随着全球农产品需求量逐渐增加、人均耕地面积逐渐减小，粮食安全问题日益突出。准确掌握粮食产量信息既有助于指导农作物种植和粮食生产，也有助于及时调整粮食进出口方案，为全球粮食贸易提供决策信息。冬小麦既是世界范围内广泛种植的粮食作物，也是我国主要的粮食作物，它被广泛种植在我国北方地区，其中以华北平原最为典型。准确掌握我国华北平原的冬小麦产量可以为稳定我国粮食生产、及时调整农业种植结构等提供参考。

遥感技术凭借其无损、快速、准确且适用于大尺度范围的优势，已被广泛应用于农作物产量估算。然而，目前结合气象要素和植被指数的遥感统计估产方法在考虑气象要素时大多采用全生长季时间范围内的气象要素，并未将气象要素按物候阶段进行区别考虑。由于农作物在不同的生长阶段对水热条件的需求以及对气象要素的敏感性不同，直接采用全生长季时间范围内的气象要素可能会弱化或忽略某一阶段不利气象条件对农作物生长发育的影响。另一方面，在提取农作物的物候信息时，大多数研究采用已有的自然植被物候监测方法，未结合冬小麦自身的生长特点和影响因素来设计相应的提取方法，并且这些方法主要关注冬小麦的返青期，无法准确提取出与冬小麦产量密切相关的其它关键物候期。

为此，本文选择华北平原冬小麦种植区作为研究区，首先设计了一种结合遥感返青期和逐日温度数据的冬小麦关键物候期（本文特指与冬小麦产量密切相关的拔节期、抽穗期和乳熟期）提取方法；然后将冬小麦的生长过程按关键物候期划分为多个物候阶段，并基于4种常用的回归模型（多元线性回归、人工神经网络、支持向量回归、随机森林回归），从模型精度和时空迁移能力2个角度在县级尺度上对比评估了分物候阶段估产方案和全生长季估产方案；最后基于优选模型对华北平原主要区县的冬小麦单产进行了估算，并评估了优选模型的应用能力。获得的主要研究结论如下：

（1）从冬小麦生长发育所需热量条件的角度出发，设计了一种结合遥感返青期和逐日温度数据的冬小麦关键物候期提取方法（简称积温法），该方法不仅具有较高且较为稳定的监测精度（基于当年的训练数据），也具有较好的预测精度（基于往年的训练数据）。该方法首先根据华北平原各物候观测站点的逐日温度数据和物候观测数据，计算出冬小麦自返青期发育至其它关键物候期的最优积温指标及阈值；然后根据遥感植被指数时序数据，使用相对阈值法提取出冬小麦的遥感返青期，再结合遥感返青期和逐日温度数据，推算出冬小麦从遥感返青期开始首次达到其它某一关键物候期最优积温指标下积温阈值的日期，将其作为该关键物候期的提取结果。应用该方法提取了2016－2018年华北平原冬小麦拔节期、抽穗期和乳熟期的空间分布，并采用物候观测数据对监测和预测精度进行了定量评价。监测精度方面，拔节期和抽穗期提取结果的决定系数（R²）均高于0.55，均方根误差（RMSE）在4~6天之间，偏差（BIAS）在3天以内；相较之下，乳熟期的监测精度略微偏低，R²在0.35~0.50之间，RMSE在5~7天之间，BIAS在2天以内。预测精度方面，基于积温法预测冬小麦关键物候期的精度相较于同年的监测精度稍有下降（BIAS上升0.4~3.8天、RMSE上升0.2~2.5天），但依然维持在较高水平，且基于两年样本的预测精度相较于基于一年样本的预测精度更高（BIAS变化幅度小于0.5天、RMSE变化幅度小于0.2天）。

With the increasing demand for agricultural products and the reduction of per capita cropland around the world, the goal of food security is facing a serious threat. Accurate and timely information on grain production is helpful for guiding crop cultivation, adjusting grain import and export schemes, and providing support for global food trade. Winter wheat is a crop widely planted all over the world. In China, it is a major food crop, which is widely planted in the north, especially in the North China Plain (NCP). Accurate information on the yield of winter wheat in the NCP can provide guidance for stabilizing grain production and adjusting agricultural planting structure in China.

Remote sensing technology has become a main method for crop yield estimation by its advantages of non-destructive, fast and applicable to a large scale. However, when combining meteorological factors with vegetation indices to build a remote sensing statistical yield estimation model, most studies adopt the mean or accumulated value of meteorological factors within the whole growing season, instead of distinguishing meteorological factors according to phenological periods. Since crops have different requirements for hydrothermal conditions in different growth phases, using the mean or accumulated value of meteorological factors within the whole growing season may weaken/ignore the impacts of adverse meteorological conditions on crop growth in a certain phase. In addition, most of the existing studies use phenology monitoring methods for natural vegetation to extract crop phenological information. These methods do not consider the characteristics and influencing factors of winter wheat growth, mainly focus on the green-up date of winter wheat, and cannot accurately extract other key phenological dates which are closely related to winter wheat yield.

To solve these problems, the winter wheat planting area in the NCP was selected as the study area. Firstly, a method for monitoring the key phenological dates of winter wheat (i.e., jointing date, heading date, and milking maturity date) by combining satellite-derived green-up dates and daily temperature data was proposed. Then the developmental progression of winter wheat was distinguished as three phenophases according to the key phenological dates. Based on four commonly used regression methods, i.e., multiple linear regression (MLR), artificial neural network (ANN), support vector regression (SVR) and random forest (RF), the accuracy and spatiotemporal transferability of the phenological piecewise modelling was compared with that of the whole-season modelling at the county level. Finally, the optimal yield estimation model was selected to estimate the yield of winter wheat in the main counties of NCP, and the accuracy was assessed. The main conclusions of this study are as follows:

(1)   From the perspective of the thermal requirements for winter wheat growth, an accumulated temperature method (ATM) for monitoring the key phenological dates of winter wheat by combining satellite-derived green-up dates and daily temperature data was proposed. The monitoring accuracy (based on the training data of the current year) of the ATM method was high and stable, and the method also performed well in forecasting the key phenological dates (based on the training data of previous years). The ATM method consists of three key procedures. First, according to phenology observation samples and daily temperature data, selecting the optimal baseline temperature and calculating the accumulated temperature thresholds from the green-up date to the other key phenological dates. Then, extracting the satellite-derived green-up date of winter wheat with the dynamic threshold method from remotely sensed vegetation index time-series data. Finally, combined with the satellite-derived green-up date and daily temperature data, selecting the date when the accumulated temperature first reaches the corresponding threshold as the extraction result of the corresponding phenological date. Based on the method, the spatial distributions of jointing date, heading date and milking maturity date of winter wheat in the NCP from 2016 to 2018 were extracted, and the monitoring and forecasting accuracies were assessed using phenology observation data. In terms of monitoring accuracy, the coefficient of determination (R²) for the extracted jointing and heading dates were both higher than 0.55, the root mean square error (RMSE) were both between 4 and 6 days, and the BIAS were both within 3 days. In contrast, the monitoring accuracy of milking maturity date was slightly lower, with R² between 0.35 and 0.50, RMSE between 5 and 7 days, and BIAS within 2 days. Compared with the monitoring accuracy, the forecasting accuracy decreased slightly (BIAS increased by 0.4 to 3.8 days and RMSE increased by 0.2 to 2.5 days), but was still at a high level. Moreover, the forecasting accuracy based on the samples from the previous two years was generally higher than that based on the samples from the previous single year (BIAS changed less than 0.5 days and RMSE changed less than 0.2 days).