Detecting salinity hazards within a semiarid context by means of combining soil and remote-sensing data

Abstract The aim of this work has been to carry out salinity mapping within the environmental context of the Lower Cheliff Plain (40,000 ha) in Algeria, where soil salinity appears to be a major threat to agricultural production. Spatio-temporal monitoring of soil salinity must therefore be urgently implemented in order to evaluate the progression of salinity hazards or the effectiveness of remediation strategies. In the present study, an extensive set of 3980 soil salinity data elements, systematically sampled on a 250-m grid, was used to assess various mapping methods based on ground measurements alone (ordinary kriging) or on a combination of ground measurements and remote-sensing data (regression–kriging method from classification and salinity index images). The accuracy of the predictions was tested using a validation set of 597 points. Eleven indices were derived from a 20-m resolution Spot XS image taken during the sampling campaign in summer 1997. Vegetation indices (NDVI) proved to be poor predictors of soil salinity within this context. Salinity indices were more closely correlated with measured values, yet significantly underestimated the salinity of zones with high levels of salt exposure. Moreover, mapping based on land-use classification does not lend sufficient accuracy, even though land use categories discriminate soil electrical conductivity in highly-saline areas. Even in this latter case however, the surface areas of highly-saline zones were still underestimated. Ordinary kriging (OK) using ground data exclusively displayed better performance than classification and simple regression methods derived from the Spot image. Nevertheless, the OK method still resulted in underestimation of the high-salinity areas. The regression–kriging method, which combines remote-sensing data with EC ground measurements, was analysed herein. This method has given rise to significant improvements in salinity estimations, as compared to purely-regressive approaches. Regression–kriging systematically provided the best validation statistics (bias, accuracy, rank of method). This approach should enable more precise spatio-temporal monitoring of soil salinity in arid areas through the combination of remotely-sensed data and ground-based monitoring networks.