Spectral unmixing of straw and soil to estimate the soil salinity using non-negative matrix factorization

Abstract: Straw returning to the field has been used to promote the soil ecological environment in the Yellow River Delta. A remote sensing estimation of soil salt can greatly contribute to monitor the soil salinization in the straw covered areas. However, the mixed data has posed a great challenge on the pixel accuracy, due to both straw and soil spectral signals in the same pixel. In this study, ten groups of salinized surface scenes were firstly collected with different straw coverage. Multiple spectral measurements were then carried out on each sample, where the average value was taken as the reflection spectrum. The Non-negative Matrix Factorization (NMF) was selected to treat the mixed spectrum under the non-negative restriction in the process of spectral unmixing. The endmember matrix and abundance matrix were updated simultaneously by an iterative method, and finally two endmember spectra were extracted. The spectra of soil and straw were separated to evaluate the performance of the straw removal from the mixed spectrum, according to the retaining salinity information. Finally, the soil salt model was constructed using the Partial Least Square Regression (PLSR), further to verify the effectiveness of NMF for the spectral unmixing and extracting soil spectrum. The results show that: 1) The soil spectral reflectance decreased gradually with the increase of the salinity in the whole wavelength range, but the trend of curves remained similar. More importantly, the straw coverage increased the spectral reflectance of soil. The spectrum of straw coverage soil was much more similar to the straw spectrum, as the degree of the straw cover increased. In addition, there were the absorption characteristics near 1 730 and 2 090 nm in the straw covered soil, compared with the pure soil. 2) The soil spectrum was effectively separated from the straw mixed spectrum after NMF unmixing. There were more outstanding characteristics of straw in the separated straw spectrum on the whole with the increase of straw coverage, when the degree of the straw cover reached 15%. The typical characteristics of straw were achieved in the separated straw spectrum, when the straw coverage was less than 25%. There was a significant fluctuation in the spectral curve of the extracted soil and straw, indicating the unstable soil reflection, when the straw coverage increased to more than 30%. It infers that the performance of NMF was confined to the low straw coverage in this case. 3) The soil spectra that extracted from the straw mixed spectra was successfully utilized to estimate the soil salinity. The high accuracy of the model was achieved using the soil spectral data after NMF spectral unmixing, compared with the original mixed ones. Specifically, the average coefficient of determination, R2, of the soil salt model was 0.68, the average Root-Mean-Square Error (RMSE) was 6.34, and the average ratio of Prediction Deviation (RPD) was 1.47 after NMF unmixing. Among them, the average R2 and RPD were 0.07 and 0.25 higher than the original mixing, respectively, whereas, the average RMSE was 1.22 lower. The RPD of the improved model was also greater than 1.4, when the degree of the straw cover was less than 25%. The findings can provide a practical basis to efficiently improve the near-ground remote-sensing estimation accuracy of straw covered saline lands in the Yellow River Delta.