干旱作为世界上最严重的自然灾害之一，具有持续时间长、影响范围广、灾害损失严重等特点，对其准确连续监测一直是众多学者研究的焦点。传统干旱监测方法虽在站点尺度精度较高，但需要布置大量站点，耗时耗力。遥感技术可实现大面积实时动态监测，但其对旱情监测的准确性仍有待研究。随着机器学习不断发展，将其应用于旱情监测可以同时发挥站点数据的精度高以及遥感数据的时空尺度优势。中国华北平原是我国重要的粮食产区，其多年来一直受到旱情的影响，因此建立一个准确且可大面积连续监测的遥感干旱指数对华北平原旱情分析具有重要意义。本文以华北平原及其周围地区为研究区（包括北京市、河南省、山东省、天津市、河北省、江苏省和安徽省，后文简称华北地区），基于游程理论，分别利用标准化降水指数（Standardized Precipitation Index，SPI）和标准化降水蒸散指数（Standardized Precipitation Evapotranspiration Index，SPEI）监测了华北地区的干旱次数、历时和烈度；基于标准化异常指标对华北地区干旱状况进行了分析，并探讨其适用性；然后基于随机森林（RF）、支持向量机（SVM）以及反向传播人工神经网络（BP-ANN）模型，利用各干旱因子的标准化异常指数、数字高程模型和气象站点的SPI实现多源综合干旱指数（Multi-source Integrated Drought Index，MIDI）的构建以及模型精度的验证；最后将MIDI应用到了华北地区旱情监测分析，并将模型迁移到了中国西南和美国大平原地区的典型干旱事件进行监测，得出结论如下：

（1）2007~2018年，华北地区气象站点SPI和SPEI监测到的干旱次数、历时和烈度具有较强的一致性。SPI和SPEI监测到华北地区气象站点发生的干旱次数均超900次，其中以持续两个月的干旱事件为主，占比均达70%以上，而大于等于4个月的干旱事件总占比较少，且单站点累计烈度大多位于20~30之间。从单个干旱因子（蒸散发、地表温度、潜在蒸散发、土壤水分、降水、归一化植被指数和日光诱导叶绿素荧光）的标准化异常来看，2007~2018年华北地区单个指标受到干旱影响的占比为44.44%~53.47%，其中气象和水文因子至少有一个指标受到胁迫的比例为90.28%，从空间上来看，单个因子在像元尺度上受到干旱胁迫的次数大多超过了60次，表明研究区在此期间干旱发生较为频繁。

（2）通过气象站点SPI和标准化异常指数的相关性分析可知，各干旱因子的标准化异常指数均通过了P<0.01的显著性检验，且多因子干旱模型优于单因子模型。因此，以蒸散发、地表温度、潜在蒸散发、土壤水分、降水、归一化植被指数和日光诱导叶绿素荧光的标准化异常指数以及数字高程模型为自变量，以一个月尺度的SPI作为因变量，基于RF、SVM和BP-ANN进行模型构建。对比模型评价指标可知，RF模型的R²最高、RMSE和MAE最小，为最优旱情监测模型。三个站点验证结果也表明，RF与SPI拟合效果最好，其中以泰山站点最为明显（r=0.904，P<0.01）。因此本文用RF构建了MIDI，其与SPEI的相关性无论在月尺度还是不同植被覆盖类型中均具有强相关性（r>0.8，P<0.01）。

（3）从华北地区月均值来看，2007~2018年华北地区MIDI均值主要位于-0.5~0.5之间，低于-1的月份出现了4次，即华北地区旱情在2008年2月、2010年11月、2011年4月和2012年5月达到了重旱，较为严重。从MIDI季节特征来看，2011年主要发生的干旱为春旱和冬旱，其中春旱较为严重，在河南省、江苏省和安徽省大部分地区出现了重旱；2013年几乎每个季节都发生了不同程度的干旱；2014年以夏旱和冬旱为主，夏旱尤为严重，在河南省、山东省和河北省地区出现重旱。在干旱事件监测中，MIDI不仅可准确监测2013年和2014年华北地区所发生的干旱事件，且与SPEI具有较强的相关性（r>0.7，P<0.01）；且在迁移研究区后，MIDI仍适用于中国西南地区和美国大平原地区的干旱监测，与SPEI的r>0.5（P<0.01）。

Drought, as one of the most serious natural disasters in the world, has the characteristics of long duration, wide range of impact and serious loss. The accurate continuous monitoring of drought has been the focus of many scholars. Although the traditional drought monitoring method has high precision in site scale, it needs to arrange a large number of sites and consumes time and energy. Remote sensing technology can realize real-time dynamic monitoring of large area, but its accuracy of drought monitoring remains to be studied. With the continuous development of machine learning, applying it to drought monitoring can give full play to the advantages of high accuracy of site data and spatio-temporal scale of remote sensing data. The North China Plain is an important grain producing area in China, which has been affected by drought for many years. Therefore, it is of great significance to establish an accurate remote sensing drought index which can be continuously monitored over a large area for drought analysis in the North China Plain. This paper takes the North China Plain and its surrounding areas as the research area (including Beijing, Henan, Shandong, Tianjin, Hebei, Jiangsu and Anhui, later referred to as the North China Region). Based on the run theory, the frequency, duration and intensity of drought in North China were monitored by standardized precipitation index and standardized precipitation evapotranspiration index respectively. Based on the standardized anomaly index, drought status in North China is analyzed and its applicability is discussed. Then, based on random forest, support vector machine and back propagation artificial neural network, the standardized anomaly index of each drought factor, digital elevation model and SPI of meteorological stations were used to construct the multi-source comprehensive drought index, and the accuracy of the model was verified. Finally, MIDI was applied to drought monitoring and analysis in North China, and the model was transferred to the typical drought events in southwest China and the Great Plain of the United States for monitoring, the conclusions were drawn as follows:

(1) From 2007 to 2018, the frequency, duration and intensity of drought detected by SPI and SPEI at meteorological stations in North China showed strong consistency. According to SPI and SPEI, there were more than 900 droughts at meteorological stations in North China, among which drought events lasting for two months were the main ones, accounting for more than 70%, while drought events lasting for more than four months accounted for a relatively small proportion, and the cumulative intensity of a single station was mostly between 20 and 30. From the standardization anomaly of single drought factors (evapotranspiration, surface temperature, potential evapotranspiration, soil moisture, precipitation, normalized vegetation index and daylight induced chlorophyll fluorescence), the proportion of single index affected by drought in North China during 2007~2018 was 44.44%~53.47%. The proportion of at least one index of meteorological and hydrological factors under stress was 90.28%. From the perspective of space, most of The Times of drought stress on the pixel scale of a single factor were more than 60 times, indicating that drought occurred more frequently in the study area during this period.

(2) Based on the correlation analysis between SPEI and the state index and the standardized anomaly index, we can see that the optimal independent variable is the standardized index and the optimal dependent variable is SPI, and the multi-factor drought model is better than the single factor model. Therefore, with evapotranspiration, surface temperature, potential evapotranspiration, soil moisture, precipitation, normalized vegetation index, normalized anomaly index of sun-induced chlorophyll fluorescence and digital elevation model as independent variables, and one-month scale SPI as dependent variable, the model was constructed based on RF, SVM and BP-ANN. Compared with model evaluation indexes, RF model had the highest R² and the lowest RMSE and MAE, which was the optimal drought monitoring model. The results of three sites also showed that the fitting effect of RF and SPI was the best, among which Mount Tai sites was the most obvious (r=0.904, P<0.01). Therefore, MIDI was constructed with RF in this paper, and its correlation with SPEI was strongly correlated in both monthly scale and different vegetation cover types (r>0.8, P<0.01).

(3) From the perspective of monthly mean value of North China, MIDI average value of North China from 2007 to 2018 was mainly between -0.5 and 0.5, and there were four months lower than -1, that is, the drought in North China reached severe drought in February 2008, November 2010, April 2011 and May 2012. According to the seasonal characteristics of MIDI, the main droughts in 2011 were spring drought and winter drought, among which spring drought was more serious and severe drought occurred in most parts of Henan Province, Jiangsu Province and Anhui Province. Nearly every season in 2013 saw some degree of drought; In 2014, there were mainly summer and winter droughts, with the summer drought being particularly severe, with severe droughts in Henan, Shandong and Hebei provinces. In drought event monitoring, MIDI could not only accurately monitor the drought events in North China in 2013 and 2014, but also had a strong correlation with SPEI (r>0.7, P<0.01). MIDI was still suitable for drought monitoring in southwest China and the Great Plain of the United States after the relocation of the study area, and SPEI's r>0.5 (P<0.01).