Salt content retrieval in on leaf texture and color using machine learning-driven

To facilitate the rapid and cost-effective assessment of soil salinity in drip-irrigated cotton fields, this study presents an analytical approach leveraging smartphone-captured RGB imagery. Phenotypic color and texture features were extracted from cotton leaf images and used to invert and predict soil electrical conductivity (EC) via Decision Tree (DT) and Random Forest (RF) models. Texture features were derived using multi-scale Gabor filters and subsequently quantified using gray-level co-occurrence matrices. The results show that the fusion of color and texture features significantly improved prediction accuracy and further confirm that the RF model outperformed the DT model. The RF-based “RGB+TF” inversion model achieved a mean coefficient of determination (R²) of 0.79, a mean root mean square error (RMSE) of 243.87 μS/cm, and a mean absolute error (MAE) of 190.52 μS/cm on the validation set, representing an 11.27% improvement in prediction accuracy over the best color-only model. Overall, this method demonstrates that integrating color and texture features can effectively enhance soil salinity inversion in cotton fields and provides a valuable technical reference for the rapid, non-destructive monitoring of other field parameters in different cropping systems. • Introduced a CCM for leaf image preprocessing to reduce light interference for accurate color phenotyping across phenophases. • We combined Gabor and GLCM features to extract cotton leaf texture for soil salinity inversion. • A RF multi-model with color-texture features estimated soil salinity, achieving R²=0.79±0.05 for accurate dynamic prediction.