Mapping of Soils Salinity with Landsat 8 OLI Imagery and Random Forest Algorithm

Soil salinity mapping is essential for sustainable land development and water resources management. In situ sampling is time-consuming, laborious, and restricted by geographical conditions. Therefore, an efficient and accurate model is necessary to monitor and assess the spatio-temporal dynamic salinization at regional a scale. In this study, Shule River Basin (SLRB) is taken as an example to develop the soil salinity mapping model based on Landsat 8 OLI images using random forest (RF) algorithms. A series of extended soil salinity indexes (ESSIs) were generated by combining any two, three, or four spectral bands were combined in expressions that include one or more of the arithmetic operations: addition, subtraction, multiplication, division, square and rooting form. The features selected from ESSIs outperformed the features selected from soil salinity indexes (SSIs) used in references. The best selected indexes are (B7^2-B5^2)^0.5, (B4^2+B5^2-B6^2)^0.5, (B1*B5-B4*B6/(B1*B5+B4*B6))^0.5,(B2*B6-B3*B7/( B2*B6+B3*B7))^0.5. In addition, three partition sampling methods of the training set and validation set for long-tail distribution problems are compared. The results showed that the resampling method considering the long-tail distribution performs better than systematic resampling and random k-fold cross-validation. The regional soil salinity mapping results showed that most areas are seriously salt-affected in the whole basin, especially along the river and the southeast mountainous area, where the soil salinity classes are highly and even over-extremely saline. This study could have implications for agricultural schemes planning and salinization control.