A Novel Self-Supervised Learning Method for Accurate Multigranularity Prediction of Soil Organic Carbon

An accurate prediction of the soil organic carbon (SOC) is crucial for better understanding the global carbon cycle. Conventional methods using machine learning (ML) approaches primarily focus on the use of covariates at the sample position level, which ignores scene information from the surroundings and wastes severely the collected global covariates, eventually achieving limited model performance. Additionally, the corresponding mapping results are generated with single granularity, which are no longer suitable for broader multiscale research supported by SOC. Therefore, this study proposed a scene-aware self-supervised learning (SASSL) framework that contains two stages: pixel-based mask autoencoders (PMAEs) and image-based fine-tuning (IFT). PMAE extracts latent representations from the source covariates through global covariate mask reconstruction. The IFT receives samples organized into images and outputs SOC prediction through model fine-tuning. Experiments conducted in South Africa indicated that the SASSL can achieve accurate SOC prediction, reducing root mean square error (RMSE) by 9.29%–14.63% and improving $R ^{2}$ by 42.06%–106.43% compared with a series of advanced models. The SASSL exhibited accelerated convergence, increased stability in prediction, and good generalization to different study areas and data sources. Additionally, SASSL can generate accurate mapping results with multiple granularities, rendering that it is more suitable for multiscale digital soil mapping (DSM). This study emphasized the effectiveness of exploiting scene information surrounding the samples in DSM, both theoretically and experimentally, and the importance of introducing a self-supervised learning (SSL) paradigm into DSM to construct efficient models that can achieve high-performance prediction of SOC and other soil properties. Code and data are now available at https://doi.org/10.17605/OSF.IO/UM3YN.