Estimating salt content of vegetated soil at different depths with Sentinel-2 data.

ABSTRACT

The accurate and timely monitoring of the soil salt content (SSC) at different depths is the prerequisite for the solution to salinization in the arid and semiarid areas. Sentinel-2 has demonstrated significant superiority in SSC inversion for its higher temporal, spatial and spectral resolution, but previous research on SSC inversion with Sentinel-2 mainly focused on the unvegetated surface soil. Based on Sentinel-2 data, this study aimed to build four machine learning models at five depths (0∼20 cm, 20∼40 cm, 40∼60 cm, 0∼40 cm, and 0∼60 cm) in the vegetated area, and evaluate the sensitivity of Sentinel-2 to SSC at different depths and the inversion capability of the models. Firstly, 117 soil samples were collected from Jiefangzha Irrigation Area (JIA) in Hetao Irrigation District (HID), Inner Mongolia, China during August, 2019. Then a set of independent variables (IVs, including 12 bands and 32 spectral indices) were obtained based on the Sentinel-2 data (released by the European Space Agency), and the full subset selection was used to select the optimal combination of IVs at five depths. Finally, four machine learning algorithms, back propagation neural network (BPNN), support vector machine (SVM), extreme learning machine (ELM) and random forest (RF), were used to build inversion models at each depth. The model performance was assessed using adjusted coefficient of determination (R2 adj ), root mean square error (RMSE) and mean absolute error (MAE). The results indicated that 20∼40 cm was the optimal depth for SSC inversion. All the models at this depth demonstrated a good fitting (R2 adj ≈ 0.6) and a good control of the inversion errors (RMSE < 0.16%, MAE < 0.12%). At the depths of 40∼60 cm and 0∼20 cm the inversion performance showed a slight and a great decrease respectively. The sensitivity of Sentinel-2 to SSC at different depths was as follows 20∼40 cm > 40∼60 cm > 0∼40 cm > 0∼60 cm > 0∼20 cm. All four machine learning models demonstrated good inversion performance (R2 adj > 0.46). RF was the best model with high fitting and inversion accuracy. Its R2 adj at five depths were between 0.5 to 0.68. The SSC inversion capabilities of all the four models were as follows RF model > ELM model > SVM model > BPNN model. This study can provide a reference for soil salinization monitoring in large vegetated area.