Contribution of Sentinel-1 radar backscatter/coherence and Sentinel-2 optical data to digital mapping of soil organic carbon in the Iberian Peninsula

Accurate mapping of soil organic carbon (SOC) using optical remote sensing is often constrained by persistent cloud cover, which limits data availability in many regions. While recent studies have explored the feasibility of radar sensors for SOC mapping to overcome this limitation, they have predominantly relied on backscatter features, largely overlooking the potential of interferometric coherence. To address this gap, this study assessed the potential of synergistically using backscatter/coherence observations from Sentinel-1 and optical data from Sentinel-2 for mapping SOC across the Iberian Peninsula. Backscatter, coherence, optical, and traditional auxiliary data (terrain and climate) were utilized as input features, and their various combinations were integrated with the LUCAS 2018 soil database to develop machine learning-based SOC prediction models. We evaluated how the temporal interval of backscatter composites and the temporal baseline of coherence data affected model performance. Both radar metrics showed strong predictive power for SOC, and their temporal configurations substantially affected modeling performance. Backscatter images with a monthly interval achieved the best performance, whereas longer intervals progressively decreased predictive accuracy. Models trained on coherence with shorter temporal baselines outperformed those with longer temporal baselines. The joint use of these two radar metrics improved predictive accuracy (R 2 = 0.42), surpassing models that only used Sentinel-2 optical data (R 2 = 0.38). Our results demonstrate promising prospects of coherence/backscatter data as substitutes or complements to optical data for SOC mapping. Integrating these three complementary and relatively independent remote sensing sources notably improved model performance, achieving accuracy no lower than models based on traditional auxiliary data. Variable importance analysis indicated that radar-derived backscatter and coherence were crucial input features for SOC mapping. The contribution of backscatter to SOC prediction was influenced by polarization modes and orbital directions, with cross-polarization and ascending-orbit backscatter showing greater importance than co-polarization and descending-orbit backscatter, respectively. The mapping results derived solely from coherence and backscatter data exhibited spatial patterns broadly consistent with those obtained from optical and traditional auxiliary data. The proposed cloud computing-based workflow utilizing freely available Sentinel optical and radar imagery provides a cost-effective and reproducible approach for large-scale SOC mapping.