Leveraging EnMAP for building soil reflectance composites with Sentinel-2

Abstract

Information on European soils and their chemical and physical characteristics are essential to achieve the ambitious goal to have all European soils in a healthy condition by 2050 (defined in the EU soil health law). In recent years, soil compositing techniques based on multispectral satellite archives have been developed and established to generate input data for spectral and digital soil mappings. The surface reflectance composites (SRC) select bare soil pixels from a multitemporal data stack by using spectral index thresholds. However, due to the limited spectral information of multispectral systems (e.g. Sentinel-2), residuals from non-photosynthetically active vegetation (NPV) cannot be fully excluded. This might also impact the quality of the soil parameter models. The novel idea presented here is to use the quantitative outputs of the semi-operational fractional vegetation cover processor (fCover) to select bare soil pixels from a Sentinel-2 time series and thus, overcome threshold-based indices. fCover provides quantitative measures of photosynthetically active vegetation (PV), non-photosynthetically active vegetation (NPV) and bare soil (BS) from hyperspectral satellite images (e.g. EnMAP, PRISMA). However, the EnMAP-based outputs cover a small portion of a Sentinel-2 scene and also just provide information for selective Sentinel-2 scenes in time. In this work, a modified deep learning model Hybrid-SN was trained using S2 images as inputs and EnMAP-based fCover maps as labels to predict fCover for the complete Sentinel-2 scene. The predicted optimized S2 fCover outputs are then used to define bare soils in each Sentinel-2 scene as input for the subsequent temporal compositing. The resulting SRCs are compared to those developed by the threshold-based Soil Composite Mapping Processor (SCMaP). SCMaP is a fully automated approach to make use of per-pixel based bare-soil compositing. The difference is quantified based on an evaluation technique developed for comparing different SRCs. By exploiting synergies of hyperspectral derived products and the comprehensive S2 archive using an innovative Deep Learning approach, the selection of undisturbed bare soil areas can be enhanced and thus, the derivation of soil information can be improved.