Estimating the soil salinity over partially vegetated surfaces from multispectral remote sensing image using non-negative matrix factorization

• DOI: 10.1016/j.geoderma.2019.113887
• Publication Date: 2019-07-30T16:56:21Z
• Authors: Ya Liu; Fangfang Zhang; Changkun Wang; Shiwen Wu; Jie Liu; Aiai Xu; Kai Pan; Xianzhang Pan
• Keywords: Thematic Mapper; Multispectral Scanner; Multispectral pattern recognition

Abstract Multispectral remote sensing technique has been extensively applied in recent years for the detection of soil salinity on bare soil; however, multispectral remote sensing is restricted in areas covered with vegetation, largely due to the mixed pixel problem. In the present study, non-negative matrix factorization (NMF) was implemented to separate soil spectral signal from mixed pixels of Landsat 5 Thematic Mapper (TM) to further estimate the soil salinity in a partially vegetated area. Four methods, namely, partial least squares regression (PLSR), least-squares support vector machine (LS-SVM), back propagation neural network (BPNN), and random forest (RF), were applied and compared. The results showed that the NMF-separated soil spectra could greatly improve the prediction accuracy compared with the mixed spectra, and among the four modeling methods, RF performed better than the rest of the methods, with the averaged results of determination of the prediction R2p = 0.67, a root mean square error of the prediction RMSEp = 0.73 ms cm−1, and the ratio of the standard deviation to RMSEp RPD = 1.61 after 100 times of random sampling and modeling. This approach could propose a new method for accurate and timely monitoring of soil salinity in a partially vegetation-covered area.