Digital mapping of soil organic carbon in permafrost regions over the Qinghai-Tibet Plateau based on deep learning and hyperspectral imaging

The Qinghai-Tibet Plateau (QTP) harbors significant amounts of soil organic carbon (SOC) in the permafrost regions, which are at risk of release as carbon dioxide or methane under global warming, amplifying the greenhouse effect. Despite this, long-term investigations into the spatiotemporal dynamics of SOC in the QTP's permafrost regions remain scarce. Furthermore, spatial scale mismatches between SOC maps and thermokarst landscape maps hinder a comprehensive understanding of carbon cycling mechanisms in these landscapes. Hyperspectral data, with its superior spectral richness, offers the potential to more precisely capture soil spectral characteristics, enhancing the accuracy of SOC estimations. However, the limited availability of long-term hyperspectral datasets for the QTP presents a major challenge to leveraging this technology for SOC estimation.

In this study, we developed a physically constrained hyperspectral generative model that integrated spectral response functions and diffusion models, utilizing satellite data from Landsat 5 TM, Landsat 7 ETM+, Landsat 8 OLI, and EO-1 Hyperion imagery. This method generated high-accuracy hyperspectral data (MSSIM = 0.96, PSNR = 38.65) for the permafrost regions of the QTP from 2000 to 2020, with a spatial resolution of 30 m and a spectral resolution of 10 nm. Leveraging these generated hyperspectral data, we constructed spectral indices and incorporated climate, topography, and soil characteristics into a dual-input convolutional neural network model. This model enabled the mapping of the spatiotemporal distribution of SOC in the 0-3 m layer across the QTP’s permafrost regions from 2000 to 2020 with resolution of 30 m. Compared to existing approaches, our model achieved a 22.9% improvement in the accuracy of SOC estimation in permafrost regions, highlighting its potential for advancing carbon estimation.