Skip to main content
ARS Home » Research » Publications at this Location » Publication #317867

Title: Spatial interpolation quality assessments for soil sensor transect datasets

item SCUDIERO, ELIA - University Of California
item Corwin, Dennis
item MORARI, FRANCESCO - University Of Padua
item Anderson, Raymond - Ray
item Skaggs, Todd

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/22/2016
Publication Date: 3/5/2016
Publication URL:
Citation: Scudiero, E., Corwin, D.L., Morari, F., Anderson, R.G., Skaggs, T.H. 2016. Spatial interpolation quality assessments for soil sensor transect datasets. Computers and Electronics in Agriculture. 123:74-79. doi: 10.1016/j.compag.2016.02.016.

Interpretive Summary: When planning agronomical practices at the field-scale, maps of soil properties are extremely useful decision making-tools for farmers, land resource managers, agriculture consultants, extension specialists, and Natural Resource Conservation Service field staff. Maps are an approximation of reality and therefore should be interpreted accordingly. The current best practices to map soil properties use intense geospatial surveys carried out with mobile soil sensors (and, alternatively, remote sensing imagery) as proxies for soil properties. Based on limited soil samples, whose location is established from spatial variation in the geo-referenced soil sensor (or remote imagery) data, a relationship between sensor data and soil analyses data can be established. To assign a value for the target soil property across the entire field, the sensor data is interpolated using geostatistics techniques, such as kriging. Maps are characterized by two types of error: measurement and interpolation error. The measurement error comes from: human error in the soil analyses, imperfections in the sensor functionality, and from the imperfect relationships between soil analyses and soil sensor measurements. The interpolation errors are given by the performance of the geostatistical model used to estimate the soil variable in un-sampled locations. In this manuscript, we focus on the latter error type. The quality of spatial interpolations is generally assessed with leave-one-out (loo) cross validation CV. This procedure removes one observation from the dataset to validate the map created with the remaining points (i.e.,training dataset). This procedure is repeated for all points in the dataset. In this manuscript we show how the classical loo CV should be avoided as it returns CV errors that are unrealistically lower than the actual errors at un-sampled locations. This is due to the fact that the neighbors of the observation removed for validation, are very similar (in value) to the latter. In practice, the validation and training datasets are highly inter-dependent in the vicinity of the validation location. To overcome this issue, we present an R (open source statistical software) application for a spatial loo (SLOO). In the SLOO, neighbors of the validation observation are removed from the training dataset. When the radius of this circular neighborhood is selected appropriately, the SLOO CV returns errors that are the same of those from un-sampled locations. The best threshold radius for the SLOO CV can be estimated as a function of the spatial structure of the modeled variable. This manuscript provides an effective tool for assessing the accuracy of soil maps made from intensive geospatial measurements of soil sensors. The tool can be used by agriculture consultants, extension specialists, and Natural Resource Conservation Service field staff supporting farmers and land resource managers.

Technical Abstract: Near-ground geophysical soil sensors provide extremely valuable information for precision agriculture applications. Indeed, their readings can be used as proxy for many soil parameters. Typically, leave-one-out (loo) cross-validation (CV) of spatial interpolation of sensor data returns overly optimistic low prediction errors, because the left out data point has values very close to that of its neighbors in the training dataset. To obtain unbiased interpolation error estimations, a spatial loo (SLOO) CV application is proposed (in the R environment), in which the neighbors of the validation observation are excluded from the training set. The SLOO CV was tested over soil apparent electrical conductivity (4 fields in California, USA) and reflectance (1 field in Italy) sensor data. The results indicate that: i) the SLOO CV is a useful tool to assess the error of kriging prediction and ii) the SLOO threshold distance (t.dist) for neighbors exclusion is proportional to the semivariogram range and partial sill. This tool provides research scientists with an improved means of cross-validating spatial interpolations.