Abstract:

[Objective] Soil salinization is a major environmental issue confronting global agriculture, posing a serious threat to land productivity and ecological sustainability. Soil electrical conductivity (EC) serves as a critical indicator of salinization degree, and its rapid and accurate monitoring is of great significance for the prevention and control of salinization and soil improvement. Current EC inversion models based on single data sources, such as spectra or images, have limited predictive accuracy due to constrained characterization information and susceptibility to environmental interference. The purpose is to overcome the limitations of single data sources and improve the accuracy of soil EC inversion by constructing a spectral-image fusion model. [Methods] The study area was in Baicheng city, Jilin province, within the Songnen Plain, a typical distribution area of soda saline-alkali soil. Sampling points were arranged at equal intervals in a 15 m × 15 m grid, with samples collected from a depth of 10－20 cm, resulting in a total of 209 soil samples. After indoor air-drying and sieving through a 2 mm mesh, spectral data and image data were acquired from the soil samples using an ASD FieldSpec 4 spectroradiometer and a Xiaomi 14 smartphone, respectively. Following combined preprocessing of the spectral data using first derivative (D1) and standard normal variate (SNV) transformation, the Recursive Feature Elimination with Cross-Validation (RFECV) algorithm was employed to select characteristic wavelengths. For the image data, 13 color features from five color spaces (RGB, HSV, CIELAB, CIELUV, CIELCH) were extracted, alongside 6 texture features derived using the Gray-Level Co-occurrence Matrix (GLCM). Subsequently, the Variable Importance in Projection (VIP) algorithm was used to select core image features. Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) were both utilized to construct data-level and feature-level fusion models respectively. For data-level fusion, the selected image features and characteristic wavelengths were directly concatenated as model input. The feature-level fusion models employed a symmetric dual-branch architecture, where the selected image features and spectral features were fed into two independent sub-networks, with feature fusion occurring in the intermediate layers. Model predictive performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE). [Results and Discussions] The RFECV algorithm selected 12 characteristic wavelengths, covering salt-sensitive bands and carbonate absorption regions. The VIP algorithm screened out 9 image features significantly correlated with soil EC. The absolute values of Pearson correlation coefficients between spectral and image features were mostly in the range of 0.01 to 0.38, indicating a low correlation level, which suggested strong independence and complementarity between the two data sources in characterizing soil EC. Among models using a single data source, the CNN model based on characteristic wavelengths achieved a test set R² of 0.895 7, slightly higher than the full-spectrum model (R²=0.892 8); the GRU model based on characteristic wavelengths achieved a test set R² of 0.877 0. In contrast, models based on image features performed lower, with the optimal CNN model achieving a test set R² of 0.666 3. All fusion models achieved test set R² values exceeding 0.9, significantly outperforming single-spectrum or single-image models. Comparing different fusion strategies, the feature-level fusion models generally demonstrated superior performance compared to data-level fusion models. Among these, the CNN feature-level fusion model performed best, with a test set R² of 0.924 9 and an RMSE of 73.908 8 μS/cm; the GRU feature-level fusion model was next, with an R² of 0.921 4 and an RMSE of 75.616 7 μS/cm; among the data-level fusion models, the optimal was the CNN data-level fusion model, with an R² of 0.920 2 and an RMSE of 76.149 9 μS/cm. [Conclusions] Extracting characteristic wavelengths and image features strongly correlated with soil EC through feature selection can effectively enhance model efficiency and generalization capability. The multi-source information fusion model based on spectral-image features significantly improved prediction accuracy compared to single-data-source models, providing a new method for the rapid detection of key parameters in saline-alkali land. While this study achieved promising results under controlled indoor conditions, the models still face challenges in field applications, including variations in illumination, interference from soil moisture, and the limited representativeness of surface measurements for characterizing salt content in the plow layer. Future work should incorporate moisture correction strategies, image color correction, and data augmentation methods to further enhance the models' robustness and applicability in real-world environments.

Key words: sodic saline-alkali soil, electrical conductivity, characteristic wavelength, image feature, feature screening, spectrum-image fusion