Deep learning using inductively coupled plasma spectroscopy spectra accurately predicts various soil physicochemical properties for soil diagnosis

Improving soil diagnosis-based agriculture can help reduce fertilizer utilization and its environmental impact. However, conventional soil diagnostic methods are time-consuming and expensive, which limits their application. Although various rapid soil testing methods have been suggested, their accuracy remains largely unexplored. Herein, multiple soil parameters were predicted using the spectral data obtained from inductively coupled plasma (ICP) spectroscopy combined with deep learning. We analyzed 1941 soil samples from seven countries with various land-use patterns and histories. All ICP wavelength spectral data from the 1 M NH₄OAc extract were used for deep learning. The targeted soil properties included exchangeable bases (Ca, Mg, K, and Na); pH (H₂O); pH (KCl); electrical conductivity; available P (Bray1-P); exchangeable Al; cation exchange capacity; total carbon, nitrogen, clay, and sand contents. The predicted soil parameters were consistent with the observations. Most soil parameters had determination coefficients (R²) of > 0.9, and the lowest R² (0.81; total carbon) was relatively high. To our knowledge, this is the first study to demonstrate the prediction of multiple soil parameters using the ICP spectra of soil extracts. Our accurate predictions indicate that this method can be applied for precise, affordable, and rapid soil diagnosis, which could enhance soil-diagnosis-based agriculture.