Significance of Planet SuperDove and refined Sentinel-2 imagery fusion for enhanced soil organic carbon prediction in croplands

When RS images from multisource specifically at high spatial and spectral resolution, are integrated, the generated imagery is believed to provide higher spatial, spectral, and temporal resolutions. Though image fusion techniques have been employed in many other fields, their applicability in soil science for the estimation of soil properties, including soil organic carbon (SOC), remains limited, especially where digital soil mapping (DSM) models using machine learning algorithms (MLA) are employed. This study explores the viability of enhancing the spectral capability of high spatial resolution imagery acquired from the PlanetScope SuperDove (PSD), which has low spectral capability, by integrating it with high spectral resolution imagery from the Sentinel-2 (S2B) satellite through an image fusion technique. The main aim is to use the fused data and topographic features from the STRM DEM to assess the predictive performance of SOC across large, diverse, and erodible cropland. Prediction models were established using the data sets separately, fused, and with or without the STRM data. Two MLAs were used, including regularised random forest (RRF) and Gaussian process regression (GPR). Correlation and homogeneity tests were conducted between the S2B bands and measured SOC values before their incorporation to obtain refined S2B data for the raster fusion approach. The results show that the optimal SOC content prediction comprised the incorporation of STRM data to the fused data, as input, using the GPR model, where the lowest RMSE of 3.3 gkg-1, the highest coefficient of determination (R²) of 0.83, and the MAE of 3.6 gkg-1 were obtained. In terms of SOC spatial distribution map, the fused datasets supplemented by STRM data employing the GPR model performed better than the other alternatives. In summary, this study highlights the promising potential of image fusion of high spatial and spectral RS images to improve the estimation model of SOC, which has the potential to be widely implemented in erodible cropland area.