Estimating soil salinity using Gaofen-2 imagery: A novel application of combined spectral and textural features.

Gaofen-2 (GF-2) imagery data has been playing an important role in environmental monitoring. However, the scarcity of spectral bands makes GF-2 difficult to use in soil salinity estimation. In this paper, we combined spectral and textual features for soil salinity estimation from GF-2 imagery. The spectral features comprised five classes of predictors: spectral value, vegetation index, salinity index, brightness index, and intensity index. Four gray-level co-occurrence matrix (GLCM) indices were used as the textural features. The least absolute shrinkage and selection operator (LASSO) was applied to select features. Four methods, namely, Random forest (RF), support vector machine (SVM), back propagation neural network (BPNN), and partial least squares regression (PLSR) were applied and compared. To this end, 211 soil samples were collected in the Yellow River Delta through field investigation. The results showed that GF-2 imagery could successfully estimate soil salinity by integrating spectral and texture features, and among the four methods, the RF had the highest accuracy with the determination coefficient for cross-validation (R2CV), a root mean square error for cross-validation (RMSECV), and the ratio of the standard deviation to the root mean square error of prediction (RPD) of 0.82, 2.36 g kg-1, and 2.28, respectively. Especially, the impact of different scale features on the soil salinity estimation accuracy was evaluated. The optimal window size for features was 9 × 9 pixels, and increasing or decreasing the window size will decrease the estimation accuracy. The study provides a novel application to soil salinity estimation from remote sensing imagery.