关键词: 无人机遥感, 土壤盐渍化, 深度, 混合模型, 随机森林, 克里金插值, 盐分预测

Abstract:

【Background】Soil salinization severely constrains crop growth and ecological balance, and its accurate monitoring is essential for saline-alkali land reclamation, yield forecasting, and precision farmland management. Driven by natural and anthropogenic factors, salinization is governed by the redistribution of water and salt within the soil profile, exhibiting pronounced vertical migration and strong spatial heterogeneity. Although unmanned aerial vehicle (UAV) remote sensing is now widely used for field-scale salinity mapping, it essentially captures surface information and fails to characterize salt gradients in deeper layers. 【Objective】To develop a UAV-image-based, layer-specific modeling framework that integrates machine learning with Kriging interpolation for high-resolution 3-D mapping of subsurface soil salinity.【Method】Firstly, the UAV was equipped with multispectral sensors to obtain high-resolution images of the test field, and the soil salinity data at different depths were measured synchronously, supplemented by real-time dynamic differential positioning technology to ensure spatial accuracy. Then, a spectral feature set including the red-edge band was constructed, and the feature optimization was carried out based on the random forest algorithm. On this basis, machine learning and Kriging interpolation method were combined to establish a stratified soil salinity prediction model and generate a high-resolution salinity distribution map. Finally, the advantages of the proposed method in the spatial representation of deep salinization were verified by comparing it with the cubic fitting depth function prediction method.【Result】The prediction accuracy R² of each depth of deep soil salinization spatial prediction by the mixed model hierarchical modeling was 0.68 (0-10 cm), 0.51 (10-20 cm), 0.58 (20-40 cm), 0.56 (40-60 cm) and 0.52 (60-80 cm), respectively, and the prediction effect of 0-10 cm surface layer was the best. The red-edged salinity index was an important predictor at all depths, which verified the applicability and effectiveness of the constructed red-edged index. By comparing the prediction results of the mixed model with the cubic fitting depth function, the spatial prediction accuracy of deep soil salinization in the layered model of the mixed model was higher, and it could more truly reflect the salinization degree at different depths in the experimental area.【Conclusion】UAV remote sensing technology is the best in shallow (0-10 cm) soil salinity prediction, and the prediction accuracy of soil properties decreases with the increase of depth, and the depth accuracy still needs to be improved. From the prediction results, the average soil salinity gradually increases with the increase of depth, indicating that there is an accumulation phenomenon of salt in the soil profile. Compared with the cubic fitting depth function method, the hybrid model based on random forest stratification modeling and Kriging residual correction shows higher spatial prediction accuracy in each soil layer, which is more reasonable and practical, and provides a scientific basis for dynamic monitoring of regional soil salinization and accurate layered soil salinity mapping.

Key words: UAV remote sensing, soil salinization, depth, hybrid model, random forest, Kriging interpolation, salinity prediction