion of total area with 57.75%, meadow came the second with 25.31%, the proportion of arable was the least in 6.50%, and 14.43% of total area were covered with snow all the year round.(2) The distribution of relief, mean annual precipitation and mean annual temperature using middle and low resolution satellite imagesRelief relative environmental covariates including elevation, slope, aspect, length of slope, curvature, valley depth, terrain roughness index, terrain wetness index and multi-resolution valley bottom flatness were calculated through DEM at 90 m resolution. Mean annual precipitation from TRMM at 0.25 degree resolution and mean annual temperature from MODIS at 0.05 degree resolution were down-scaled to 90 m resolution by geographically weighted regression. The results demonstrated that geographically weighted regression, based on longitude, latitude and elevation, performed well in down-scaling mean annual precipitation and mean annual temperature in the Linzhi, and R2 in the calibration were 0.91 and 0.99 respectively. The local R2 of mean annual precipitation and mean annual temperature in the Mount Sejila were 0.63-0.72 and 0.84-0.98 respectively.(3) The distribution of principle components about visible near-infrared spectra at topsoil through field surveySoil organic carbon of topsoils sampled in the field survey were analyzed in the laboratory. First three principle components, obtained from principle component analysis of soil visible near-infrared spectra, could explain 98.63% of the total variance, thus they could reflect soil information. The first two principle components of soil spectra could distinguish arable and natural land covers (forest and meadow). However, it’s difficult to separate from forest and meadow. Taking elevation as a covariate, the distribution of first three principle components about spectra at topsoil were mapped using ordinary cokriging finally.(4) The best predictive model of soil organic carbon at topsoil and the estimation of soil organic carbon storage using Scorpan functionCubist algorithm and random forest algorithm were compared to modelling soil organic carbon at topsoil. With a lowest RMSE in 7.62 g kg-1, the predictive model using random forest gained higher accuracy than Cubist algorithm. Random forest model revealed variable importance of each environmental covariates in predictive model of soil organic carbon at topsoil. Explaining nearly 70% of total variance, the first three principle components were the most important factors in the model. The mean soil organic carbon at topsoil in the Mount Sejila was 49.51 g kg-1 while the mean soil organic carbon density at topsoil was 11.43 kg m-2 which is far more than that in the Tibet in 4.27 kg m-2. The mean soil organic carbon and mean soil organic carbon density of forest were the highest among three land covers, reaching 51.71 g kg-1 and 12.10 kg m-2 respectively. Meadow had the second mean soil organic carbon and mean soil organic carbon density, pointing at 51.15 g kg-1 and 10.45 kg m-2 respectively. With mean soil organic carbon at 36.18 g kg-1 and mean soil organic carbon density at 9.70 kg m-2, arable ranked last. The soil organic carbon storage estimated by Scorpan function in the Mount Sejila was 2.79 × 109 g which is far more than that estimated by the method of combining soil classification and GIS. The soil organic carbon storage in the Southeastern Tibet was highly underestimated because of high spatial variation.Digital soil mapping has been proved an effective method for predicting soil organic carbon at topsoil in the Mount Sejila. Moreover, the method can generate a high resolution map of soil organic carbon so as to improve prediction accuracy and decrease the prediction error in the modelling. Our research can provide a basis for improving the prediction accuracy of soil organic carbon in the Qinghai-Tibet Plateau.