Simulation of Sentinel-2 data using Hyperspectral Data for Bare Surface Soil Moisture Estimation

Soil moisture is one of the driving variables for precision irrigation management, drought assessment etc. Various approaches for measuring and monitoring soil moisture have been devised in the past, including laboratory techniques, point-based sensors, and remote sensing. However, the use of simulated optical data for bare surface soil moisture estimation is limited to date. The main objective of this study is the assessment of simulated Sentinel-2 data for bare surface soil moisture estimation. Hyperspectral soil moisture data provided by [1] in the spectral range of 450-950 nm over the test site in Germany has been utilized in this study. Hyperspectral images (50 x 50 pixels) were captured using the Cubert UHD 285 hyperspectral snapshot camera. The data available for public use consist of mean spectra of soil surface (average of 50 x 50 pixels) in 125 spectral bands. Simulation of 8 bands (B2-B8A) of Sentinel2 covering the same spectral range as that of hyperspectral camera was carried out using spectral response curve of Sentinel-2. Random Forest Regression based model was developed using all bands. Moreover, important features were selected based on a percent increase in mean squared error (%IncMSE). Selected bands include B2, B4, B5 and B7. Evaluation of models developed using all the 8 bands and selected 4 bands was carried out using the independant testing data based on Root Mean Sqaure Error (RMSE). We have obtained an RMSE of 1.01 and R
²
value of 0.91 for a model with 8 bands. However, RMSE was reduced to 0.96 and R
²
was increased to 0.94 in the case of selected 4 bands. In addition to this, validation of the model developed using the selected simulated bands was carried out using actual Sentinel-2 satellite observations on the TCS Demo Farm located in Pune, India. Estimated soil moisture was compared with the In-situ soil moisture observations available during Nov-Dec 2017. The difference between actual and estimated soil moisture was found to be between -2.10 to 3.18 %. The obtained outcomes depict that the results are significant and models developed on simulated Sentinel-2 data can be applied on Sentinel-2 satellite observations to estimate bare surface soil moisture.