Comparative Study of Back-Propagation Artificial Neural Network Models for Predicting Salinity Parameters Based on Spectroscopy Under Different Surface Conditions of Soda Saline–Alkali Soils

• Published in: Agronomy (Basel) 2024-10, Vol.14 (10), p.2407
• Main Authors: Jing, Yating, You, Xuelin, Lu, Mingxuan, Zhang, Zhuopeng, Liu, Xiaozhen, Ren, Jianhua
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial neural networks; back propagation artificial neural network (BP-ANN); Back propagation networks; Comparative studies; Cracking (chemical engineering); Crop yield; Desiccation; desiccation cracks; Electric properties; Electrical conductivity; Electrical resistivity; Environmental monitoring; Ethylenediaminetetraacetic acid; Measurement methods; Neural networks; Precipitation; Prediction models; Principal components analysis; Process parameters; Remote monitoring; Remote sensing; Saline soils; Salinity; Salinity effects; Salinization; Salt; Salts; Soil productivity; Soil remediation; Soil salinity; soil salinization; Soil surfaces; Soils; Soils, Salts in; Songnen Plain; Spectral reflectance; Spectral sensitivity; Spectroscopic analysis; Spectroscopy; Spectrum analysis; Surface layers
• ISSN: 2073-4395; 2073-4395
• DOI: 10.3390/agronomy14102407

Soil salinization typically exerts a highly negative influence on soil productivity, crop yields, and ecosystem balance. As a typical region afflicted by soil salinization, the soda saline–alkali soils in the Songnen Plain of China demonstrate a clear cracking phenomena. Nevertheless, the overall spectral response to the cracked soil surface has scarcely been studied. This study intends to study the impact of salt parameters on the soil cracking process and enhance the spectral measurement method used for cracked salt-affected soil. To accomplish this goal, a controlled desiccation cracking experiment was carried out on saline soil samples. A gray-level co-occurrence matrix (GLCM) was calculated for the contrast (CON) texture feature to measure the extent of cracking in the dried soil samples. Additionally, spectroscopy measurements were conducted under different surface conditions. Principal component analysis (PCA) was subsequently performed to downscale the spectral data for band integration. Subsequently, the prediction accuracy of back-propagation artificial neural network (BP-ANN) models developed from the principal components of spectral reflectance was compared for different salt parameters. The results reveal that salt content is the dominant factor determining the cracking process in salt-affected soils, and that cracked soil samples had the highest model prediction accuracy for different salt parameters rather than uncracked blocks and 2 mm comparison soil samples. Furthermore, BP-ANN prediction models combining spectral response and CON were further developed, which can significantly enhance the prediction accuracy of different salt parameters with R2 values of 0.93, 0.91, and 0.74 and a ratio of prediction deviation (RPD) of 3.68, 3.26, and 1.72 for soil salinity, electrical conductivity (EC), and pH, respectively. These findings provide valuable insights into the cracking mechanism in salt-affected soils, thereby advancing the field of hyperspectral remote sensing for monitoring soil salinization. Furthermore, this study also aids in enhancing the design of spectral measurements for saline–alkali soils and is also helpful for local soil remediation with supporting data.