An innovative approach for integrating two-dimensional conversion of Vis-NIR spectra with the Swin Transformer model to leverage deep learning for predicting soil properties

The application of visual-near-infrared and shortwave-infrared (VNIR-SWIR) diffuse reflectance spectroscopy for soil properties analysis is increasingly gaining popularity due to its rapid, cost-effective, and non-destructive nature. In particular, deep learning models have been found to perform exceptionally well for large spectra libraries. This study proposes a novel approach to enhance the deep learning approach that involves converting one-dimensional spectra into two-dimensional (2D) spectral images. We investigated several methods, such as cutting reshape (CR), Gramian angular difference field (GADF), Gramian angle sum field (GASF), and Markov transition field (MTF). We then combine these converted images with the Swin Transformer model to predict a range of soil properties. Furthermore, we compare our proposed method with existing techniques reported in the literature. The results showed that the root mean square error (RMSE) of predicting soil organic carbon content (OC (g·kg−1)), nitrogen content (N (g·kg−1)), cation exchange capacity (CEC (cmol+ ·kg−1)), pH, and sand (%) and clay content (%) using Gram's angular difference field (GADF) and Swin Transformer were 23.25, 1.26, 8.55, 0.54, 15.33, 6.14, and determination coefficients R2 of 0.95, 0.92, 0.79, 0.90, 0.74 and 0.84, respectively. This study introduces a new perspective to enhance deep learning models for soil spectroscopy.