Multi-Modal Vision Transformer for high-resolution soil texture prediction of German agricultural soils using remote sensing imagery

The quantification and mapping of important soil properties, such as soil texture, are vital for effective crop management and the assessment of overall soil health in agricultural systems. In this study, we propose a multi-modal Visual Transformer (MMVT) architecture to predict and map the soil particle size distribution of agricultural topsoils in Germany at a high spatial resolution of 10 meters. Our modeling utilized multi-source bare soil satellite image composites with terrain and soil-related covariates. To optimize the model’s ability to capture spatial soil context, various image sizes were evaluated. The study findings highlighted the effectiveness of our MMVT model, demonstrating improved estimation accuracies compared to a two-dimensional Convolutional Neural Network (2D CNN) and a Random Forest (RF) model. Specifically, the proposed transformer network achieved the highest averaged validated accuracy in predicting the soil texture when incorporating a contextual image surrounding of 320 × 320 m around the soil sampling positions (Sand: R 2 = 0.74, RMSE = 14.78%, and RPIQ = 3.52, Silt: R 2 = 0.73, RMSE = 12.36%, and RPIQ = 3.50, Clay: R 2 = 0.52, RMSE = 6.30%, and RPIQ = 1.95). This integrated approach underscores the potential of advanced deep learning techniques and multi-modal learning in providing comprehensive insights into soil characteristics with high resolution and at a large scale. • Predicted soil texture across Germany at 10-m resolution. • Introduced a modified Vision Transformer (MMVT) for multi-modal soil analysis. • Improved accuracies for MMVT when spatial context is integrated. • MMVT mapping showed general soil texture trends and detailed field-scale variances.