Improving the prediction accuracy of soil organic matter: Addressing the challenge of soil moisture variability

Abstract

• Developed 3 soil moisture correction strategies using Sentinel-2 time series; MinSM combined with random forest achieved high-precision SOM mapping (R 2 = 0.73, RMSE = 7.19 g/kg, MAE = 5.46). • Identified mean annual temperature and spectral variables as key drivers of SOM distribution, providing a basis for regional soil health management. • Our 10-m resolution SOM map provides a crucial data for implementing Nature-based Solutions (NbS), particularly for agricultural ecosystems facing soil degradation and climate change. Soil organic matter (SOM) serves as a critical indicator for assessing land degradation and regulating the global carbon cycle, particularly in the Songnen black soil region—a critical agricultural zone where accurate SOM mapping is vital for sustainable land management. However, soil moisture variability significantly affects spectral reflectance in remote sensing data, thus hindering the precision of SOM estimation and limiting support for evidence-based ecological management. To address this limitation, in this study, the Google Earth Engine cloud platform was utilized to extract bare soil pixels from over 20,000 Sentinel-2 images (2020–2023) via multiple index thresholds. Three image compositing strategies were developed to mitigate the influence of soil moisture: maximum bare soil composite (MaxBS), maximum reflectance composite (MaxRef), and minimum soil moisture composite (MinSM). Machine learning models, incorporating these composites and environmental covariates, were constructed to predict SOM. Results revealed that the MinSM strategy most effectively reduced the influence of soil moisture, with the random forest model produced the most robust predictions (R 2 = 0.73, RMSE = 7.19 g/kg, MAE = 5.46)— outperforming other available open-source SOM products. Annual mean temperature was identified as the most important individual predictor of SOM, whereas the spectral variables collectively accounted for 46.21 % of model variance. The derived high-resolution SOM map revealed values ranging from 10.8 to 87.3 g/kg, with a significant southwest-to-northeast gradient. These findings demonstrate that the rigorous identification of bare soil pixels and moisture mitigation improve SOM mapping accuracy. By providing precise SOM spatial distributions, our approach strengthens the implementation of Nature-based Solutions (NbS) (e.g., degradation restoration) in the region, particularly for land degradation remediation, and sustainable agricultural management of human-dominated landscapes.

Topics & Concepts