Enhancing saline soil quality using enriched sheep manure compost and organic fertilizer: A FTIR-Deep learning...

Soil salinity restricts nutrient uptake, degrades soil structure, and ultimately reduces crop productivity. In this study, electrical, physical, and chemical soil properties were measured using standard laboratory procedures, whereas Fourier Transform Infrared (FTIR) spectroscopy was applied to characterize functional group variations induced by soil amendments. Machine learning and deep learning (ML-DL) approaches were employed to establish quantitative relationships between spectral features and soil properties. Principal Component Analysis (PCA) was used to assess spectral clustering and compositional variation among treatments, whereas Partial Least Squares Regression (PLSR) modelled the relationship between FTIR spectra and soil parameters. A Long Short-Term Memory (LSTM) and A One-dimensional Convolutional Neural Network (1D-CNN) with SHapley Additive exPlanations (SHAP) was further implemented to enhance predictive performance and interpret influential spectral regions respectively. PCA revealed clear separation between untreated saline soils and amendment treated soils, indicating substantial compositional shifts toward productive agricultural conditions. PLSR demonstrated strong correlations between FTIR spectra and soil properties, with coefficients of determination (R²) ranging from 0.27 to 0.99 for organic fertilizer and 0.82 - 0.99 for sheep manure compost. Sheep manure showed comparatively higher predictive accuracy for dielectric constant (DC), exchangeable sodium percentage (ESP), organic carbon (OC), and available carbon (AC), with R² values ranges between 0.96 and 0.98. The 1D-CNN model outperformed than LSTM and PLSR across all parameters, achieving exceptionally high predictive accuracy (R² > 0.99 for DC and >0.995 for ESP), while OC and AC also demonstrated strong performance (R² = 0.98). SHAP analysis confirmed meaningful spectral property relationships. Overall, integrating FTIR spectroscopy with ML- DL models provide a rapid, accurate, and interpretable framework for saline soil assessment and sustainable soil reclamation strategies.