A Rule‐based System to Map Soil Properties

Abstract

Conventional soil mapping is limited in its capabilities in that it presents a summary of the soil surveyor's conceptual view of soil variation. As such, the method conveys little regarding what is known about the variation of individual soil properties, or the quantitative nature of their variation. We developed a new method for soil mapping, based on the concepts employed in the PROSPECTOR mineral exploration system, which builds on existing soil surveyor knowledge to construct quantitative statements about individual soil properties via the development of a network of rules. These rules operate within a system of Bayesian inference to assign the varying probability of occurrence of a soil property of interest within an area, given evidence that relates to it in a known way. Permissible evidence includes the range of attributes normally used by a soil surveyor, such as landform, vegetation, land use, or parent material, and can also include remotely sensed digital data. Evidence is weighted according to the uncertainty associated with it, and combined to produce a single estimate of probability of a given attribute. The relationship between the evidence and prediction is stated explicitly at each stage of the procedure and is thus repeatable in a consistent manner. The system has the advantage that while it does not discard the evidence and knowledge used in conventional soil survey, it produces quantitative estimates of the distribution of soil properties, which can be used for a wide range of applications. The data produced is amenable to storage in geographic information systems and related data bases. As such, it can be updated or enhanced as new information or knowledge becomes available.

Soil Spectral Mapping and Its Correlation With the Traditional Methodology

“…Thus, it was possible to obtain the map of Figure 6. By the observation of Figure 6 there was no relationship between the clusters generated by multivariate analysis and soil classes found in the field by conventional methodology, confirming the difficulty of cartographic representation of the wide range of neighboring classes already observed by Cook et al (1996). Eight clusters were obtained by clustering analysis, larger number when compared to the soil classes(six).…”

Updating the 1:50,000 Dutch soil map using legacy soil data: A multinomial logistic regression approach

“…It has become widely recognized that tacit knowledge of the soil-landscape system provides valuable information that should be integrated into the digital soil mapping process (Heuvelink and Webster, 2001;McKenzie and Gallant, 2007;Walter et al, 2007). Such knowledge can be used to build expert systems for mapping soils (Cook et al, 1996;Zhu et al, 2001) or to define a conceptual model of pedogenesis that forms the foundation of a quantitative (statistical) model for digital soil mapping (McKenzie and Ryan, 1999;McKenzie and Gallant, 2007). In case of regression modeling, use of knowledge of the soil-landscape system should be fully integrated throughout the process of model-building.…”

Untangling crop management and environmental influences on wheat yield variability in Bangladesh: An application of non-parametric approaches

“…A range of non-parametric analytical methods can be applied to comprehend spatial data variability, including Bayesian rule-based systems (Cook et al, 1996;Skidmore et al, 1996), artificial neural networks (Park and Vlek, 2002), linear mixed effects and generalized linear models (Bolker et al, 2008), and recursive partitioning approaches such as classification and regression tree analysis, or CART (Tittonell et al, 2008). Of these, linear mixed effects models and binary recursive partitioning approaches that construct decision 'trees' weighting the importance of predictor variables and their effect on yield responses, appear to be among the most commonly used in the agricultural sciences (Ferraro et al, 2009;Raman et al, 2011;Virk and Witcombe, 2008;Yang, 2010;Zheng et al, 2009).…”