벼-밀 작부체계에서의 기후 맞춤형 물 관리를 위한 데이터 기반 APSIM?AI 융합 프레임워크 개발 = Development of a Data-Driven APSIM-AI Fusion Framework for Climate-Resilient Water Management in Rice-Wheat Cropping Systems

다국어 초록 (Multilingual Abstract)

Effective water management is essential for improving winter wheat yield, grain quality, and climate resilience in rice-wheat double-cropping systems across South and East Asia. However, wheat production on paddy-derived soils is limited by spring drought, post-anthesis waterlogging, and the lack of region-specific irrigation guidelines. This study evaluated precision irrigation strategies integrating real-time soil-moisture monitoring to enhance water-use efficiency (WUE) and crop performance under variable hydroclimatic conditions. Over three seasons (2021-2024) in South Korea, conventional rainfed management (CRF) was compared with two sensor-guided strategies: soil moisture based irrigation triggered at 55% of available soil water (SIA) and a saturation-oriented strategy triggered at 55% of saturation water content (SIS). SIA consistently outperformed CRF and SIS, increasing grain yield by 20-27%, WUE by 10-22%, and leaf area index by up to 16%. Maintaining 0-40 cm soil moisture within the plant-available range from jointing to grain filling improved growth and resource use, whereas SIS promoted oversaturation stress and CRF incurred deficit stress.
Building on these observations, APSIM-Wheat was calibrated and validated for Korean rice-wheat rotations and used to diagnose management sensitivities under present and future climates. Soil-moisture performance was generally good, with larger errors in an anomalously wet season that amplified waterlogging impacts and boundary-condition sensitivity. Leaf area index was reproduced with moderate-to-high fidelity; flowering and maturity were typically predicted within ~1-3 days; simulated yields aligned with regional benchmarks. Grain-protein errors were within commonly accepted ranges but reflected concentration dilution effects in wet years and carry-over of soil water and mineral nitrogen. Future climate simulations projected yield increases under both low and high-forcing pathways due to elevated CO₂ and warmer temperatures that advanced phenology, alongside protein declines consistent with nitrogen dilution. Scenario analysis indicated that a rainfall-deficit irrigation rule reliably increased yield but tended to reduce grain protein, whereas phenology-anchored split nitrogen top-dressing approached an efficiency plateau near 120 kg/ha. Combining these management strategies, split nitrogen top-dressing (60:60 kg/ha at booting and five days before heading) and rainfall-triggered irrigation resulted in the most stable, simultaneous improvements in both yield and protein content.
To enable operational use, a data-driven APSIM-Machine Learning fusion framework was developed, coupling APSIM's mechanistic grounding with a compact, stage-resolved feature set and an ensemble of learners (Random Forest, XGBoost, Lasso, Decision Tree, Extra Trees, and Gradient Boosting). Mechanistically informed phenophase-centered features reduced dependence on complete high-frequency meteorological inputs while preserving interpretability. Across sites and management scenarios, the fusion framework achieved accurate multi-target predictions for yield (R² = 0.96; RMSE = 203 kg/ha), grain protein (R² = 0.71; RMSE = 0.4%), and phenology (days after sowing). The system integrates historical climate-path simulations with real-time observations and short-term forecast augmentation to support mid-season updating and scenario testing. To assess extrapolation performance, winter wheat yields for the 2026-2056 period were predicted using data from 18 GCMs under SSP1-2.6 and SSP5-8.5. Across these climate realizations, the mean RMSE was 340 kg/ha, and model skill under SSP1-2.6 reached R² = 0.91, indicating that the framework retains appreciable predictive accuracy when extrapolated beyond the historical training conditions.
Overall, this study integrates sensor-triggered field irrigation, validated process-based simulation, and scalable machine learning into a coherent pipeline for paddy-upland wheat systems. It provides field-based evidence that SIA is a robust, stage-aware irrigation strategy, confirms the suitability of APSIM-Wheat for decision support in Korean rice-wheat rotations, and delivers an AI-application APSIM-ML fusion framework that bridges mechanistic biophysical simulation with real-time operational demands. Collectively, these contributions outline a practical pathway to improve domestic wheat productivity and resource-use efficiency, while offering a transferable template for climate-resilient agriculture in data-scarce and forecast-uncertain environments.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1.1 Research background and motivation 1
• 1.2 Study objectives and content 5
• Chapter 2. Literature review 9
• 2.1 Importance and challenges of field water management for winter wheat 9
• 2.2 Application of crop models in agriculture 17
• 2.3 AI and machine learning applications in agricultural systems 24
• Chapter 3. Materials and methods 30
• 3.1 Experimental site and design 30
• 3.2 Measurement and sampling 35
• 3.3 Crop water and yield quantification 39
• 3.4 APSIM model description and structure 41
• 3.4.1 Model overview 41
• 3.4.2 Model structure 42
• 3.5 Integration of APSIM and machine learning 52
• 3.5.1 APSIM and machine learning fusion framework 52
• 3.5.2 Feature engineering and input variable processing 54
• 3.5.3 Dataset construction and prediction models 55
• 3.6 Model performance evaluation 57
• 3.7 Model-based scenario analysis 60
• 3.7.1 Irrigation and fertilization scenarios 60
• 3.7.2 Future climate scenarios 62
• 3.7.3 Expanded scenario applications 66
• Chapter 4. Improving winter wheat yield and water use efficiency using soil moisture sensor-driven precision furrow irrigation 68
• 4.1 Monitoring of winter wheat soil water status, growth, and productivity 68
• 4.1.1 Soil moisture dynamics 68
• 4.1.2 Leaf area index 73
• 4.1.3 Crop water indices 74
• 4.1.4 Biomass and grain development 76
• 4.1.5 Yield, protein, and water use performance indicators 79
• 4.2 Discussion 83
• 4.2.1 Precision irrigation and soil moisture regulation 83
• 4.2.2 Physiological responses and canopy functionality 84
• 4.2.3 Biomass and yield 85
• 4.2.4 Water use efficiency and agronomic performance 86
• 4.2.5 Irrigation implications in paddy soils 87
• 4.3 Conclusion 89
• Chapter 5. Assessing future climate change impacts on winter wheat yield and protein content using the APSIM model 90
• 5.1 Model performance evaluation 90
• 5.1.1 Crop genetic parameters 90
• 5.1.2 Soil moisture content 91
• 5.1.3 Leaf area index and crop phenology 96
• 5.1.4 Yield and protein content 98
• 5.2 Evaluation of scenario impacts 100
• 5.2.1 Irrigation scenario impacts under future climate 100
• 5.2.2 Fertilization scenario impacts under future climate 102
• 5.2.3 Combined irrigation and fertilization scenario impacts under future climate 104
• 5.3 Discussion 106
• 5.3.1 Simulated soil moisture dynamics and responses 106
• 5.3.2 Canopy development and phenology 107
• 5.3.3 Grain yield-protein content responses and trade-offs 107
• 5.3.4 Model performance and applicability 109
• 5.3.5 Implications of future climate scenarios 110
• 5.3.6 Implications for irrigation and fertilization management strategies 111
• 5.4 Conclusion 114
• Chapter 6. Development of an APSIM-Machine Learning fusion framework for winter wheat water management 116
• 6.1 Model performance evaluation 116
• 6.1.1 Internal training performance 116
• 6.1.2 Target prediction performance 117
• 6.1.3 Importance of climate features 120
• 6.2 Evaluation of scenario impacts 122
• 6.2.1 Irrigation scenario analysis 122
• 6.2.2 Fertilization scenario analysis 123
• 6.2.3 Initial moisture scenario analysis 124
• 6.2.4 Sowing time scenario analysis 124
• 6.2.5 Future climate scenario analysis 125
• 6.3 Application in real-time decision making 126
• 6.4 Discussion 130
• 6.4.1 Interpretation of model performance 130
• 6.4.2 Interpretation of climate feature importance 132
• 6.4.3 Insights from scenario analysis 133
• 6.4.4 Assessment of model extrapolation and generalizability 135
• 6.5 Conclusion 138
• Chapter 7. Comprehensive conclusion 139
• References 142