Improved spatiotemporal monitoring of soil salinity using filtered kriging with measurement errors: An application to the West Urmia Lake, Iran

Real-time monitoring of soil salinity based on field samples and laboratory analyses is a costly and time demanding procedure, so that sound methods that could reduce the burden by making use of cheaper data would be a step toward a more sustainable salinity hazards monitoring system on the long run. Typically, this involves replacing presumably error-free laboratory salinity measurements with indirect measurements that are however affected by various source of uncertainties, and these uncertainties need to be accounted for in order to avoid compromising the quality of the final results. More specifically, in a spatiotemporal prediction framework where salinity maps need to be drawn repeatedly at various time instants and where salinity, values need to be compared over time for agricultural areas that are prone to salinity hazards, it is of major importance to process these uncertainties in a sound way, as failing to do so would impair our ability to detect salinity changes at an early stage for taking preventive actions. The aim of this paper is to propose a filtered kriging framework that allows the user to rely on cheap field sampled electrical conductivity (EC) measurements, that cannot however be assumed as error-free. Field EC measurements need to be calibrated from laboratory measurements and the corresponding calibration errors cannot be neglected. Moreover, when sampling is repeated over time, positioning errors are quite common and can adversely impact the results due to the inclusion of an extra variability source. It is shown how these uncertainties can be quantified and successfully processed afterwards for improving both the reliability of the spatial predictions and temporal comparisons of soil salinity. The idea is to rely on a same general optimal linear predictor that can be easily adapted to get rid of these unwanted effects. The procedure is illustrated by using a rich data set of EC measurements that cover a time span of seven years in the western part of Urmia Lake, northwest Iran. From these data, it is shown how calibration errors can be considered as spatially independent and zero-mean Gaussian distributed, while laboratory measurements exhibit a clear spatial structure but are also affected by a not inconsiderable spatial nugget effect, which is in turn impacting the errors for field EC measurements due to the positioning errors. By relying on a linear optimal predictor that reduces here to filtered kriging with measurement errors, it is shown that filtering out these two random effect components clearly improves the quality of the results when it comes to map EC values and to detect changes that occurred over time. Comparing filtered values for the successive sampling campaigns provided evidence that a major salinity shift did occur between autumn 2011 and autumn 2014 while the other parts of the area were left unchanged by comparison. From this study, it can be concluded that even if the only errors involved in this work were linked to calibration and positioning errors, the methodology is general enough to process various sources of uncertainties in general. It is thus a valuable tool for practitioners, with a field of potential applications that goes beyond the framework of salinity monitoring. (C) 2017 Elsevier B.V. All rights reserved.