【作者】 黄友昕；

【导师】 刘修国；

【作者基本信息】 中国地质大学 ， 测绘科学与技术， 2021， 博士

【摘要】 农业干旱是一种反复出现、持续时间长、无结构化的自然灾害。在全世界所有土地上都曾经历过不同程度的农业干旱事件,尤其是经济来源大部分依靠于农业生产的国家。农业干旱不仅直接造成农作物大面积减产绝收,而且对社会经济造成巨大的损失,严重地影响了农业的可持续发展和社会的稳定。因此,如何有效地监测农业干旱,因地制宜地选取合适的干旱监测方法已成为抗旱减灾部门和农业管理部门面临的一项紧迫的任务。基于遥感技术的农业干旱监测方法,常常通过农业干旱遥感监测指数来实现,它具有客观、及时、覆盖范围广等优点,弥补了地面站点的不足,已被证明是农业干旱监测中最有前景的技术手段。但是不同的农业干旱遥感监测指数具有明显相异的时空适应性。这些指数从遥感光谱信息中提取干旱发生与发展的特征,它既受空间方面区域下垫面的影响;又受时间方面作物不同物候期的生长形态的影响。如何根据不同区域下垫面、作物不同物候期选取适合的农业干旱遥感监测指数,是精确评估和监测农业干旱的基础。评价农业干旱遥感监测指数的适应性主要采用光谱特征匹配法、多元统计分析法、主成分分析法和人工神经网络分析法等方法。然而,由于作物生长的土壤、物候期、地形、气候等时空多因素相互作用的复杂性,从环境依赖的角度,对农业干旱遥感监测指数的时空适应性与敏感性还需进一步研究;另外对农业干旱监测指数的适应性评价常常带有一定的人为主观性和经验性问题,客观性与自动化水平也有待改进。针对上述问题,本文利用多源遥感数据、气象站点实测数据、土壤湿度数据等等,从不同区域下垫面特征、作物不同物候期以及综合时空多因素等多个角度,利用机器学习技术,研究不同农业干旱监测参量与作物环境多因子之间的关联关系,提出顾及时空多因素的农业干旱遥感监测方法,并评价这些方法在不同应用场景的适应性。论文主要研究工作和成果有:(1)提出一种顾及下垫面改进标准化降雨蒸散指数的农业干旱监测方法针对标准化降雨蒸散指数(SPEI)对区域不同下垫面特征的差异性考虑不足,对干旱的响应具有明显的区域差异性问题,本文在综合考虑地形高程、土地类型(灌丛、草地、耕地、裸地)等下垫面多因素条件下,提出一种顾及下垫面的改进SPEI。并以实测降雨量、气温以及有效土壤水含量数据计算得到的PDSI、sc PDSI、SPI为基准,对该方法进行验证。研究结果表明:改进的SPEI指数对内蒙古雨养农业区近40年的农业干旱演变情况与实际旱情相符;其监测结果与44个旗县级行政区的站点数据sc PDSI、实测SPI(1月、3月尺度)指数相关系数均通过了显著性检验。该方法更适合于雨养农业区的干旱监测场景。(2)提出一种顾及物候期反演土壤湿度的农田墒情监测方法土壤湿度能够反映作物土壤含水量和作物生物量的状态,反演土壤湿度对评估作物干旱状况和生长环境条件至关重要。而土壤湿度是一个复杂的非线性耦合系统,受土壤复杂结构和作物环境多因子影响显著,如何分析多源输入与输出间的非线性映射关系,提高土壤湿度反演精度是值得研究的问题。而人工神经网络模型能自动分析多源输入与输出间的非线性映射关系,基于此,本文在顾及作物物候期的条件下,以冬小麦返青期为例,基于MODIS干旱指数与径向基神经网络方法,提出一种顾及物候期反演土壤湿度的农田墒情遥感监测方法。研究结果表明:反演的土壤湿度应用在河南省农业干旱墒情监测中效果较好;相比线性模型与BP神经网络反演土壤湿度精度更高,该模型回归分析相比1:1线的偏差最小;反演的平均预测精度达到93.27%,相关系数为0.846,决定系数为0.862 6。这表明MODIS干旱指数结合径向基神经网络协同反演冬小麦返青期的土壤湿度模型有效。该方法较适合于区域农田土壤墒情的干旱监测场景。(3)提出一种综合时空多因素的复合农业干旱遥感监测方法针对综合遥感干旱监测指数的权重设定客观性与自动化不足问题。本文基于深度学习方法,在非显式定义下垫面特征的情况下,引入卷积神经网络方法,自动学习时空多环境因素与多农业干旱遥感监测参量之间的关系与规则,构建一种农业干旱遥感监测指数重要性评价及复合指数深度学习网络模型(Ieci Net)。Ieci Net模型相比其它传统机器学习模型的精度更高。Ieci Net模型在拟合实测旱情参量的同时,还能够自动从MODIS遥感数据中定量地获取农业干旱遥感监测指数的重要性系数。以此重要性系数为权重,提出一种综合时空多因素的复合农业干旱遥感监测方法。同时以站点干旱指数sc PDSI、u SPEI和土壤湿度数据为基准,验证该复合农业干旱监测指数的有效性。研究结果表明复合农业干旱监测指数在不同气候干湿分区均有较好的监测效果,该方法适合于区域局部下垫面较复杂的干旱监测场景。更多还原

【Abstract】 Agricultural drought is a recurrent natural disaster,and it lasts for a long time and has no structure.There have been experienced agricultural drought events of different frequency and severity on the land all over the world,especially in the countries whose economic sources mostly depend on agricultural production.Agricultural drought not only caused directly a large area of crop yield reduction,but also caused huge losses to the social economy,seriously affected the sustainable development of agriculture and social stability.Therefore,how to effectively monitor agricultural drought and select appropriate drought monitoring methods according to local conditions has become an urgent task for drought relief departments and agricultural management departments.The agricultural drought monitoring method based on remote sensing technology is often implemented through remote sensing derived agricultural drought monitoring indiecs.It has the advantages of objectivity,timeliness and wide coverage,and it makes up for the shortage of ground stations and has been proved to be the most promising technical means in agricultural drought monitoring.However,different remote sensing derived agricultural drought monitoring indices have different spatiotemporal adaptability.The characteristics of drought occurrence and development from the remote sensing spectral information are extracted by these indices.But the characteristics of drought are affected by the underlying surface in space;it is also affected by the growth pattern of crops in different phenological periods.How to select suitable remote sensing monitoring index of agricultural drought is the basis of accurate evaluation and monitoring of agricultural drought according to different underlying surface and different crop phenology.Spectral feature matching method,multivariate statistical analysis method,fuzzy comprehensive evaluation method,principal component analysis method and artificial neural network analysis method are used to evaluate the adaptability of agricultural drought remote sensing monitoring index.However,due to the complexity of the interaction of soil,phenology,topography,climate and other spatiotemporal factors of crop growth,from the perspective of environmental dependence,the spatiotemporal adaptability and sensitivity of agricultural drought monitoring index need to be further studied.In addition,the adaptability evaluation of agricultural drought monitoring index often has some human subjectivity and empirical problems,and its objectivity and automation level also need to be improved.Based on the above problems,this paper uses multi-source remote sensing data,meteorological station measured data,soil moisture data and so on,from the characteristics of different regional underlying surface,different phenological periods of crops and comprehensive spatiotemporal factors and other aspects,using machine learning technology to study the correlation between different agricultural drought monitoring parameters and crop environmental factors.This paper proposes remote sensing derived agricultural drought monitoring methods concerning spatiotemporal multi-factor.In addition,the adaptability of these methods in different application scenarios is evaluated.The main research work and achievements are as follows:(1)An improved standardized rainfall evapotranspiration index for agricultural drought monitoring is proposed considering underlying surfaces.In view of the obvious spatial-temporal difference of standardized rainfall evapotranspiration index response to drought under different regional underlying surface conditions,an improved SPEI is proposed concerning different underlying surface environmental factors,such as terrain elevation,land type(shrub,grassland,cultivated land and bare land)and other factors.At the same time,The PDSI,sc PDSI and SPI calculated from the measured rainfall,air temperature and available soil water content data are used as the benchmark to verify the method.The results show that the improved SPEI is consistent with the actual drought situation in the rain fed agricultural area of Inner Mongolia in recent 40 years;the correlation coefficients between the monitoring results and the sc PDSI and SPI(January and March scale)of 44 counties have passed the significance test.This method is more suitable for drought monitoring scenarios in rain fed agricultural areas.(2)In this paper,a method of soil moisture monitoring is proposed considering phenological periods.Soil moisture can reflect the state of crop soil water content and crop biomass,and retrieving soil moisture is very important to evaluate crop drought and growth environment.Soil moisture is a complex nonlinear coupling system,which is significantly affected by the complex soil structure and crop environment factors.How to analyze the nonlinear mapping relationship between multi-source input and output and improve the accuracy of soil moisture inversion is a problem worthy of study.The artificial neural network model can automatically analyze the nonlinear mapping relationship between multi-source input and output.Based on this,this paper proposes a remote sensing monitoring method of agricultural drought considering phenological periods,taking the turning green period of winter wheat as an example,based on MODIS drought index and radial basis function neural network method.This studies reveal that the inversion of soil moisture has a better effect in agricultural drought monitoring in Henan Province;compared with the linear model and BP neural network,the inversion of soil moisture has higher accuracy,and the regression analysis of the model has the smallest deviation compared with the 1:1 line;the average prediction accuracy of inversion is 93.27%,the correlation coefficient is 0.846,and the determination coefficient is 0.8626.The results show that the soil moisture model of Winter Wheat in green period can be effectively retrieved by using MODIS multi band drought index and radial basis function neural network.This method is more suitable for remote sensing monitoring of regional farmland soil moisture.(3)In this paper,a remote sensing monitoring method for agricultural drought is proposed,which integrates spatial and temporal multi-factor.In order to solve the problem of objectivity and automation in the weight setting of comprehensive remote sensing drought monitoring index.Based on deep learning method,convolution neural network method is introduced to automatically learn the relationship and rules between spatiotemporal multi environmental factors and multi agricultural drought remote sensing monitoring parameters without explicitly defining underlying surface features.A deep learning network model(Ieci Net)for agricultural drought remote sensing monitoring and its adaptability evaluation is proposed.Compared with other machine learning models,Ieci Net model has the highest accuracy.While fitting the measured drought parameters,Ieci Net model can also automatically obtain the importance coefficient of agricultural drought monitoring index from MODIS multi-band data.Taking the importance coefficient as the weight,a composite remote sensing monitoring method of agricultural drought based on spatial-temporal multi factors is proposed.At the same time,based on soil moisture data,site drought index PDSI,sc PDSI and u SPEI,it is verified that the composite agricultural drought monitoring index had good monitoring results in different climate dry and wet regions.This finding indicates that it is a suitable method for agricultural drought monitoring scenarios with complex local underlying surface.更多还原