Improving soil organic matter mapping using transfer learning and satellite-simulated samples from bare soil hyperspectral imagery

Calibrating an accurate soil organic matter (SOM) satellite mapping model requires a sufficient number of representative satellite samples (including actual ground-based SOM values and corresponding satellite spectral data). However, collecting these samples is challenging due to complex geological environments and inconvenient transportation conditions across large areas. To address this issue, a satellite sample simulation strategy was developed to transform ground-based local-area soil spectral samples into satellite-simulated spectral samples. Subsequently, a transfer learning (TL) approach was implemented to encode the spectra-SOM relationship learned from the satellite-simulated sample set into a basic neural network model. Finally, by fine-tuning this model with only a small number of satellite spectral samples, a robust SOM satellite mapping model was established. The results indicated that a total of 846 satellite-simulated spectral samples were generated, and the differences between the satellite-simulated spectral samples and the satellite spectral samples were minimized. The optimal TL-based SOM satellite mapping model (TL-half model) had an R2 of 0.90 and RPIQ of 4.02, representing a 12.50% improvement over the traditional SOM mapping model (R² = 0.80; RPIQ = 3.10). The TL-based SOM satellite mapping method proposed in this study offers an effective technical approach for regional SOM monitoring and global carbon storage management.