Latent Conditional Diffusion-Based Data Augmentation for Small-Sample Hyperspectral Prediction of Forest Soil Organic Carbon.

Accurate forest soil organic carbon (SOC) monitoring is essential for forest soil quality assessment and carbon-sink evaluation. Visible-near-infrared (Vis-NIR) hyperspectral sensing provides rapid and information-rich measurements for SOC prediction, but obtaining sufficiently large labeled soil-spectral datasets remains difficult because field sampling, sample preparation, and reference SOC determination are labor and time intensive. This study developed a latent conditional diffusion-based data augmentation framework for SOC prediction from hyperspectral sensor data. A total of 248 forest red-soil samples from Guangxi, China, were measured using laboratory Vis-NIR reflectance spectroscopy over 350-2500 nm and divided by the Kennard-Stone algorithm into a 174-sample modeling set and a fixed 74-sample validation set. Four generative models, including VAE, GAN, WGAN-GP, and the proposed hyperspectral latent conditional denoising diffusion implicit model (HsDDIM), were evaluated using spectral visualization, t-SNE distributions, maximum mean discrepancy (MMD), Fréchet Inception Distance (FID), and downstream prediction performance. Unlike joint spectral-label generation, HsDDIM treats SOC as an external condition and generates spectra in the latent space under specified SOC conditions; the SOC condition itself is not generated by the diffusion process. Among the compared augmentation strategies, HsDDIM showed the closest distributional agreement with the real spectral samples according to MMD and FID, with values of 0.0806 and 0.5281, respectively. Without augmentation, FD1-SVR achieved the best validation result (R