Researchers in China used a generative AI model called WGAN-GP to improve soil organic carbon mapping from drone hyperspectral data, even with only 76 lab-analyzed soil samples. The method lifted prediction accuracy by 13.7% over the best conventional model, pushing the performance ratio into the good range. The approach could lower the cost of measuring soil carbon across farmland, which matters for soil health and carbon credit verification. Not every AI method worked. A diffusion model produced no usable synthetic samples at this sample size, and borrowing 3,000 samples from an external spectral library actively hurt predictions, a reminder that more data is not always better. The study's authors note the results come from one region and one soil type, so real-world gains will depend on testing across more landscapes.
