Recent studies demonstrated the capability of Sentinel-2 (S2) data to estimate topsoil properties and highlighted the sensitivity of these estimations to soil surface conditions depending on the S2 acquisition date. These estimations are based on Bottom of Atmosphere (BOA) reflectance images, obtained from Top of Atmosphere (TOA) reflectance values using Atmospheric Correction (AC) methods. AC of optical satellite imagery is an important pre-processing stage before estimating biophysical variables, and several AC methods are currently operational to perform such conversion. This study aims at evaluating the sensitivity of topsoil clay content estimation to atmospheric corrections along an S2 time series. Three AC methods were tested (MAJA, Sen2Cor, and LaSRC) on a time series of eleven Sentinel-2 images acquired over a cultivated region in India (Karnataka State) from February 2017 to June 2017. Multiple Linear Regression models were built using clay content analyzed from topsoil samples collected over bare soil pixels and corresponding BOA reflectance data. The influence of AC methods was also analysed depending on bare soil pixels selections based on two spectral indices and several thresholds: the normalized difference vegetation index (NDVI below 0.25, 0.3 and 0.35) and the combination of NDVI (below 0.3) and Normalized Burned Ratio 2 index (NBR2 below 0.09, 0.12 and 0.15) for masking green vegetation, crop residues and soil moisture.

First, this work highlighted that regression models were more sensitive to acquisition date than to AC method, suggesting that soil surface conditions were more influent on clay content estimation models than variability among atmospheric corrections. Secondly, no AC method outperformed other methods for clay content estimation, and the performances of regression models varied mostly depending on the bare soil pixels selection used to calibrate the regression models. Finally, differences in BOA reflectance among AC methods for the same acquisition date led to differences in NDVI and NBR2, and hence in bare soil coverage identification and subsequent topsoil clay content mapping coverage. Thus, selecting S2 images with respect to the acquisition date appears to be a more critical step than selecting an AC method, to ensure optimal retrieval accuracy when mapping topsoil properties assumed to be relatively stable over time.

最近的研究证明了 Sentinel-2 (S2) 数据估计表土特性的能力，并强调了这些估计对土壤表面条件的敏感性，具体取决于 S2 采集日期。这些估计基于使用大气校正 (AC) 方法从大气顶部 (TOA) 反射值获得的大气底部 (BOA) 反射率图像。光学卫星图像的 AC 是估计生物物理变量之前的重要预处理阶段，目前有几种 AC 方法可以执行这种转换。本研究旨在评估表土粘土含量估计对沿 S2 时间序列的大气校正的敏感性。测试了三种 AC 方法（MAJA、Sen2Cor、和 LaSRC）对 2017 年 2 月至 2017 年 6 月在印度（卡纳塔克邦）的一个耕地区域采集的 11 张 Sentinel-2 图像的时间序列进行了分析。使用从裸土像素上收集的表土样本分析的粘土含量建立了多元线性回归模型和相应的 BOA 反射率数据。还根据基于两个光谱指数和几个阈值的裸土像素选择分析了 AC 方法的影响：归一化差异植被指数（NDVI 低于 0.25、0.3 和 0.35）以及 NDVI（低于 0.3）和归一化烧毁比的组合2 指数（NBR2 低于 0.09、0.12 和 0.15）用于掩盖绿色植被、作物残留物和土壤水分。使用从裸土像素上收集的表土样本和相应的 BOA 反射率数据分析的粘土含量建立了多元线性回归模型。还根据基于两个光谱指数和几个阈值的裸土像素选择分析了 AC 方法的影响：归一化差异植被指数（NDVI 低于 0.25、0.3 和 0.35）以及 NDVI（低于 0.3）和归一化烧毁比的组合2 指数（NBR2 低于 0.09、0.12 和 0.15）用于掩盖绿色植被、作物残留物和土壤水分。使用从裸土像素上收集的表土样本和相应的 BOA 反射率数据分析的粘土含量建立了多元线性回归模型。还根据基于两个光谱指数和几个阈值的裸土像素选择分析了 AC 方法的影响：归一化差异植被指数（NDVI 低于 0.25、0.3 和 0.35）以及 NDVI（低于 0.3）和归一化烧毁比的组合2 指数（NBR2 低于 0.09、0.12 和 0.15）用于掩盖绿色植被、作物残留物和土壤水分。

首先，这项工作强调回归模型对采集日期比对 AC 方法更敏感，这表明土壤表面条件对粘土含量估计模型的影响比大气校正中的变异性更大。其次，在粘土含量估计方面，没有任何 AC 方法优于其他方法，回归模型的性能主要取决于用于校准回归模型的裸土像素选择。最后，同一采集日期 AC 方法之间 BOA 反射率的差异导致 NDVI 和 NBR2 的差异，从而导致裸土覆盖识别和随后的表土粘土含量绘图覆盖。因此，根据采集日期选择 S2 图像似乎比选择 AC 方法更关键，