Double-stage linear spectral unmixing analysis for improving accuracy of sediment concentration estimation from MODIS data: the case of Tekeze River, Ethiopia

Abstract

Dynamic spectral properties of sediment types together with less wide river dimension compared to spatial resolution of public imageries are the challenges for modeling suspended sediment concentration (SSC) using the conventional regression remote sensing analysis. In this study, a new subpixel analysis approach called double-stage linear spectral unmixing analysis (DLSUA) was proposed for modeling the variability of SSC from the moderate-resolution imaging spectroradiometer (MODIS) data. In the first stage, the linear spectral unmixing analysis was used to unmix the ground cover components (rock, bare land and turbid water) which contributed to the reflectance of the mixed pixels. In the next stage, the spectral mixing coefficients (SMCs) of the constituents in the turbid water were determined. Finally, nonlinear regression models correlating the SSC and its SMC were generated. Furthermore, the field and laboratory-based observed SSC and reflectance data were used for verifying the model. The coefficients of determination (R2) and relative mean square error (RMSE) were used to evaluate the performance of the model. The DLSUA approach improved the simulations of SSCs (R2 = 0.83 and RMSE ± 9.96) compared to the simulations using the empirical regression remote sensing model (R2 = 0.74 and RMSE ± 16.2). Overall, the study showed that understanding the variability of spectral properties of primary components is important to use remote sensing for modeling and monitoring the SSCs using coarser remote sensing data (MODIS).