Exploring soil databases: a self‐organizing map approach

Abstract

A soil quality database (SQDB) is a collection of soil samples described by a given set of parameters, allowing farmers, scientists and other stakeholders to make informed decisions about practices, processes and policies for soil use and management. If each parameter is considered as a dimension of the space spanned by the SQDB, extracting information becomes a difficult task when the number of parameters is >3. A widely used approach to explore multidimensional data sets is the self‐organizing map (SOM) method, which is suitable for clustering, visualization and extraction of information from multidimensional data. We applied the SOM method as an exploratory technique to an unlabelled SQDB to extract knowledge – data patterns and data associations – from the data set (the time and location of each sample were unknown). The SQDB used in this study is a set of 1240 unlabelled samples within the Central Valley of Chile, covering ca 7500 km2. The predominant soils are Andisols with a large organic matter content (7–12%), small bulk densities (0.6–1.0 g/cm3) and large water‐holding capacity. We identified three patterns: (i) isolated region within the map with close neurons (smooth transitions), (ii) two or more regions with predominantly large or small values and (iii) homogeneous map with small values with an isolated region of large values. These patterns show that the data set represented more than two groups that were not necessarily related. For pH, no important associations with other investigated parameters were observed. Previous studies carried out by the local agricultural research station showed that pH values below 5.5 constrain nutrient uptake. Thus, locations presenting pH<5,5 should be subject to seasonal monitoring to assess management practices that mitigate soil acidity. The component plane for organic matter indicates that ca. 50% of the soil samples had contents <8% related to soil series characteristics and management practices. As the k‐means is initialized by random partitions, the two‐step approach (clustering the map representing the input data) is less sensitive to variations in the input data (subsamples) than is the direct application of k‐means to the input data, but it also reduces the computational cost. The ability of SOMs to visualize multidimensional data sets helps gain an understanding of the data in the exploratory phase, such as the association and integration of physical, chemical and biological parameters.

Using self-organizing maps for determination of soil fertility (case study: Shiraz plain)

“…Similarly, Merdun (2011) and Rivera et al (2015) used the SOM for clustering the soil properties. The results of the research show that the SOM represents a powerful technique for Digital Soil Mapping (Merdun 2011;Rivera et al 2015). …”

Is pH increasing in the noncalcareous topsoils of France under agricultural management? A statistical framework to overcome the limitations of a soil test database

“…Because these data sets often include many soil properties, Rivera et al . () proposed using the self‐organizing map (SOM) method for clustering, visualization and information extraction from such multidimensional data. Villamil et al .…”

“…Because many soil parameters are available in these types of data sets, similar analyses can be conducted using large sets of soil properties and can be related to changes in land use, agricultural practices or environmental conditions. Because these data sets often include many soil properties, Rivera et al (2015) proposed…”

Assessment of methods to determine soil characteristics for management and design of irrigation systems

“…Also farming practices, such as tillage and cropping, decrease bulk density on the surface due to the presence of roots and the addition of organic matter from pruning or soil amendments. Rivera et al (2015) and Dec and Dörner (2014).…”

“…Soils in Chile's Central Valley are highly variable in space, posing major problems for irrigation system design and management (Rivera et al, 2015).…”