Bridging the gap between laboratory VNIR-SWIR spectra and Landsat-8 bare soil composite image for soil organic carbon prediction

International interest is focusing on how to better manage soil organic carbon (SOC) to increase resilience to climate change and reinforce food security. Remote sensing imagery and visible, near-infrared, and short-wave infrared (VNIR-SWIR) spectroscopy combined with advanced modeling algorithms can monitor SOC in an efficient, low-cost, and environmentally friendly manner. However, the limited bands (multi-spectral imagery) as well as the confounding effects such as soil moisture and surface roughness from the satellite may jeopardize the spectral prediction of SOC. We proposed a framework to integrate laboratory spectra with Landsat-8 multispectral data collected in natural real-environments to implement the satellite hyperspectral simulation for SOC prediction and mapping. A soil spectral library (SSL) with 873 samples containing SOC and VNIR-SWIR reflectance was developed in Northeast China. Two pixelwise temporal mosaicking approaches (i.e., averaged bare soil conditions [Median] and dry soil conditions to exclude anomalous values [R90]) were compared for creating bare soil composite images. Fractional-order derivative was explored for spectral preprocessing. Results indicated Median approach outperformed R90 regarding both spectral correlation and the validation performance for SOC. Raw reflectance after hyperspectral image simulation with 201 bands developed by random forest algorithm improved the validation R² from 0.38 and 0.29 to 0.49 and 0.32 for Median and R90 approaches, respectively, as compared to raw reflectance with six bands. Laboratory spectra yielded higher predicted accuracy for SOC than Median-simulated, followed by R90-simulated hyperspectral images. The convolutional neural network models developed with 0.75-order, 0.75-order, and 1.25-order one-dimensional spectra had the optimal validated performances for laboratory, Median-simulated, and R90-simulated hyperspectral data, respectively, with the corresponding validation R² values of 0.79, 0.68, and 0.54. Our study highlights that legacy SSLs show huge potential for integrating with current and forthcoming satellite multispectral images to maximize the predictive performance for SOC monitoring and mapping.