Multi-parameter Soil Property Prediction Incorporating Mid-infrared Spectroscopy and Dropout Sequential Artificial Neural Network

Given the facilitation of efficient soil data acquisition, light diffraction in both field and laboratory settings allows for applying infrared spectroscopy. This leads to the development of soil spectral library at regional and international levels owing to the extensive interest in the mid-infrared spectroscopy (MIR) domain. Spectroscopy practices meritoriously evaluate various soil constituents such as total nitrogen (TN), organic carbon (OC), potassium (K), and phosphorus (P) within the mid-infrared range, utilizing direct spectral responses and advanced modeling, mostly while analyzing fresh soil samples (undisturbed, wet). Machine and deep learning approaches potentially revolutionize soil spectral data modeling, demonstrating their transformative impact in various fields of study. A novel technique called DrSeq-ANN (dropout sequential artificial neural network), which falls under DL algorithms for predicting soil properties based on raw soil spectra, is proposed and evaluated in this investigation. The National Soil Survey Center-Kellogg Soil Survey Laboratory of the United States Department of Agriculture (USDA) database, comprising nearly 860 topsoil measurements from Kansas State having biological and physicochemical parameters, was employed. DrSeq-ANN outperformed other algorithms when fed to the pre-processed data with the help of techniques such as initial derivative, inverse derivative, logarithmic transformation with a base of 10 (Log10x), and logarithmic derivative. Specifically, while forecasting soil organic carbon, the DrSeq-ANN algorithm achieved R2 value of 0.79 and RMSE value of 0.03 with the logarithmic pre-processed method. With the competencies of the ANN model, DrSeq-ANN proved to be more accurate in prediction. The study confirmed that DrSeq-ANN can be trained in a multi-task setting to forecast the 4 soil factors (TN, OC, P, and K).