Regional prediction of soil organic carbon content over temperate croplands using visible near-infrared airborne hyperspectral imagery and synchronous field spectra

• A VIS-NIR 1 m—AISA-Eagle image was evaluated for mapping soil organic carbon.
• 74 field samples/spectra (28 synchronous) spread over 221 km2-croplands were used.
• Best model = cross-validated PLSR constructed with the synchronous raw spectra.
• Regional validation RMSE range of 3.73–4.49 g Kg−1 and ∼4 g Kg−1 in median.
• Best model achieved regional 3.73/within-field 1.44 g Kg−1 validation RMSEs.

This study aimed at identifying the potential of Vis-NIR airborne hyperspectral AISA-Eagle data for predicting the topsoil organic carbon (SOC) content of bare cultivated soils over a large peri-urban area (221 km2) with both contrasted soils and SOC contents, located in the western region of Paris, France. Soil types comprised haplic luvisols, calcaric cambisols and colluvic cambisols. Airborne AISA-Eagle data (400–1000 nm, 126 bands) with 1 m-resolution were acquired on 17 April 2013 over 13 tracks. Tracks were atmospherically corrected then mosaicked at a 2 m-resolution using a set of 24 synchronous field spectra of bare soils, black and white targets and impervious surfaces. The land use identification system layer (RPG) of 2012 was used to mask non-agricultural areas, then calculation and thresholding of NDVI from an atmospherically corrected SPOT image acquired the same day enabled to map agricultural fields with bare soil. A total of 101 sites sampled either in 2013 or in the 3 previous years and in 2015 were identified as bare by means of this map. Predictions were made from the mosaic AISA spectra which were related to topsoil SOC contents by means of partial least squares regression (PLSR). Regression robustness was evaluated through a series of 1000 bootstrap data sets of calibration-validation samples, considering 74 sites outside cloud shadows only, and different sampling strategies for selecting calibration samples. Validation root-mean-square errors (RMSE) were comprised between 3.73 and 4.49 g Kg−1 and were ∼4 g Kg−1 in median. The most performing models in terms of coefficient of determination (R2) and Residual Prediction Deviation (RPD) values were the calibration models derived either from Kennard-Stone or conditioned Latin Hypercube sampling on smoothed spectra. The most generalizable model leading to lowest RMSE value of 3.73 g Kg−1 at the regional scale and 1.44 g Kg−1 at the within-field scale and low bias was the cross-validated leave-one-out PLSR model constructed with the 28 near-synchronous samples and raw spectra.