Digital soil mapping using machine learning techniques in a varied tropical environment

Abstract

<p>Environmental models often require soil maps to represent the spatial variability of soil attributes. However, mapping soils using conventional in-situ survey protocols is time-consuming and costly. As an alternative, digital soil mapping offers a fast-mapping approach that might be used to monitor soil attributes and their interrelationships over large areas. In Brazil, conventional survey methods are still widely used, and thus maps still in development are considered as the state-of-the-art products for decades. In this study, we address this lack of updated spatial information on many soil attributes by producing regional statistical soil models using an innovative framework. This new framework attempts to reduce prediction redundancies due to high multicollinearity, by implementing a Feature Selector algorithm. This is expected to improve a model’s strength by decreasing its unexplained variance. The framework’s core is composed of the Soil-Landscape Estimation and Evaluation Program (SLEEP) and a calibrated Gradient Boosting Model capable of modelling the spatial distribution of soil attributes at multiple soil depths. These models allowed us to explain the spatial distribution of some basic soil attributes (physical and chemical), and its environmental drivers. The model training and testing approach used 30 environmental attributes, and data from 223 soil profiles for the state of Pernambuco, Brazil. Our models demonstrated a consistent potential to perform spatial extrapolations with r<sup>2</sup> ranging from 0.8 to 0.97, and PBIAS from -0.51 to 2.03. The properties related to topographic and climatic conditions were dominating when estimating the number of horizons, percentage of silt and the sum of bases (a measure of soil fertility). We believe that our framework features high flexibility, while reducing capital investments when compared to <em>in situ</em> surveys and traditional mapping protocols. These findings also have implications for the improvement and testing of pedotransfer functions. We thank FACEPE for funding this through APQ 0646-9.25/16.</p>