Analyzing the efficacy of different methodologies for constructing a learning sample of points for predictive soil mapping
Abstract
This article presents the results of a comparative study of the effectiveness of three methods for constructing a training sample for digital predictive soil mapping, using agricultural land in the Oshmyany District of Republicof Belarus as a case study. Based on Landsat 8 satellite imagery and a digital elevation model (FathomDEM), a wide range of predictors has been calculated, including morphometric characteristics of the relief, vegetation indices and moisture indices. Using the principal component analysis method, three of the most comprehensive characteristics were identified: distance to the water network, relief dissection depth and erosion base level. The study compares classical approaches (random distribution — 4.168 points; distribution on a regular grid – 4.633 points) with the authors’ method of stratified proportional distribution based on relief factors, followed by optimisation using the k-nearest neighbours method. The proposed method made it possible to reduce the sample size to 581 points (an 87.5% reduction compared with the conventional method used), whilst ensuring balanced representation of all soil types, proportional to their actual area. Prediction maps were constructed using the gradient boosting algorithm (XGBoost). Accuracy assessment using discrepancy matrices showed that the model, trained on a sample formed taking soil-forming factors into account, achieved the highest overall accuracy (87.3%) and kappa coefficient (0.873), significantly outperforming the random distribution method (82%, kappa 0.67) and grid-based methods (79%, kappa 0.62). The average accuracy across the three validation runs for the proposed method was 62.5%, which is also higher than that of the alternative approaches (59.8% and 58.3% respectively). The results obtained confirm that taking into account the spatial structure of the key soil-forming factors is critical for improving the reliability of forecasts and optimising field costs.Received: 01/26/2026
Accepted: 08/26/2026
Accepted date: 11/20/2026
Keywords: spatial sampling; digital soil mapping; k-means method; remote sensing; morphometric characteristics of relief
DOI: 10.55959/MSU0137-0944-17-2026-81-4-132-
Available in the on-line version with: 18.11.2026
-
To cite this article:

This work is licensed under a Сreative Commons Atribiution - NonCommercial 4.0 International (CC BY-NC 4.0)

