Mapping soil units within a vineyard using statistics associated with high-resolution apparent soil electrical conductivity data and factorial discriminant analysis

One of the key issues for digital soil mapping is how to make use of the large amount of information contained in data generated from high-resolution soil sensor surveys. Usually, analysis is based only on the sensor response(s) however descriptors of the data may also provide useful information for soil classification and mapping. Following a commercial soil apparent electrical conductivity survey of a vineyard in southern France, which generated more than 1000 data ha− 1, 8 non-spatial and 5 spatial statistics were derived at the block level (~ 1 ha/block) from the high-resolution sensor survey data. These block statistics were then ranked using Wilks' Criterion according to how well they explained the variance in an existing expert-derived soil map of the vineyard. The vineyard contained three distinct pedological units. Variance was best explained by a spatial statistic associated with the east–west component of a trend surface then indicators of the magnitude and variance of the sensor response within blocks. Given the nature of the three soil units in the vineyard, clear pedological reasons for the selection of these statistics were determined. A cross-validation procedure was used to identify the best factorial discriminant analysis model to predict soil units across the vineyard. This factorial discriminant analysis was compared to, and significantly outperformed, a linear discriminant analysis using only the mean sensor response. An analysis of the difference between the factorial discriminant analysis output and the original expert-defined soil map showed that the factorial discriminant analysis map actually identify errors in the original map when considering only the top 1 m of the soil profile. The results demonstrated that the statistics associated with high-resolution data within defined polygons (such as fields or soil units) provided useful information for soil classification and should be considered in prediction models. However, these relationships and models are likely to be location specific and the potential scales of application of this approach are not well understood at the moment.