Mapping soil heavy metals in a geologically complex Mediterranean forested area: an Interpretable Machine Learning approach

Mapping heavy metals (HMs) in soils has gained attention for its relevance to soil health, yet background levels in unpolluted regions remain limited. Interest is growing in complex geological settings where naturally high HM levels occur. This study maps soil chromium (Cr), nickel (Ni), and selenium (Se) in Sierra de las Nieves, Spain. Predictive feature sets included Sentinel-2 composites, spectral indices representing land cover, phenological features, terrain features derived from a Digital Elevation Model, and two geological features, the distance to both peridotites and fractures. An ensemble stacking approach combining Random Forest, Support Vector Machine, and Neural Network models was applied, with Sequential Feature Selection method for dimensionality reduction. The HM prediction was evaluated using the following Interpretable Machine Learning methods in the already-trained stacked models with Feature Selection: model-agnostic Feature Importance, Accumulated Local Effect (ALE) plots, and Shapley values. Nickel showed the highest accuracy (R 2 = 0.6, RMSE = 0.78), followed by chromium (R 2 = 0.41, RMSE = 0.68) and selenium (R 2 = 0.394, RMSE = 0.43). Distance to peridotites was key for Cr and Ni, with the samples within peridotites areas showing higher Cr and Ni values, while elevation was crucial for Se, with low content Se samples linked to higher elevations. Remote sensing features including SWIR composites for Cr and Ni and NIR composites for Se proved highly relevant. This framework offers baseline maps at the landscape scale and valuable insights into the distribution drivers.