基于半监督机器学习方法的火灾风险遥感评估研究

发布时间：2018-05-13 16:24

本文选题：自然灾害 + 火灾　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年博士论文

【摘要】：火灾是全球性常发自然灾害之一,每年发生的火灾都会破坏大量地表植被系统,极大地影响了农林生产和经济发展。中国是一个火灾多发国家,是全球受火灾影响最为严重的前20个国家之一。特别是,在全球变化背景下,中国需要考虑科学有效的降低火灾风险策略,应对不断增多的火灾风险。遥感数据作为研究大气与地表过程的一种重要信息源,具有许多无可比拟的特点和优势。目前,不同空间分辨率、光谱分辨率、时间分辨率的遥感数据提供了丰富的火情信息,利用星载或机载传感器,既可以对区域性的突发火情快速高精度地观测,也可以实现全球性的过火区域监测。许多学者开始利用不同遥感数据及其产品,如Suomi-NPP VIIRS、MODIS、ATSR、SPOTVEGETATION、ASTER、AVHRR、风云等对过火事件进行研究,并在火点识别、过火面积提取等方面取得了长足进展。本文重点关注火灾风险的遥感评估问题,主要从两个方面展开:1)讨论和研究影响中国南部火灾风险的关键气候和环境参量,2)选取关键气候和环境参量,实现火灾风险的遥感建模评估。在第一方面研究中,本文重点研究了降水量(PPT)、蒸散发(ET)、潜在蒸散发(pET)等多气候和环境参量对火灾风险的贡献机制,主要包括选取了与火灾诱发相关的系列气候和环境参量,开展火灾密度与各参量之间的网格相关分析。其中,基于网格的相关分析从三个时间尺度进行展开,即年度、冬季(12月-次年2月)和春季(3月-5月),后两个时间尺度的选择主要考虑到春季和冬季是中国南部地区易发生火灾的季节。本文研究表明,降水量、蒸散发等气候和环境参量同火灾风险在季节性尺度上存在相关性,但这种相关性还同时与时间、空间条件密切相关,即中国东南部地区气候和环境参量同火灾风险在冬季的相关性较大,而西南部地区则在春季呈现强相关。同时发现,与中国南部地区火灾活动最显著相关的参量是潜在蒸散发和蒸散发与潜在蒸散发的比值(ET/pET),其他参量如水分平衡和降水量等也与火灾活动相关,但这些参量主要影响了华东南部冬季的火灾风险。综合来看,气候和环境参量与火灾活动紧密相关,可以作为中国南部地区火灾风险评估的指示性指标。在第二方面的研究中,本文提出应用半监督机器学习方法来进行火灾风险的建模评估。该机器学习方法是一种在仅有正例和未标注样本的训练数据集下进行机器学习的特殊半监督学习方法。在本部分的研究中,考虑到网格的空间分辨率得到提高,为了方便处理大量数据,本文将研究区域缩小至中国东南部地区,并主要考虑常绿阔叶林、混合林和多树草地三种不同的地类。对于上述三类不同的地表植被覆盖类型,本文对中国东南部地区的火灾风险进行了评估。研究发现,与常绿阔叶林和多树草地相比,基于半监督机器学习方法的火灾风险评估模型在混合林区域的准确性更好。综上,本文所提出的半监督机器学习方法为理解火灾风险问题提供了新的方法贡献。当前,本文主要采取了针对不同地类分别构建单一性评估模型的方法,在未来的改进方面,还需考虑研究一种可用于整个地区、适用于多类地物的火灾风险评价模型。同时,本文提出的方法在精度上仍有提升空间,如可考虑选取一些与火灾活动具有较高相关性并表现出高空间变异性的气候和环境特征来构建模型。总体而言,本文提出的半监督机器学习方法是利用遥感数据进行火灾风险评估的有效手段。
[Abstract]:Fire is one of the most common natural disasters in the world. Every year's fires will destroy a large number of surface vegetation systems, which greatly affect the production and economic development of agriculture and forestry. China is a fire prone country and one of the most serious fire affected countries in the world. In particular, China needs to consider the section under the background of global change. It is effective to reduce fire risk strategies and respond to increasing fire risk. Remote sensing data, as an important source of information for the study of atmospheric and surface processes, has many unparalleled features and advantages. At present, remote sensing data with different spatial resolution, spectral resolution and time resolution provide rich information and use of stars. A load or airborne sensor can be used for rapid and high precision observation of a regional burst of fire and a global monitoring area. Many scholars have begun to use different remote sensing data and their products, such as Suomi-NPP VIIRS, MODIS, ATSR, SPOTVEGETATION, ASTER, AVHRR, and wind clouds to study the fire events and identify the fire points, This paper focuses on the remote sensing assessment of fire risk, mainly from two aspects: 1) discuss and study the key climate and environment parameters affecting the fire risk in southern China, 2) select the key climate and environmental parameters, and realize the evaluation of remote sensing modeling for fire risk. In this study, this paper focuses on the contribution mechanism of precipitation (PPT), evapotranspiration (ET), potential evapotranspiration (pET) and other climatic and environmental parameters to the fire risk, including the selection of a series of climate and environmental parameters related to the fire induced, and the grid correlation analysis between the fire density and the parameters. The analysis is carried out from three time scales, namely, annual, winter (December - February) and spring (March -5 months). The first two time scales are selected mainly considering that spring and winter are the prone fire seasons in southern China. This paper shows that the climate and environmental parameters such as precipitation, evapotranspiration and other climate are on the seasonal scale. There is a correlation, but the correlation is closely related to time and space conditions, that is, the climate and environment parameters in the southeast of China are related to the fire risk in winter, while the southwest region is strongly correlated in the spring, and the most significant parameter related to the fire activity in the southern part of China is the potential evapotranspiration. The ratio of hair and Evapotranspiration to potential evapotranspiration (ET/pET), other parameters such as water balance and precipitation are also related to fire activities, but these parameters mainly affect the risk of fire in winter in the south of East China. In the second aspect of the study, this paper proposes a semi supervised machine learning method for the modeling and evaluation of fire risk. The machine learning method is a special semi supervised learning method for machine learning under a training data set with only positive and unlabeled samples. In order to improve the spatial resolution, in order to facilitate the processing of a large number of data, this paper narrowed the research area to the southeast of China, and mainly considered the evergreen broad-leaved forest, mixed forest and multi tree grassland, three different types of land. For the above three types of surface vegetation cover types, this article reviews the fire risk in the southeast of China. The study found that the accuracy of the fire risk assessment model based on semi supervised machine learning method is better than that of evergreen broad-leaved forest and multi tree meadow. To sum up, the semi supervised machine learning method proposed in this paper provides a new method to understand the problem of fire risk. In the future improvement, we also need to consider a fire risk assessment model which can be used in the whole area and suitable for multi class objects in the future. At the same time, the method proposed in this paper still has the space to improve the accuracy, for example, it can be considered to have a higher correlation with the fire activity. In general, the semi supervised machine learning method proposed in this paper is an effective means of using remote sensing data to assess the risk of fire.

【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：博士
【学位授予年份】：2017
【分类号】：X932;TP79

【相似文献】