Intelligent classification of variable rate fertilization for soda salt-alkali soil based on multi-source data fusion

To achieve precise management of soda saline-alkali soil in northeast China and variable rate application of organic fertilizers, a multi-source data fusion deep learning method hybrid model driven by fertilizer recommendation index（FRI）has been proposed．A FRI-CNN-MLP hybrid model was constructed by integrating 2023 growing season Sentinel-2 time-series remote sensing data, based normalized difference vegetation index（NDVI）and soil-adjusted vegetation index（SAVI）with soil cation exchange capacity（CEC）and pH data from SoilGrids database．This model has achieved four-level classification of fertilization application intensity within plastic-mulched areas of saline-alkali soil．FRI labels for model training were generated based on expert rules derived from current technical standards．Prediction results demonstrated that light fertilizer application index（accounting for 70.53%）was prodominant in study area, while moderate and heavy application were concentrated in low-lying peripheral areas, exhibiting a heavier application at periphery and lighter application at center．Compared with baseline method, model achieved an overall accuracy rate of 89.7%, and macro average F1 score of 0.85, outperforming traditional threshold method and random forest model．FRI-CNN-MLP hybrid model effectively integrates spatiotemporal-soil features, overcoming strong subjectivity and poor generalization of traditional threshold methods．It can directly determine organic fertilizer application rates, providing technical support for intelligent saline-alkali soil improvement．