autoRA: An Algorithm to Automatically Delineate Reference Areas—A Case Study to Map Soil Classes in Bahia, Brazil

• Published in: Land (Basel) Vol. 14; no. 3; p. 604
• Main Authors: Rodrigues, Hugo, Ceddia, Marcos Bacis, Vasques, Gustavo Mattos, Grunwald, Sabine, Babaeian, Ebrahim, Villela, André Luis Oliveira
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.03.2025
• Subjects: Accuracy; Algorithms; Automation; Case studies; Datasets; Delineation; Digital Elevation Models; digital soil mapping; Environmental gradient; Geology; Geomorphology; Heterogeneity; Hypotheses; Mapping; Prediction models; previously mapped areas; Redundancy; Reproducibility; soil class mapping; Soil mapping; Soil profiles; Soil properties; Soil sciences; Soil surveys; Taxonomy; Brazil; Netherlands
• ISSN: 2073-445X; 2073-445X
• DOI: 10.3390/land14030604

The reference area (RA) approach has been frequently used in soil surveying and mapping projects, since it allows for reduced costs. However, a crucial point in using this approach is the choice or delineation of an RA, which can compromise the accuracy of prediction models. In this study, an innovative algorithm that delineates RA (autoRA—automatic reference areas) is presented, and its efficiency is evaluated in Sátiro Dias, Bahia, Brazil. autoRA integrates multiple environmental covariates (e.g., geomorphology, geology, digital elevation models, temperature, precipitation, etc.) using the Gower’s Dissimilarity Index to capture landscape variability more comprehensively. One hundred and two soil profiles were collected under a specialist’s manual delineation to establish baseline mapping soil taxonomy. We tested autoRA coverages ranging from 10% to 50%, comparing them to RA manual delineation and a conventional “Total Area” (TA) approach. Environmental heterogeneity was insufficiently sampled at lower coverages (autoRA at 10–20%), resulting in poor classification accuracy (0.11–0.14). In contrast, larger coverages significantly improved performance: 30% yielded an accuracy of 0.85, while 40% and 50% reached 0.96. Notably, 40% struck the best balance between high accuracy (kappa = 0.65) and minimal redundancy, outperforming RA manual delineation (accuracy = 0.75) and closely matching the best TA outcomes. These findings underscore the advantage of applying an automated, diversity-driven strategy like autoRA before field campaigns, ensuring the representative sampling of critical environmental gradients to improve DSM workflows.