Software implementation of a deep learning model for predicting soil properties based on spectroscopy data

The article presents the development and software implementation of a deep learning model for predicting soil properties based on visible and near-infrared spectroscopy data. The study’s relevance stems from modern agriculture’s need for rapid soil property analysis to optimize resource management and enhance crop yields. The objective is to create an efficient model utilizing the residual neural network architecture with a convolutional block attention module and soil acidity consideration. The LUCAS dataset, comprising spectra and physicochemical properties of over 20,000 soil samples, was used for training and testing. Methods involve data preprocessing (outlier removal, smoothing, normalization, interpolation to 500 points) and implementation in Python using the PyTorch library. The results show high accuracy in predicting organic carbon, calcium carbonates, and nitrogen (determination coefficient above 0.9), with improved predictions due to acidity integration into the attention mechanism. The novelty lies in incorporating acidity into the model and adapting it for portable spectrometers. Practical significance includes its potential application in precision agriculture systems and robotic platforms for soil monitoring.