Spatial prediction and dynamic change of soil organic carbon using remote sensing variables as auxiliary information in wavy plain, Northeast China

Abstract

Remote sensing (RS) data, which provide rich spatial information, can effectively reflect soil physical and chemical properties. These data can complement environmental variables for soil organic carbon (SOC) prediction, reducing the influence of covariance between topographic variables in wavy landscapes areas and improving prediction accuracy. This study focuses on improving prediction accuracy and applying SOC content prediction models in wavy plains. We constructed three RS variables, normalized difference vegetation index (NDVI), land surface temperature (LST) and temperature vegetation drought index (TVDI), based on Landsat satellite images, and combined them with climate, topography, and soil property variables. We then used random forest (RF), residual regression kriging (RRK), and ordinary regression kriging (ORK) models to predict SOC content. The spatial distribution of SOC content in a typical wavy plain area of Northeast China was mapped for the first time at a resolution of 30 m, using data from 2708 soil sampling points and 18 environmental variables. The results indicated that: ① The best prediction model in this study area was the RF model (R² = 0.68, RMSE = 2.96, MAE = 2.38) with RS data as supporting information. ② Over the past three decades, SOC content had shown an upward trend, although its spatial distribution consistently displayed higher values in the east and lower values in the west. The area with rising SOC content was twice as large as the area with falling SOC content, accounting for 63.5 % of the total area. ③ Climate and elevation were the main factors influencing the spatial distribution of SOC content. In this study, new environmental variables based on RS inversion algorithms were introduced, compensating for the loss of prediction accuracy of topographic variables in spatial modeling. The spatial prediction accuracies of the RF, RRK, and ORK models were improved by 12 %, 30 %, and 3 %, respectively. This study can serve as a reference for high-precision SOC mapping in similar areas and assist relevant authorities in managing ecosystems and improving SOC content in these regions.