STL-BiLSTM模型在多气象因子水稻NDVI预测中的应用

Application of the STL-BiLSTM Model in Predicting Rice NDVI Based on Multiple Meteorological Factors

摘要: 归一化差异植被指数(NDVI)作为表征植被生长状况的核心指标，是水稻产量动态预测与农业精准决策的关键支撑。当前水稻NDVI预测存在气候因子耦合深度不足、模型时序特征捕捉能力有限、区域适配性不强等问题，传统模型或依赖有限气象变量，或忽视NDVI时序结构与多因子协同效应，难以满足高精度预测需求。为解决上述问题，本研究以黑龙江省哈尔滨、齐齐哈尔、鸡西、佳木斯、绥化5个水稻主产区为研究区，基于1983~2022年逐日NDVI数据与气象数据，提出一种融合STL时序分解与双向长短期记忆网络(BiLSTM)的NDVI预测模型。研究首先通过STL方法将NDVI序列解构为趋势、季节及残差分量，厘清各分量与气象因子的多尺度关联特征；随后构建“STL分解-BiLSTM预测”一体化架构，融合多尺度气候指标与NDVI分层特征，形成多元输入的预测模型。预测结果表明，STL-BiLSTM模型预测精度优异，五市NDVI预测值与真实值的决定系数(R²)均≥0.994，显著优于对比模型；气象因子与NDVI分量存在尺度差异化关联，气温主要驱动季节分量波动，降水量主导趋势分量演化，验证了多尺度耦合机制的合理性。本研究的创新点在于建立了气候因子与STL分层特征的精准耦合机制、构建了时序分解与双向时序建模的一体化架构、优化了多元气象数据的输入模式，为水稻NDVI高精度预测提供了新范式，也为基于NDVI的水稻产量预测与粮食安全保障提供了参考价值。

Abstract: The Normalized Difference Vegetation Index (NDVI), a core indicator of vegetation growth, is crucial for dynamic rice yield prediction and precision agricultural decision-making. However, current rice NDVI prediction faces challenges: insufficient climate factor coupling, limited model capability in capturing temporal features, and weak regional adaptability, as traditional models either depend on limited meteorological variables or overlook NDVI’s temporal structure and multi-factor synergies, failing to meet high-precision demands. To solve these problems, this study developed an NDVI prediction model integrating STL time series decomposition and Bidirectional Long Short-Term Memory (BiLSTM) network, focusing on five major rice-producing cities in Heilongjiang Province (Harbin, Qiqihar, Jixi, Jiamusi, Suihua) and using daily NDVI and meteorological data from 1983 to 2022. The STL method first decomposed NDVI sequences into trend, seasonal, and residual components to clarify their multi-scale correlations with meteorological factors; an integrated “STL decomposition-BiLSTM prediction” framework was then constructed to fuse multi-scale climate indicators with NDVI layered features, forming a multi-variable input model. Results showed that the STL-BiLSTM model achieved superior accuracy, with the coefficient of determination (R²) between predicted and observed NDVI values ≥ 0.994 in all five cities, significantly outperforming comparison models. Moreover, meteorological factors exhibited scale-differentiated correlations with NDVI components: temperature primarily drove seasonal component fluctuations, while precipitation dominated trend component evolution, verifying the rationality of the multi-scale coupling mechanism. This study innovates by establishing a precise coupling mechanism between climate factors and STL layered features, building an integrated framework for time series decomposition and bidirectional temporal modeling, and optimizing multi-source meteorological data input modes. It provides a new paradigm for high-precision rice NDVI prediction and valuable references for NDVI-based rice yield prediction and food security guarantee.