| 谢鹏程,刘小刚,崔宁博,邢立文.基于改进长短时记忆深度学习模型的黄土高原苹果蒸腾量估算[J].干旱地区农业研究,2025,(6):259~270 |
| 基于改进长短时记忆深度学习模型的黄土高原苹果蒸腾量估算 |
| Estimation of apple transpiration in the Loess Plateau based on an improved long short\|term memory deep learning model |
| |
| DOI：10.7606/j.issn.1000-7601.2025.06.25 |
| 中文关键词: 苹果树 蒸腾估算 深度学习算法 分位数回归 通径分析 黄土高原 |
| 英文关键词:apple tree, transpiration estimation deep learning algorithm quantile regression path analysis Loess Plateau |
| 基金项目:国家自然科学基金项目（52309057，51922072，51779161，51779161）；“十三五”国家重点研发计划项目（51009101）；“十四五”国家重点研发计划项目（2022YFD1900805）；四川省科技计划项目（2024YFHZ0217） |
|
| 摘要点击次数: 170 |
| 全文下载次数: 53 |
| 中文摘要: |
| 苹果树的蒸腾量是影响其生长、产量、果实质量和水分管理的关键因素，同时对气候适应性、生态平衡和可持续农业具有重要作用。然而苹果树蒸腾量的精确估算在实际应用中面临许多挑战，尤其是在黄土高原地区，复杂的气候条件和土壤特性使得蒸腾量的预测变得更加困难。针对黄土高原地区复杂气候与土壤特性下苹果树蒸腾量估算精度低的问题，本研究提出一种结合分位数回归、时间卷积网络与双向长短期记忆网络（QRTCN-BiLSTM）的深度学习模型，并与传统经验模型（MLR、MJS、S-W）及经典深度学习模型（LSTM、Bi-LSTM）进行对比。基于2020—2022年山西省晋中市太谷县苹果园的实测数据（气象、土壤含水量、叶面积指数等），通过通径分析量化蒸腾驱动因子的动态影响，发现冠层部分阶段叶面积指数（LAI）和水汽压差（VPD）为主导因素，而茂密冠层阶段VPD影响最为显著。QRTCN-BiLSTM模型在各生长阶段的估算精度显著优于传统模型，全输入条件下决定系数R2=0.96，相对均方根误差RRMSE降低61%以上，且对输入变量缺失具有较强鲁棒性。该模型通过融合时序卷积与分位数回归，有效捕捉了环境因子的非线性关系及历史依赖特征，为黄土高原旱作苹果园水分精准管理及区域水资源高效利用提供了可靠技术支撑。 |
| 英文摘要: |
| The transpiration of apple trees is a key factor affecting their growth, yield, fruit quality, and water management, and it also plays an important role in climate adaptability, ecological balance, and sustainable agriculture. However, the precise estimation of apple tree transpiration faces many challenges in practical applications, especially in the Loess Plateau of China, where complex climate conditions and soil characteristics make transpiration prediction more difficult. To address the issue of insufficient estimation accuracy of apple tree transpiration under the complex climate and soil conditions of the Loess Plateau, this study proposed a deep learning model combining Quantile Regression, Temporal Convolutional Network, and Bidirectional Long Short\|Term Memory network (QRTCN-BiLSTM), and compared it with traditional empirical models (MLR, MJS, S-W) and classical deep learning models (LSTM, Bi-LSTM). Based on the measured data (meteorological, soil moisture, leaf area index, etc.) from apple orchards in Taigu County, Jinzhong City, Shanxi Province, during 2020-2022, path analysis was used to quantify the dynamic effects of transpiration driving factors. It was found that leaf area index (LAI) and vapor pressure deficit (VPD) at certain stages of the canopy were the dominant factors, with VPD having the most significant impact during the dense canopy stage. The results showed that the QRTCN-BiLSTM model significantly outperforms traditional models in terms of estimation accuracy at all growth stages, with a coefficient of determination (R2) of 0.96 under full input conditions, and a relative root mean square error (RRMSE) reduced by more than 61%. Moreover, the model demonstrated strong robustness to missing input variables. By integrating temporal convolution with quantile regression, the model effectively captured the nonlinear relationships of environmental factors and historical dependence characteristics, providing reliable technical support for precise water management and efficient regional water resource utilization in dryland apple orchards of the Loess Plateau. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| | |