Integrating multi-source remote sensing and UAV data for landscape-scale soil carbon mapping: uncertainty hotspots at ecotone boundaries in forest-steppe ecosystems

In forest–steppe landscapes, soil organic carbon (SOC) varies sharply with land cover, moisture, and topography, yet its spatial heterogeneity in transitional zones remains quantitatively under-represented in landscape-scale models. This study maps topsoil (0–15 cm) SOC and prediction uncertainty in the Oka–Don Lowland of the East European Forest-steppe (Tambov Region, Russia). We collected 413 field samples using a multiple-properties-oriented representative sampling design. Predictors included UAV-LiDAR micro-topography (2 m digital terrain model), multi-year Sentinel-2 phenological parameters, vegetation and moisture indices, and soil pH. Deep learning models combining convolutional neural networks, long short-term memory networks, and multilayer perceptrons (CNN-LSTM-MLP) processing raster inputs, time-series, and point inputs were compared with conventional machine-learning methods. Model performance was assessed using spatially clustered five-fold cross-validation and uncertainty was quantified via 1000-iteration bootstrap resampling. The CNN-LSTM-MLP achieved the highest accuracy (coefficient of determination R2 = 0.67, root mean square error = 5.8 g kg−1) and narrowest 95% prediction intervals (mean prediction interval range = 23 g kg−1), with CatBoost slightly behind (R2 = 0.61). All models exhibited increased uncertainty at ecotones, with prediction interval ranges up to 50% higher than in interior biotopes. Neighbourhood-aware modelling and phenological information substantially improve SOC estimation in heterogeneous forest–steppe landscapes and demonstrate the need for targeted ecotone sampling for precise carbon accounting and land-management decisions.