Uncertainty of soil reflectance retrieval from SPOT and RapidEye multispectral satellite images using a per-pixel bootstrapped empirical line atmospheric correction over an agricultural region

- ELM validation accuracy and uncertainty was assessed for soil reflectance retrieval with SPOT4, SPOT5, RapidEye images.
- Field spectra at ∼30 sites across 221 km2 enabled to build 1000 bootstrap training/validation datasets for OLS regressions.
- Median validation RMSE was <2-3% reflectance, bias <|1|% reflectance and uncertainty of error around 14%.
- Best models were achieved for RapidEye and SPOT5 with spectra measured within small field site size (7 m2).
- Worst models were obtained for oblique SPOT4, late winter plough, with spectra measured within large (100 m2) field sites.

Many authors have reported the use of empirical line regression between field target sites and image pixels in order to perform atmospheric correction of multispectral images. However few studies were dedicated to the specific reflectance retrieval for cultivated bare soils from multispectral satellite images, from a large number (≥15) of bare field targets spread over a region. Even fewer were oriented towards additional field targets for validation and uncertainty assessment of reflectance error. This study aimed at assessing ELM validation accuracy and uncertainty for predicting topsoil reflectance over a wide area (221 km2) with contrasting soils and tillage practices using a set of six multispectral images at very high (supermode SPOT5, 2.5 m), high (RapidEye, 6.5 m) and medium (SPOT4, 20 m) spatial resolutions. For each image and each spectral band, linear regression (LR) models were constructed through a series of 1000 bootstrap datasets of training/validation samples generated amongst a total of about 30 field sites used as targets, the reflectance measurements of which were made between −6 days/+7 days around acquisition date. The achieved models had an average coefficient of variation of validation errors of ∼14%, which indicates that the composition of training field sites does influence performance results of ELM. However, according to median LR-models, our approach mostly resulted in accurate predictions with low standard errors of estimation around 1-2% reflectance, validation errors of 2-3% reflectance, low validation bias (<|1|% reflectance). The best results were obtained for SPOT5 and RapidEye images the spatial resolution of which is likely to better match the size of the sampled field sites. The worst results (higher median RMSE values 3.1-4.8%) were yielded for shortwave-infrared bands of SPOT4 images acquired in March: in agricultural areas, images programmed during periods when most field tillage operations have resulted in smooth seedbed conditions (April in this study) are in favour of better performances of soil reflectance prediction. Nevertheless, directional effects appear to mainly and moderately affect the global performance of near-infrared and SWIR bands-models except for oblique viewing images (viewing angle > |20°|). The predictions obtained from median LR-models through per-pixel bootstrapped ELM approach were as accurate as the ATCOR2 predictions with default parameters for the RapidEye image and were slightly more accurate and less biased for the SPOT4 images.