?
Updating legacy soil map reveals spatial soil-class probabilities
Legacy soil maps remain essential for national land-use planning. However, their thematic accuracy and spatial resolution are often insufficient for contemporary applications requiring reliable estimates. This study presents a digital soil-mapping approach to update legacy soil information without extensive field sampling, demonstrated for arable lands in the Republic of Tatarstan (4.5 million ha). The proposed approach integrates the 1:2.5 M Soil Map of Russia with multisource predictors derived from principal component analysis of MODIS time-series imagery (2013–2025). These predictors represent landscape invariants associated with stable productivity, thermal, and moisture regimes, complemented by topographic, climatic, and parent material covariates. Principal component analysis was applied to extract stationary environmental regimes, while linear discriminant analysis was used to identify stable boundaries between soil-forming conditions within this invariant space. Linear discriminant analysis produced a refined predictive soil map (250-m resolution) with four units corresponding to Russian soil types and World Reference Base classes: Albic Retisols, Albic Luvisols/Greyic Phaeozems, Chernozems, and Fluvisols. Validation against 166 independent soil profile observations yielded an overall accuracy of 59.6% and a Kappa coefficient of 0.4, indicating moderate agreement beyond chance. Unit-specific accuracies ranged from 75.0% (Albic Retisols) to 50.6% (Albic Luvisols/Greyic Phaeozems). The lowest performance for Albic Luvisols/Greyic Phaeozems reflects their transitional characteristics in the forest–steppe ecotone. Probability maps showed spatial variation in confidence for soil-unit assignments. Unlike purely predictive machine-learning approaches, this framework emphasizes the identification of stable relationships between environmental regimes and soil units rather than maximizing classification accuracy alone.