Using deep neural networks for evaluation of soil quality based on VIS–NIR spectroscopy.

This study is focused on evaluating the potential of deep neural networks for assessing soil properties based on VIS–NIR spectroscopy with spectral wavelength ranges of 350–2500 nm and 10 nm resolution on a global scale. The dataset was provided by the ICRAF-ISRIC soil spectral library and consists of 4438 samples from 58 countries. In This research we used the powerful one-dimensional (1D) convolutional neural network (CNN) models to predict soil organic carbon (OC), pH, calcium carbonate (CaCO3), cation exchange capacity (CEC), effective CEC (ECEC), sum of cations (SC), base saturation (BS), exchangeable acidity (ACe), exchangeable cations (aluminum (Al), calcium (Ca), magnesium (Mg), sodium (Na) and‌ potassium (K)), silt, sand and clay content. Also, traditional regression approaches of spectral data including partial least squares regression (PLSR), multilayer perceptron (MLP) and random forest (RF) with optimum preprocessing of spectral data were also tested and the models' performances were evaluated and compared. The optimum structure of models was determined by testing different components for PLSR, selecting the best number of neurons in the hidden layer, activation functions, and solvers for MLP, and finding the proper number of trees and maximum depth in RF. The CNN-based model estimated the OC, pH, CaCO3, CEC, ECEC, SC, BS, ACe, Al, Ca, Mg, Na, K, silt, sand and clay content with the ratio of percent deviation (RPD) of 2.98, 1.9, 2.48, 2.01, 2.05, 2.39, 1.97, 1.01, 1.59, 2.06, 1.72, 1.69, 1.29, 1.38, 1.8 and 1.97, respectively. We evaluated the performance of CNN-based methods with an effective architecture for soil spectral data analysis. The CNN model outperforms other regression algorithms in terms of accuracy and low predicting errors using non-preprocessed data. However, the use of diverse soil properties for modeling spectral data can provide sufficient information on the advantages and disadvantages of VIS–NIR data in predictions. [ABSTRACT FROM AUTHOR]