Pedometrics Forums 10 PM challenges Can we incorporate mechanistic pedological knowledge in digital soil mapping?

Viewing 3 posts - 1 through 3 (of 3 total)
  • Author
  • #2547

    [Proposed by Gerard Heuvelink]
    Most digital soil mapping algorithms are to a high degree empirical. And this is only increasing, now that we entered the data science era and rely heavily on machine learning algorithms. Pedological knowledge only creeps in when we adopt the CLORPT model to identify relevant covariates. Structural equation modelling makes an attempt to move away from purely data-driven approaches and Bayesian networks may be useful too, but ideally we would make use of dynamic, mechanistic models of soil forming processes. Can we do that? This is a huge challenge because the input variables and parameters of these models are often poorly known, and also the model structure (and ‘optimal’ degree of complexity) is far from obvious. Hydrologists are much further than we are with methods to deal with parameter and structural uncertainties, such as through Bayesian calibration and Bayesian model averaging. Is this the way forward? Or should we be looking for ways to incorporate expert knowledge that is in the heads of soil surveyors and pedologists?

    Nicolas Saby

    [By Lin Yang]
    Soil is the result of the interaction of environmental factors including climate, parent material, topography, and organisms over millions of years. Environmental factors differ across years and locations. The forming processes of soil are very complicated. Knowledge on those processes relevant for soil formation is limited, thus it is not easy to build mechanistic pedological models.

    Despite the difficulty of building mechanistic models, there are still some efforts on this topic. For example, Finke and Hutson (2008) developed a SoilGen1 model to simulate soil development in calcareous loess at Holocene (15, 000 BP-present) temporal extent. The core of this model is the LEACHM model for water and solute transport. Several functions (such as cycling of C) were added to the core model to cover the wider range of pedogenetical processes. Vanwalleghem et al. (2013) presents a model, named Model for Integrated Landscape Evolution and Soil Development (MILESD), which describes the interaction between pedogenetic and geomorphic processes. They believed that soil formation is closely related to Landscape evolution. Their model included soil formation processes including weathering and clay translocation, and the lateral redistribution of soil particles through erosion and deposition was combined. The model also simulated the vertical variation in soil horizon depth, soil texture and organic matter content.
    Due to the key role of soil in critical zone, mechanistic mathematical models have been developed considering soil functions and processes. We can borrow experiences from critical zone research. For example, Giannakis et al. (2017) developed the 1D Integrated Critical Zone (1D-ICZ) mode. This model simulated the coupled processes that underpin major soil functions including water flow and storage, biomass production, carbon and nutrient sequestration, pollutant transformation, and supporting biological processes. It coupled with pedotransfer functions to predict bulk soil properties, thus dynamically links soil structure characteristics and hydraulic soil properties. Furthermore, the model can simulate and quantify four main soil ecosystem functions.

    Mechanistic models for soil genesis are not easily operationalized because the temporal variation of boundary conditions is not easy to reconstruct, calibration is difficult and computational demands may be high. Some “compromise” approaches between mechanistic models and machine learning methods have been applied for soil mapping . Structural equation modelling (SEM) is one of those, which draws people’s attention recently. SEM has its roots in social sciences, and integrates empirical information with mechanistic knowledge by deriving the model equations from known causal relationships (Angelini et al., 2016). It also includes a graphical form. Besides, it uses knowledge about interrelationships between soil properties and predicts these properties simultaneously. SEM cannot reproduce the true physical, chemical and biological processes. Instead, the model structure in SEM is based on hypotheses of the functioning of a soil-landscape system that is formalized in a conceptual model. SEM thus can be used as a tool to understand the interactions of the soil-landscape system, its genesis and functioning.
    SoLIM (Soil Land Inference Model) (Zhu et al., 2001) is also a framework to incorporate pedological knowledge into soil mapping. In SoLIM, expert knowledge is converted and formalized as membership functions (curves) based on fuzzy logic. The success of SoLIM first depends on whether the knowledge of soil experts is comprehensive. Extensive field work is usually needed to obtain the understanding of relationships on soil and its environmental factors.

    Although pedological knowledge is not used as input for machine learning approaches, those approaches could generate knowledge that is helpful to understand the relationships between soil and its environment factors. The generated decision trees and partial dependence plots based on training samples could indicate those relationships. However, the reliability of those relationships relies on the representativeness of training samples.

    In my opinion, to incorporate pedological knowledge is a promising direction of digital soil mapping. However, we need to have pedological knowledge at first. Experiences of the previous studies on existing models could be a good start. Go to the field to know soil is the direct way to understand the soil-landscape system. Also, machine learning is useful as it helps to generate knowledge. References: Angelini, M.E., Heuvelink, G.B.M , Kempen, B , Morrás, H.J.M., 2016. Mapping the soils of an Argentine Pampas region using structural equation modelling. Geoderma 281, 102-118.

    Finke, P.A., Hutson, J.L., 2008. Modelling soil genesis in calcareous loess. Geoderma 145, 462–479.

    Giannakis, G.V., Nikolaidis, N.P., Valstar, J., Rowe, E.C., Banwart, S.A., 2017. Chapter Ten: Integrated Critical Zone Model (1D-ICZ): A Tool for Dynamic Simulation of Soil Functions and Soil Structure. Advances in Agronomy 142, 277-314.

    Vanwalleghem, T., Stockmann, U., Minasny, B., McBratney, A.B., 2013. A quantitative model for integrating landscape evolution and soil formation. Journal of Geophysical Research Atmospheres 118(2).

    Zhu, A.X., Hudson, B., Burt, J.E., Lubich, K., 2001. Soil mapping using GIS, expert knowledge, and fuzzy logic. Soil Science Society of America Journal 65, 1463-1472.

    Nicolas Saby

    [From Philippe Lagacherie, INRA, France)

    Having begun my career as a soil surveyor, I appreciate well the loss of information from the soil surveyor’s mental model to any empirical DSM model, however sophisticated it is. Most of this loss lays in the story of soil cover genesis that a soil surveyor builds along his (her) prospection and uses as a powerful way for extrapolating the observations. Rather than searching the Holly Grail, i.e. the mechanistic model of soil forming process that would predict any soil properties at any locations that have little chance to become operational for the reason mentioned by Gerard, I prefer looking to less ambitious solutions e.g. considering in our models the old idea stating that the soil cover is made up of nested systems or completing the quantitative evaluation of DSM models by assessing the plausibility of the predicted soil patterns considering our knowledge of soil forming processes.

Viewing 3 posts - 1 through 3 (of 3 total)
  • You must be logged in to reply to this topic.