| 引用本文: | 陆丽芳,刘春伟,张宝忠,等.基于机器学习的南京地区冬小麦蒸散发估算研究[J].灌溉排水学报,2025,44(8):65-76. |
| LU Lifang,LIU Chunwei,ZHANG Baozhong,et al.基于机器学习的南京地区冬小麦蒸散发估算研究[J].灌溉排水学报,2025,44(8):65-76. |
|
| 摘要: |
| 【目的】降低南京地区冬小麦蒸散发（ETc）估算的不确定性并提高农业水资源管理水平。【方法】采用机器学习方法优化冬小麦ETc的估算，并与大型蒸渗仪实测数据进行对比分析。研究基于2011—2018年南京地区大型称重式蒸渗仪实测冬小麦ETc及相关气象数据，利用贝叶斯优化（Bayesian Optimization, BO）算法优化Lasso回归、自适应增强（AdaBoost）、随机森林（RF）和梯度提升决策树（GBDT）4种机器学习模型的参数。通过决定系数（R2）、平均绝对误差（MAE）、均方根误差（RMSE）和SHAP可解释性分析等评估各个模型在ETc估算中的精度及ETc估算的关键决定因素。【结果】GBDT模型和RF模型的估算效果最佳，R2分别为0.951和0.926，GBDT模型的MAE和RMSE分别为0.370 mm/d和0.541 mm/d，RF的MAE和RMSE分别为0.506 mm/d和0.694 mm/d。SHAP可解释性分析显示，叶面积指数（LAI）、太阳辐射（Rs）和净辐射（Rn）是ETc估算中关键因素。此外，基于Logistic模型的LAI估算方法，能够通过生长天数和有效积温准确估算LAI，从而实现对ETc的间接估算。【结论】通过优化机器学习模型参数，显著提高了南京地区冬小麦ETc的估算精度。GBDT模型和RF模型在ETc估算中表现最佳，能够有效预测冬小麦的ETc。基于机器学习的ETc估算，不仅减少了对高成本观测设备的依赖，还可为农业水资源管理提供科学支持。 |
| 关键词: 冬小麦；蒸散发；模型；机器学习；可解释性；贝叶斯优化算法 |
| DOI：10.13522/j.cnki.ggps.2024409 |
| 分类号: |
| 基金项目: |
|
| Machine learning-based methods for estimating evapotranspiration of winter wheat field in the regions of Nanjing |
|
LU Lifang, LIU Chunwei, ZHANG Baozhong, WANG Ranghui, ZHANG Pei, QIU Rangjian
|
|
1. Jiangsu Key Laboratory of Agricultural and Ecological Meteorology, Nanjing University of Information Science and Technology; Open Laboratory of Ecosystem Carbon Source and Sink, China Meteorological Administration, Nanjing 210044, China;
2. State Key Laboratory of Basin Water Cycle Simulation and Regulation, China Institute of Water Resources and
Hydropower Research, Beijing 100038, China; 3. Jiangsu Climate Center, Nanjing 210008, China;
4. State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan 430072, China
|
| Abstract: |
| 【Objective】Accurate estimation of crop evapotranspiration (ETc) is essential for improving irrigation and water resource management. This study aimed to reduce the uncertainty in ETc estimation for winter wheat in the region of Nanjing and improve the efficiency of agricultural water use. 【Method】 Four machine learning models: the Lasso Regression, the Adaptive Boosting (AdaBoost), the Random Forest (RF) and Gradient Boosting Decision Tree (GBDT) were used to calculate ETc using meteorological data measured from 2011 to 2018. Bayesian optimization (BO) was used to fine-tune the model parameters. The calculated results were compared with data measured from lysimeters. Shapley Additive Explanations (SHAP) analysis was used to identify key meteorological factors. A logistic growth model for leaf area index (LAI) based on degree-day and effective accumulated temperature was developed to indirectly estimate ETc. 【Result】 ① The GBDT and RF models outperformed other model, with their R2 values being 0.951 and 0.926, respectively. The GBDT model was most accurate, with a MAE of 0.370 mm/d and RMSE of 0.541 mm/d. ② SHAP analysis showed that LAI, solar radiation (Rs) and net radiation (Rn) were the most influential factors in ETc estimation. ③The logistic-based LAI model enables indirect prediction of ETc using temperature-driven metrics.【Conclusion】The integration of machine learning with Bayesian optimization significantly improves ETc estimation for winter wheat. The GBDT and RF models offer robust alternatives to traditional methods, reducing dependence on costly instrumentation and dense meteorological networks. This scalable method supports data-driven irrigation scheduling and improves sustainable agricultural water management in winter wheat production in the study region. |
| Key words: winter wheat; evaporation; model machine learning; explainability; Bayesian optimization algorithm |
