Mapping soil organic carbon content in two contrasted pedoclimatic regions by combining time series of Sentinel-2 and Sentinel-1 with Vis-NIR laboratory spectra

International audience

英语. Mapping Soil organic carbon (SOC) is essential for continuous monitoring of its spatial and temporal dynamics. In this study, we developed a method using time series of Sentinel-2 (S2) data combined with Sentinel-1 (S1) and vis-NIR laboratory spectra to map SOC content of agricultural soils of two contrasting sites. Site 1, in temperate Brittany (northwest France), and site 2, in semi-arid central Tunisia, are agricultural areas of 1.5 km² and 40 km², respectively. From each site soil samples were collected, their SOC content measured and their spectra recorded under laboratory conditions. The SOC content ranged from 15.2 to 49.4 g.kg-1 in site 1 and from 4.1 to 11.9 g.kg-1 in site 2. Deep neural network algorithms were implemented after dividing the data set constructed from the time series of S2 and S1 images into calibration (70%) and validation (30%) sets. Three random draws of the validation sets were performed to assess model robustness. Four approaches were tested: (1) models developed using S2 bands as a single input, (2) applying multiple factor analysis (MFA) to select 12 of 40 indices calculated from S2 data and adding them to the S2 bands, (3) adding soil moisture derived from the time series of S1 (SM1), and (4) progressively adding five indices calculated from laboratory spectra in descending order of their correlation with SOC. Model performance was compared based on validation results, and semi-variograms for observed and predicted SOC were then used to analyze the maps generated. Results showed that for site 1, only models using approach 4 were validated (RPIQ = 1.78±0.19 - 3.07±0.6), whereas for site 2, the models showed good performance in validation regardless of the approach used (RPIQ = 1.95±0.03 - 2.12±0.17). The S2 indices used in approach 2 differed between sites due to different spatial distributions in the MFA plane. The addition of SM1 improved model robustness for site 1, as predictive performance was stable over the 3 random draws of validation sets. In approach 4, laboratory indices showed significant correlations with SOC content only for site 1. Thus, we were able to validate our models once we added the two indices with the highest correlation. Semivariograms of the predicted values showed lower sill-to-nugget ratios but similar shapes to the observed values. Finally, the developed method allowed us to map 70% of the area studied in Site 1 and 97% of the area studied in Site 2.