Modeling sunflower yield and soil water–salt dynamics with combined fertilizers and irrigation in saline soils using APSIM and deep learning

• Published in: Environmental sciences Europe Vol. 37; no. 1; pp. 104 - 19
• Main Authors: Miao, Qingfeng, Yu, Dandan, Shi, Haibin, Feng, Zhuangzhuang, Feng, Weiying, Li, Zhen, Gonçalves, José Manuel, Duarte, Isabel Maria, Li, Yuxin
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 20.06.2025; Springer Nature B.V; SpringerOpen
• Subjects: Accumulation; Agricultural land; Agricultural production; Agrochemicals; Agronomic crops; Algorithm; Algorithms; Animal manures; Artificial neural networks; Carbon; Computer simulation; Crop growth; Crop model; Crop yield; Deep learning; Design of experiments; Earth and Environmental Science; Ecotoxicology; Environment; Environmental protection; Fertilizer application; Fertilizers; Field tests; Growth models; Helianthus; Irrigation; Irrigation water; Leaf area; Leaf area index; Learning algorithms; Machine learning; Manures; Mineral fertilizers; Moisture content; Neural networks; Nitrates; Nitrogen; Optimization; Organic fertilizer; Organic fertilizers; Organic nitrogen; Osmotic stress; Phosphorus; Photosynthesis; Physiology; Plant stress; Pollution; Productivity; Ratios; Regression analysis; Resource utilization; Saline soils; Salinity; Salinity effects; Salt; Salts; Soil dynamics; Soil moisture; Soil salinity; Soil water; Substitutes; Sunflowers; Supplemental irrigation; Support vector machines; Water; Water shortages; China; Crop growth; Organic fertilizer; Crop model; Algorithm
• ISSN: 2190-4715; 2190-4715
• DOI: 10.1186/s12302-025-01145-2

Understanding the interactions between crop growth and abiotic stressors (water, salt, and nitrogen) is crucial for optimizing fertilizer use, improving plant stress resistance, and promoting agricultural productivity and environmental sustainability. Herein, we investigated the effects of organic fertilizer type, organic fertilizer ratio, and supplemental irrigation on soil water and salt transport, crop growth, and yield in mildly to moderately salinized soils. Using the APSIM model, we simulated crop growth and soil moisture under different organic fertilizer application ratios in mildly to moderately saline soils. Based on sunflower field experiments, four machine learning models (regression trees, random forest, support vector machines, and XGBoost) and two deep learning models (deep neural networks and neural networks) were developed to predict soil salinity. Results showed that reducing nitrogen application and using organic fertilizers decreased soil salinity by 11.1–22.8% at a 0–60 cm depth. A 50% organic to inorganic fertilizer ratio minimized salt accumulation. In mildly salinized soils, supplemental irrigation increased leaf area index (LAI) and biomass by 1.8–7.1% and 9–35%, respectively. Moreover, in mildly salinized farmlands, the combination of 75% organic fertilizer and 44 mm of supplemental irrigation resulted in relatively lower soil salinity. In moderately salinized farmland, lower soil salinity accumulation was observed with 25% organic fertilizer and 44 mm supplemental irrigation. In mildly saline–alkali soils, maximum yield was achieved with 50% organic nitrogen substitution + 22 mm supplemental irrigation. In moderately saline–alkali soils, the same substitution rate (50%) yielded peak production but required 44 mm irrigation to counteract osmotic stress. Compared to natural farm manure, commercial organic fertilizer with supplemental irrigation increased crop yield, agronomic efficiency (Ac), and harvest index (Hi). The maximum crop yield and yield components were achieved with 50% organic fertilizer and 22 mm supplemental irrigation. In the moderately salinized soil, the highest irrigation productivity was achieved with 75% organic fertilizer. Although the APSIM-sunflower model can be used to simulate growth and development ( R 2 = 0.7–0.9; NRMSE = 0.1–0.2), its simulation of soil water dynamics is unsatisfactory ( R 2 = 0.4–0.5; NRMSE = 0.3). In simulating soil salinity, deep learning models generally outperform machine learning models (EVS ≤ 0.3; R 2 ≤ 0.42), with the deep neural network (DNN (EVS ≤ 0.3; R 2 ≤ 0.82)) algorithm demonstrating the best simulation performance. The adjustment of the organic–inorganic fertilizer ratio and supplemental irrigation strategies can optimize resource utilization in saline-alkali soils. DNN provides a more accurate method for predicting soil salinity, achieving a balance between productivity improvement and environmental protection in salt-affected areas.