Spatial Mapping of Soil Potassium in a Hilly Area of Southern China Based on the Geospatial Transformer Model Adapted to Small-Sample Data

As the underlying model of ChatGPT-4V and scientific computing models (e.g., AlphaFold 3), the application of Transformer in the field of soil property mapping is worth exploring. In this study, we aim to explore the application value of Transformer models in soil potassium mapping. To achieve this objective, random forest (RF), feedforward neural network (FNN), one-dimensional convolutional neural network (1D-CNN), and geospatial Transformer (GT) models were used to map the spatial distribution of total potassium and available potassium in regional soil. In addition, the explanatory SHAP model was used to elucidate the contribution of each environmental variable to the model. The GT model exhibited the best effect and accuracy in modeling total and available potassium. The model improved the total potassium prediction accuracy to 0.77, representing increases of 208%, 156.67%, and 41.54% over the RF model, FNN model, and 1D-CNN model, respectively. It also improved the available potassium prediction accuracy to 0.42, representing improvements of 16.02%, 88.34%, and 19.66% over the RF model, FNN model, and 1D-CNN model, respectively. The geographical-feature explanatory model revealed that the red edge index exhibited a synergistic response to the spatial distribution of total potassium and available potassium. In the modeling of available potassium, it was found that the accuracy of GT model driven by ‘ physical laws ‘ rather than pure data was not only higher than that of ordinary deep learning model, but also higher than that of machine learning model, which broke the traditional rule that the accuracy of machine learning model was higher than that of deep learning model using small samples and a few feature variables, and confirmed the huge mapping potential of GT model for soil attribute data.