基于Transformer‑BiLSTM混合模型的冬小麦估产与特征贡献度分析

Winter Wheat Yield Estimation and Feature Contribution Degree Analysis Based on Transformer‑BiLSTM Hybrid Model

摘要:

在全球气候变化持续加剧、粮食安全挑战日益严峻的背景下，精准、及时地估算作物产量至关重要。传统基于反射率光谱的植被指数难以实时捕捉作物光合生理状态，且单一Transformer和双向长短期记忆网络（Bi?directional long short?term memory，BiLSTM）模型在提取与产量相关的时序特征方面亦有一定局限性。为此，本研究提出一种日光诱导叶绿素荧光（Solar?induced chlorophyll fluorescence，SIF）以及实际蒸散发（Actual evapotranspiration，Aet）、降水量（Precipitation，Ppt）和帕尔默干旱指数（Palmer drought severity index，PDSI）等数据融合的混合深度学习估产模型，通过结合Transformer提取全局依赖关系的优势以及BiLSTM捕捉局部细节变化方面的优势，构建了Transformer?BiLSTM冬小麦估产混合模型，并评估了模型泛化能力及特征贡献度。结果表明，在河南省2013—2019年县级尺度样本测试集数据中，Transformer?BiLSTM混合模型拟合性能优越（决定系数R2为0.89，归一化均方根误差（NRMSE）为8.18%，相对预测偏差（RPD）为2.90），各项指标均优于单一Transformer和BiLSTM模型（R2均提高0.04，NRMSE分别降低1.46、1.22个百分点，RPD由2.46、2.53提升至2.90）。2020—2022年河南省县级数据跨时间试验中，Transformer?BiLSTM混合模型仍保持较高精度（R2为0.89，NRMSE为8.44%，RPD为2.77），R2较单一Transformer和BiLSTM模型分别提高0.05和0.07，NRMSE分别降低1.92、1.56个百分点，RPD则由2.25、2.33提升至2.77，体现了该模型良好的时间泛化能力。进一步将混合模型应用至产量分布更为复杂的安徽省，模型表现依然稳健（R2为0.87，NRMSE为11.07%，RPD为2.73），R2分别提高0.07和0.08，NRMSE分别降低2.33、3.10个百分点，RPD由2.25、2.13提升至2.73，验证了Transformer?BiLSTM混合模型具有较强的区域泛化能力。此外，将模型在像元尺度生成的高分辨率产量分布图汇总至县级后进行验证，结果亦表明其与统计产量具有较高的一致性（R2>0.8）。基于SHAP（Shapley additive explanations）特征重要性分析结果表明，1—2月最低温和3—6月SIF对模型输出贡献最高，其中SIF在整个时序中始终保持较高重要性。同时，冬小麦拔节至灌浆期PDSI、Ppt和Aet等气象因子对产量预测亦具有显著影响，反映模型能有效捕捉作物生长过程与环境因子间的协同作用。

Abstract:

Against the backdrop of intensifying global climate change and escalating food security challenges, accurately and timely estimating crop yields is important. Traditional vegetation indices based on reflectance spectra are difficult to capture the photosynthetic physiological state of crops in real time, while single?model approaches like Transformers and bi?directional long short?term memory network (BiLSTM) also exhibit limitations in extracting yield?related temporal features. Therefore, a hybrid deep learning yield estimation model that integrated data such as solar?induced chlorophyll fluorescence (SIF), actual evapotranspiration (Aet), precipitation (Ppt), and Palmer drought severity index (PDSI) was proposed. By leveraging the advantages of Transformer in extracting global dependencies and the BiLSTM in capturing local detail changes, a Transformer?BiLSTM wheat yield estimation model was constructed. The generalization ability and feature contribution of the model were also evaluated. Results indicated that the Transformer?BiLSTM hybrid model demonstrated superior fitting performance on the 2013—2019 county?level sample test dataset from Henan Province (R2=0.89, NRMSE was 8.18%, RPD was 2.90). All metrics outperformed those of both the single Transformer and BiLSTM models (R2 was increased by 0.04, NRMSE was decreased by 1.46 and 1.22 percentage points respectively, and RPD was improved from 2.46 and 2.53 to 2.90). In the 2020—2022 cross?temporal experiment of county?level data in Henan Province, the Transformer?BiLSTM hybrid model maintained high accuracy (R2=0.89, NRMSE was 8.44%, RPD was 2.77). Compared with single Transformer and BiLSTM models, R2 was improved by 0.05 and 0.07, respectively, while NRMSE was decreased by 1.92 and 1.56 percentage points. The RPD was risen from 2.25 and 2.33 to 2.77, demonstrating the model's robust temporal generalization capability. Further application of this hybrid model to Anhui Province, where yield distributions were more complex, exhibited robust performance (R2=0.87, NRMSE was 11.07%, RPD was 2.73). The R2 values were increased by 0.07 and 0.08, respectively, while the NRMSE was decreased by 2.33 and 3.10 percentage points. The RPD was improved from 2.25 and 2.13 to 2.73, confirming the strong regional generalization capability of the Transformer?BiLSTM hybrid model. Furthermore, aggregating the high?resolution yield distribution maps generated at the pixel scale to the county level for validation demonstrated high consistency with statistical yields (R2>0.8). Based on Shapley additive explanations (SHAP) feature importance analysis, the minimum temperatures from January to February and the SIF from March to June contributed most significantly to the model outputs, with SIF maintaining consistently high importance throughout the entire time series. Concurrently, meteorological factors such as PDSI, Ppt, and Aet during the winter wheat jointing to grain filling stage also exerted significant influence on yield prediction, indicating the model's ability to effectively capture synergistic interactions between crop growth processes and environmental factors.