多源遥感数据尺度转换的夏玉米蒸散发融合模型研究

Fusion Model of Evapotranspiration of Summer Maize for Scale Conversion of Multi-source Remote Sensing Data

摘要:

蒸散发（Evapotranspiration, ET）是作物需水量的核心组分，也是区域水资源优化配置的关键依据。本文以陕西关中宝鸡峡灌区夏玉米为研究对象，采用BP神经网络（Back propagation neural network，BPNN）、支持向量机（Support vector machine, SVM）、极限学习机（Extreme learning machine, ELM）和极致梯度提升树（eXtreme gradient boosting, XGBoost）4种机器学习算法构建无人机-卫星多源遥感数据协同校正模型，并以最优算法建立的模型校正卫星多光谱数据，实现无人机和卫星数据的尺度转换。利用校正后高精度卫星数据反演夏玉米叶面积指数（Leaf area index, LAI）与株高（Crop height, hc）为蒸散发模型提供数据输入。分别采用双作物系数法、METRIC模型及Penman-Monteith（P-M）冠层阻力模型进行夏玉米蒸散发估算，引入贝叶斯模型平均（Bayesian model averaging, BMA）实现不同生育阶段各方法/模型权重的动态分配，最终得到玉米拔节-完熟期性能稳健的蒸散发BMA融合模型。结果表明：XGBoost算法在夏玉米拔节-完熟期的B/G/R/NIR波段建模精度均为最高，四波段建模结果决定系数（Coefficient of determination, R2）较算法ELM高出8.43%、8.67%、6.79%和10.41%；校正后的卫星多光谱数据LAI与hc反演结果R2较原始卫星数据分别平均提高97%和67.5%；BMA融合模型在夏玉米拔节-抽雄期和蜡熟-完熟期较单一最优方法/模型（METRIC模型）均方根误差（Root mean squared error, RMSE）降低39.3%~58.5%。本研究利用“协同校正-动态融合”显著提升了蒸散发遥感监测精度，可为水资源精细化管理提供理论支撑。

Abstract:

Evapotranspiration (ET) is a core element of crop water requirement and serves as a key basis for optimizing regional water resource allocation. Focusing on summer maize in the Baojixia Irrigation District located in the Guanzhong Plain of Shaanxi Province, four machine learning algorithms were employed, back propagation neural network (BPNN), support vector machine (SVM), extreme learning machine (ELM), and eXtreme gradient boosting (XGBoost), to develop a collaborative correction model utilizing multisource remote sensing data from unmanned aerial vehicles (UAVs) and satellites. Subsequently, the model constructed by the optimal algorithm was selected to correct satellite multispectral data, ultimately achieving scale conversion between UAV and satellite data. The calibrated high-precision satellite data were utilized to retrieve the leaf area index (LAI) and crop height (hc) of summer maize, providing essential data inputs for the evapotranspiration (ET) model. Evapotranspiration (ET) of summer maize was estimated using three distinct approaches: the dual crop coefficient method, the mapping evapotranspiration at high resolution with internalized calibration (METRIC) model, and the Penman-Monteith (P-M) canopy resistance model. Subsequently, the Bayesian model averaging (BMA) method was introduced to dynamically assign weights to each method/model across different growth stages. Ultimately, this process led to the development of a robust BMA-merged ET model for the maize growing period spanning from the jointing stage to physiological maturity. The results demonstrated that the XGBoost algorithm consistently achieved the highest modeling accuracy for B/G/R/NIR bands during the jointing to maturity stages of summer maize, with R2 values in the four-band modeling outperforming the suboptimal ELM algorithm by 8.43%, 8.67%, 6.79%, and 10.41% respectively. The retrieval of LAI and hc using the calibrated satellite multispectral data exhibited an average improvement in R2 of 97% and 67.5%, respectively, compared with retrievals based on the original satellite data. Compared with the single best-performing method/model (the METRIC model), the BMA-merged model significantly reduced the root mean squared error (RMSE) by 39.3% to 58.5% during both the jointing-tasseling stage and the dough-physiological maturity stage of summer maize. The “collaborative calibration-dynamic fusion” framework proposed significantly improved the accuracy of remote sensing-based evapotranspiration (ET) monitoring, thereby providing theoretical support for precision water resource management.