Geoderma (Sep 2024)
Spectra-based predictive mapping of soil organic carbon in croplands: Single-date versus multitemporal bare soil compositing approaches
Abstract
Sustainable cropland management requires quantitative and up-to-date information on the spatial pattern of soil organic carbon (SOC) at scales relevant for implementing targeted conservation measures. Spectra-based remote sensing of SOC in croplands is promising, but it requires the extraction of high-quality bare soil pixels that enable spatially continuous coverage. Here, we aim to compare the SOC predictive capability of single-date versus multitemporal compositing of bare soil images in an intensively cultivated region (4,700 km2) of northeast China. A series of 12 bare soil images within 2017–2022 were processed and passed onto three multitemporal compositing approaches (geometric median, and univariate mean and median) to create bare soil mosaics. Both single-date and multitemporal spectral images, together with laboratory-simulated Sentinel-2 benchmark data, were used to develop partial least squares regression, Cubist and random forest models via 100 bootstrapped validations. With Cubist consistently being the best performing algorithm for all three data sources, results show that the 12 single-date images exhibited temporally unstable performance (R2: 0.30–0.67). Among the three compositing approaches, the high-dimensional geometric median composite was the most suitable because of (i) its close resemblance to laboratory reference and robustness to outliers, which yielded a Cubist model (R2: 0.64; RMSE: 2.24 g/kg) outperforming 11 out of 12 single-date models; and (ii) its ability to retain between-band spectral relationships that allowed further incorporation of SOC-relevant bare soil indexes, which led to a further 6.5 % increase in model prediction accuracy. The resultant SOC map highlighted the capability of Sentinel-2 bare soil imaging to reveal field-scale soil degradation patterns. Future work should explore the possibility of extending the purely spectra-based framework to integrated SOC mapping and monitoring with additional soil biophysical and management information.
