【作者】 江昊；

【导师】 林涛；

【作者基本信息】 浙江大学 ， 生物系统工程， 2022， 博士

【摘要】 利用数据驱动方法融合多源物候、气象和遥感数据,为准确估计大尺度玉米单产和评估气象因子的影响提供了切实可行的工具和方法,对保障粮食安全至关重要。目前玉米估产模型大多基于单一时空尺度,无法体现玉米的生长动态过程和时空交互因素,在模型时空结构上的不足为数据驱动估产模型在大尺度应用带来不确定性。本文针对目前大尺度数据驱动估产模型研究中存在的玉米物候信息时空分辨率不足、模型时空尺度效应影响不明、气象因子对于玉米单产的时空异质性影响不明和缺少融合生长动态和多源数据的估产框架等问题开展工作。以美国玉米带为研究区域,实现不同尺度玉米物候信息遥感提取、气象估产模型尺度效应评估、贝叶斯时空结构下非平稳关系拟合和基于深度学习的多源数据融合估产框架,为大尺度玉米估产和气象影响评估提供了新的方案。主要的研究结论如下:（1）针对大尺度物候信息时空分辨率不足问题,使用MODIS遥感观测WDRVI植被指数,结合改进的形态模型法实现对6个玉米关键物候期在像元-县-农业区-州级不同尺度的提取。经过统计数据的验证,形态模型法可将传统阈值法在州尺度玉米物候提取的RMSE平均降低11.9天;在农业区尺度,形态模型法提取的RMSE最低为出苗进度的5.1天,RMSE最高的是乳熟进度为7.8天;在县级尺度,提取的玉米物候空间分布符合纬度规律,表明该方法在大尺度玉米物候提取上有良好效果。（2）针对模型时空尺度导致的定量分析模型解释性差异,评估研究区气象估产的最佳时空尺度并定量分析关键生育期和关键气象因子。结果表明以农业区尺度进行建模,校正R²相比州和县尺度平均高出11.04%。玉米单产对气象因子响应的敏感性分析表明,热害积温（KDD）是最主要的影响因子,玉米带南部在抽丝-乳熟阶段,每单位KDD升高会造成20.14 kg·hm^-2的单产损失,而玉米带北部地区在乳熟-蜡熟阶段,每单位KDD升高会导致21.72 kg·hm^-2的单产损失。播种-抽丝期的过量降水会对玉米单产造成负面影响,每升高1mm降水平均降低2.19 kg·hm^-2的单产。（3）针对大尺度玉米单产对气象变化响应的时空自相关和非平稳过程建模问题,提出基于层次贝叶斯结构的时空可变系数（STVC）模型,量化分析玉米单产-气象因子在时间和空间维度上异质性变化规律。该模型考虑气象因子和玉米单产之间的时空非平稳性,拟合时空可变回归系数,分析产量对气象因子响应的局部和时空关联过程。结果表明,STVC模型相比其他五种结构的时空效应模型,RMSE平均降低了35.02%,空间自相关度量指标Moran’s I平均降低0.05,同时在气象灾害严重的年份依旧能有效拟合气象因子变化的影响。实验结果表明,STVC模型可拟合气象因子对产量影响的时空特征和变化趋势。（4）针对玉米产量形成的累积过程及复杂非线性关系表征问题,构建融合物候、气象、遥感多源数据的深度学习估产框架,并以生长特征指导深度学习模型结构。结果表明,长短期记忆神经网络（LSTM）可以有效融合多源气象和遥感数据进行估产,相比单独使用气象或遥感数据的方法,估产误差分别降低了41.21%和20.25%。LSTM模型能从不断增加的样本量中学习有效单产特征,对比1年和10年训练数据,测试集的RMSE由1340 kg·hm^-2下降至870 kg·hm^-2。对比LASSO和RF模型,LSTM在极端干旱年份和动态估产方面的稳定性和精度都有显著提高。季中动态估产实验结果表明,抽丝-乳熟期是影响玉米单产的关键阶段,这一阶段的输入信息对玉米估产精度影响最大。更多还原

【Abstract】 Data-driven models are practical approaches to conflate multi-source phenological,meteorological and remote sensing data for estimating large-scale corn yields and assessing the impact of meteorological factors,which are vital to ensuring food security.However,most corn yield estimation models are based on a fixed spatiotemporal scale,which cannot reflect the dynamic processes of corn growth and the interactions across time and space.The insufficiency of the model structure brings uncertainty to the large-scale application of data-driven yield estimation models.This paper aims to address the problems such as the insufficient high-resolution corn phenology information,the unclear influence of scale effect,the complexed spatiotemporal autocorrelation processes of yield-meteorological interactions,the lack of efficient framework for multi-source data fusion and yield estimation.Taking the U.S.corn belt as the research area,this research focused on extracting multi-scale corn phenology information using remote sensing data;evaluating the scale effect on regression models between yield and meteorological factors;constructing the non-stationary process between yield and meteorological factors based on Bayesian spatiotemporal structure;and developing a deep learning framework for yield estimation using multi-source data.Solutions for large-scale corn yield estimation and meteorological impact assessment are provided in this research.Here presents the main researches and conclusions in this work as following:（1）A refined shape model approach based on MODIS data is proposed for multi-scale（pixel-county-district-state）corn phenology retrieval.This approach uses the morphological characteristics of WDRVI time-series to estimate corn phenological dates for 6 key stages.Comparing with the traditional threshold method,the shape model method reduces the RMSE by 11.9 days on average at the state-level.At the agricultural district level,the shape model method has the lowest RMSE as 5.1 days for emergence dates detection.The highest RMSE is6.9 days for planting dates detection.At the county level,the spatial distribution of the detected corn phenology is correlated with the latitude,indicating that this method is eligible for large-scale corn phenology detection.（2）Developed and compared statistical models with various spatiotemporal resolutions to understanding the scale effect on corn models.Multilinear regression models are trained using cross-sectional datasets pooled at three spatial resolutions（state,district,county）with temperature and precipitation related predictors according to two temporal resolutions（growing season,growing phase）.The results show that when modeling at the agricultural area scale,the adjusted R² is 11.0%higher than that at the state and county scales on average.Sensitivity analysis of the response of corn yield to meteorological factors showed that the killing degree days（KDD）is the most important factor.In the southern part of the corn belt in the silking-dough period,one unit of increase in KDD could cause a yield loss of 20.14 kg·hm^-2.In the northern part of the belt,during the dough-dented period,one unit of increase in KDD could result in a yield loss of 21.72 kg·hm^-2.Excessive precipitation during the planting-silking period has a negative impact on the corn yield,and the yield could be reduced by an average of 2.19kg·hm^-2 for every 1 mm increase of precipitation.（3）A Bayesian hierarchical structure-based spatiotemporally varying coefficient（STVC）model was constructed to quantitatively analyze the heterogenous interaction structure between corn yield and meteorological factors across space and time.The model based on temporal and spatial non-stationary processes,and fits the local yield response to meteorological factors.The results show that the STVC can reduce the degree of spatial aggregation for estimated residuals.Compared with the other five structures of time-space effect models,the average RMSE is reduced by 35.02%,and the measure of spatial autocorrelation Moran’s I is reduced by an average of 0.05.Even in the years when meteorological disasters are severe,the impact of meteorological factors can still be effectively fitted by STVC model.The results show that the STVC model can fit the spatiotemporal correlation and non-stationary process of the impact of meteorological factors on yield,which is of reference significance for understanding the response of corn yield to meteorological factors.（4）A deep learning yield estimation framework that integrates phenological,meteorological,and remote sensing data is developed based on the long short-term memory neural network（LSTM）structure.Compared with methods using single meteorological or remote sensing data,the use of multi-source meteorological and remote sensing data can reduce 41.21%and 20.25%of RMSE in county-level corn yield estimation.According to the sensitivity analysis results of the changes in the number of training samples,the LSTM model can learn effective yield characteristics from the increasing sample size.The LSTM model can improve the feature extraction ability by increasing the sample size.Compare the training data with 1-year and 10-year samples,the RMSE of the LSTM model in the test year decreased from 1450 kg·hm^-2 to870 kg·hm^-2.The LSTM model also outperformed LASSO and random forest approaches in the drought year.According to the results of in-season yield prediction,the observations at the silking-dough period is the most critical for prediction accuracy.更多还原