Image-based deep learning enables flexible, holistic 3D mapping of multiple soil properties

Digital soil mapping (DSM) increasingly relies on machine learning to predict soil properties from environmental covariates, yet systematic comparisons of modern image-based deep learning (DL) architectures remain limited. Self-supervised learning (SSL) DL approaches are particularly appealing for DSM due to their flexibility and potential to pre-learn soil forming factor maps. We compared a random forest (RF) with four DL models including single-task convolutional neural networks (CNNs), multi-task CNNs, Vision Transformer (ViT) Autoencoders (AE) with SSL pre-training, and ViT Masked Autoencoders (MAE) with SSL pre-training. The AE and MAE approaches leveraged unlabeled pixel patches for pre-training, learning spatial environmental patterns before fine-tuning for soil properties. Using 2,042 soil observations and environmental rasters we evaluated models for predicting CaCO 3 , clay, ec, pH, sand, sar, silt, and soc at seven depths (0–150 cm). The best DL models had very similar Lin’s concordance (CCC) values to RF models (∼0.65 to 0.67), but improved preservation of correlation structure between soil properties. The best DL model predictions had pairwise correlations 0.02 units closer, on average, to the actual correlations in the data than RF models. Multi-task DL models trained with transformed right-skewed soil properties specifically improved maintenance of correlation structures. Of the DL architectures, MAE models and an 11x11 pixel patch size were slightly more accurate across metrics. The MAE models also were able to better represent complex patterns of regional petrocalcic horizons and landform patterns that are smoothed over by RFs. The popular horizon-based 3D RF training strategy had similar accuracy metrics to other models, but exhibited unrealistic trends in surface soil organic carbon mirroring issues documented previously with properties that exhibit strong depth signatures. We conclude that new DL approaches are strong options for operational DSM development.