SDIDM: An agricultural irrigation decision model based on Chain-of-Thought Prompts enhancement

Existing Chain-of-Thought (CoT) prompting methods often lack systematic modeling of reasoning depth when addressing multi-step complex inference tasks, resulting in limited ability to correct error accumulation during the reasoning process. Moreover, current approaches suffer from limited semantic guidance and weak domain adaptation mechanisms, which constrain the logical consistency and adaptability of large language models (LLMs) in complex decision-making scenarios. To address these limitations in high-complexity tasks such as agricultural irrigation decision-making, we propose CoTIrr, a CoT-based intelligent decision framework that introduces a three-dimensional evolution strategy: (1) Reasoning Task Deepening:we formulate ordinary / partial di ff erential equations (ODE / PDE) to model the temporal and spatial dynamics of multiple variables (e.g., soil moisture, evaporation rate, and crop water demand), enhancing the depth and structural stability of reasoning chains;(2) Semantic Diversification Guidance: we re-construct the background and key factors of tasks in diversified ways to enrich the semantic structure and improve task representation;(3) Domain Adaptation Optimization: we refine and expand reasoning steps to achieve fine-grained enhancement of the CoT process, improving reasoning accuracy and overall alignment with domain tasks. To ensure logical validity, we introduce a dual-filtering mechanism consisting of evolution e ff ectiveness evaluation and accuracy verification, and construct the first CoT prompt generation-validation framework tailored for agricultural decision-making, resulting in a dataset of 70,000 high-quality reasoning chains. Based on this dataset, we train the Structure-Driven Irrigation Decision Model (SDIDM) using GLM-4 9B.To mitigate error accumulation in reasoning, SDIDM innovatively incorporates a progressive argumentation mechanism that enables multi-role collaborative correction of intermediate steps, thereby improving reasoning consistency and stability in complex scenarios.Experiments conducted on heterogeneous data sources (meteorological, soil moisture, crop parameters, historical irrigation records) from two demonstration zones—Qitai and Manas—show that SDIDM outperforms mainstream methods by 3.27% in accuracy and 4.18% in reasoning consistency, demonstrating the e ff ectiveness of the proposed approach.