Spatial Prediction and Performance Comparison of Soil Organic Carbon in High-Latitude Forest Soils Based on Multiple Feature Selection Methods and Machine Learning Models

Soil organic carbon (SOC) in permafrost regions is a critical component of terrestrial carbon reservoirs, and spatially explicit multi-depth mapping is essential for identifying subsurface carbon emission hotspots and supporting carbon neutrality accounting. The Kamalan River Basin in the Greater Khingan Mountains, situated in a discontinuous permafrost zone with relatively high SOC stocks and strong sensitivity to climate warming, was selected as the study area. A total of 90 field sampling points were collected and integrated with Sentinel-1 Synthetic Aperture Radar (SAR) imagery, Sentinel-2 multispectral imagery, and topographic data, yielding 238 environmental features. Three predictive models, Random Forest (RF), Geographically Weighted Random Forest (GWRF), and One-Dimensional Convolutional Neural Network (1D-CNN), were combined with three feature selection methods (Boruta, RF importance, and Shapley) to construct 27 model–feature–depth combinations across three soil layers (0–10 cm, 10–20 cm, and 20–30 cm). RF feature selection combined with RF model achieved the highest prediction accuracy at 0–10 cm (R2 = 0.93, RMSE = 16.58 g/kg), with its bootstrap aggregation strategy effectively suppressing overfitting under small-sample and high-dimensional conditions. 1D-CNN showed relatively unstable performance, which may be associated with feature selection strategy, feature ordering sensitivity, a potential mismatch between its inductive bias and tabular environmental data. Feature importance analysis suggested that bulk density (BD) and soil moisture content (SMC) were among the most influential predictors across all depth layers, with elevation, radar backscatter coefficients, and Sentinel-1 texture features demonstrating relatively high cross-method and cross-depth stability. Boruta excluded BD and SMC at 0–10 cm, which may be partly attributable to inter-predictor redundancy, a strict independent-contribution threshold, and insufficient statistical power under small-sample conditions. For spatial mapping, the best-performing combination for each model was applied: RF feature selection with RF model, RF feature selection with GWRF model, and RF feature selection with 1D-CNN model. SOC maps revealed considerable spatial heterogeneity, with relatively high values concentrated in valley lowlands and lower values in high-altitude forest zones; mean SOC declined from 93.08 g/kg at 0–10 cm to 69.88 g/kg at 20–30 cm. Despite the limited number of sampling points (n = 90), these findings provide practical guidance for model and feature selection in SOC mapping of permafrost forests and a scientific basis for carbon sink assessment under climate warming scenarios. Conservation efforts should prioritize riparian zones and high-elevation areas with elevated SOC concentrations, while vegetation protection and reduced soil disturbance are recommended in low-SOC regions to help limit further carbon loss.