Random Forest-Based Fusion of Proximal and Orbital Remote Sensor Data for Soil Salinity Mapping in a Brazilian Semi-arid Region

Proximal and remote sensor data were fused and a random forest model was used to map soil salinity levels in a farm with 11 ha in the semi-arid region of Northeast Brazil. Proximal sensor data included electrical conductivity from EM38-MK2, and remote sensor included C-band radar data from the Sentinel-1 satellite. Salinity level classes derived from electrical conductivity data measured at 35 points in the study area on a 50 × 50 m grid and at three depths (0–10, 10–30 and 30–50 cm) were used to validate the salinity level predictions. The overall accuracy of the salinity level predictions lied between 0.66 and 0.74, with Kappa values between 0.43 and 0.59. Soil salinity was relatively low at 0–10 and 10–30 cm due to the implementation of a surface drainage system, whereas the 30–50 cm depth had the highest occurrence of salic soils, with a potentially harmful effect for crop growth.