Comparing three approaches of spatial disaggregation of legacy soil maps based on the Disaggregation and Harmonisation of Soil Map Units Through Resampled Classification Trees (DSMART) algorithm

Abstract

Abstract. Enhancing the spatial resolution of pedological information is a great challenge in the field of digital soil mapping (DSM). Several techniques have emerged to disaggregate conventional soil maps initially and are available at a coarser spatial resolution than required for solving environmental and agricultural issues. At the regional level, polygon maps represent soil cover as a tessellation of polygons defining soil map units (SMUs), where each SMU can include one or several soil type units (STUs) with given proportions derived from expert knowledge. Such polygon maps can be disaggregated at a finer spatial resolution by machine-learning algorithms, using the Disaggregation and Harmonisation of Soil Map Units Through Resampled Classification Trees (DSMART) algorithm. This study aimed to compare three approaches of the spatial disaggregation of legacy soil maps based on DSMART decision trees to test the hypothesis that the disaggregation of soil landscape distribution rules may improve the accuracy of the resulting soil maps. Overall, two modified DSMART algorithms (DSMART with extra soil profiles; DSMART with soil landscape relationships) and the original DSMART algorithm were tested. The quality of disaggregated soil maps at a 50 m resolution was assessed over a large study area (6775 km2) using an external validation based on 135 independent soil profiles selected by probability sampling, 755 legacy soil profiles and existing detailed 1:25 000 soil maps. Pairwise comparisons were also performed, using the Shannon entropy measure, to spatially locate the differences between disaggregated maps. The main results show that adding soil landscape relationships to the disaggregation process enhances the performance of the prediction of soil type distribution. Considering the three most probable STUs and using 135 independent soil profiles, the overall accuracy measures (the percentage of soil profiles where predictions meet observations) are 19.8 % for DSMART with expert rules against 18.1 % for the original DSMART and 16.9 % for DSMART with extra soil profiles. These measures were almost 2 times higher when validated using 3×3 windows. They achieved 28.5 % for DSMART with soil landscape relationships and 25.3 % and 21 % for original DSMART and DSMART with extra soil observations, respectively. In general, adding soil landscape relationships and extra soil observations constraints allow the model to predict a specific STU that can occur in specific environmental conditions. Thus, including global soil landscape expert rules in the DSMART algorithm is crucial for obtaining consistent soil maps with a clear internal disaggregation of SMUs across the landscape.

Hybrid Machine Learning for Integrating Pedological Knowledge into Digital Soil Mapping to Advance Next-Generation Earth System Models

“…Numerous studies in Europe (Poggio and Gimona, 2017;Ballabio et al, 2016;Tóth et al, 2017;Adhikari et al, 2014), Africa (Ramifehiarivo et al, 2017;Akpa et al, 2016), North and South America (Padarian et al, 2017;Guevara et al, 2018; [This is a non-peer reviewed preprint submitted to EarthArXiv] Hartemink et al, 2012), and Oceania (Teng et al, 2018;Gray et al, 2016) used DSM to reduce soil mapping costs over large areas. More specifically, some of them used 3D radar products to acquire high spatial resolution soil information either through data extrapolation using regressors (Adhikari et al, 2014) or disaggregation with machine learning (ML) techniques (Ellili-Bargaoui et al, 2020). Some of these studies contributed to existing regional datasets (Teng et al, 2018) or global datasets such as the GlobalSoilMap project (Ballabio et al, 2016;Rahmati et al, 2018).…”

Digital Terrain Model Derivatives Analysis with the Aim of Identifying Specific Soil Types in Young Post-Glacial Topography with a Vector Approach

“…SOTER, SOVEUR, EGSDB). In the last decade, machine learning methods such as random decision forest (RF) (Vaysse and Lagacherie 2017, Wadoux et al 2019a, Ellili-Bargaoui et al 2020) and, more recently, artificial neural networks of deep learning (DL) (Behrens et al 2018, Wadoux et al 2019b) have become principal methods of digital soil mapping. These methods emphasize interdepend-ence between environmental and topographic variables, referring to the holistic (ecological) approach in classical soil modelling, which makes use of filtering terrain attributes based on different neighborhood sizes, as well as identification of topographic features in varied scales utilising so-called octaves (a multiscale approach).…”

Geotechnologies applied in digital soil mapping

“…Our approach of using the Digital Soil Mapping products to predict the SMUs is similar to the Disaggregating and Harmonisation of Soil Map Units Through Resampled Classification Trees (DSMART) developed by Odgers et al (2014) and extensively applied in some studies (Ellili- Bargaoui et al, 2020;Ellili Bargaoui et al, 2019;Møller et al, 2019;. The difference is that the DSMART uses as predictors the environmental variables necessary to predict the soil attributes in DSM, while our approach is the first attempt to use the predict soil attributes as predictors of the SMUs.…”