基于遥感热辐射增量的早期野火自动识别研究

Using satellite data of NOAA, MODIS and FY etc.，the absolute thermal radiation index or brightness temperature value are difficult to identify the early forest fire, of which the essential reason is that the fire lines or areas are too small in the remote sensing images with 1 km resolution to make brightness temperature of each pixel up to or close to the saturation value. Brightness temperature of thermal radiation of some of fire lines or areas is even lower than that of bare rock, bare soil, sand and valley. To solve this problem, the study propose a procedure of automatically identifying and extracting the forest fire by the Spatio-Temporal catastrophe characters of forest fires in forest regions, which will be construct a relative increment image of brightness temperature of thermal radiation based on the backgrounds average temperature baseline by the natural geographical patterns. And using the robust Spatio-Temporal index between forest fire and other mutation characteristics of high temperature heat source, we will build an intelligent neural model of Deep Stacking Network for automatically recognizing and extracting the forest fires. We will focus on the study of the relationship and models between the remote sensing instantaneous temperature of forest vegetation and the average temperature fields of background objects, a method of constructing the relative increment image of brightness temperature of thermal radiation, the pattern models of the forest fires and the Spatio-Temporal index of an increment of brightness temperature of thermal radiation. Finally we will build an artificial intelligence model for the forest fires Comprehensive recognition. The study will break through the technical limits of early forest fire monitoring by satellites. The study is expected to resolve the application difficulties of extracting the early forest fire, which have been existed for more than 20 years for forest fire monitoring. The key techniques will provide new ideas and methods for study the mechanism of abnormal changes monitoring of surface thermal infrared.

由于火线及过火区在1公里像元内所占的面积太小，无法使像元的热辐射亮温达到或接近饱和值，甚至远低于裸岩裸地、沙滩、沟谷等强增温地物，利用NOAA、MODIS和FY等遥感图像及绝对热辐射亮温指标无法识别早期森林火灾。针对以上问题，提出了以自然地理格局的背景温度为基准，构建相对热辐射亮温增量图像，并利用森林分布区小火的时空突变特征和其它高温热源的稳态特征指标，构建智能化深度神经网络模型，自动识别和提取森林火灾的方法。主要研究森林植被的遥感瞬时温度与背景均温的关系，以及热辐射亮温增量图像构建方法；研究林火像元图像时空模式结构，提出火灾亮温增量的空域及频域变化指标，森林火灾综合识别的 人工智能模型，以突破卫星遥感监测只能开展手工解译的技术瓶颈。研究有望解决20多年来卫星林火监测难以提取早期森林火灾的应用难题，将为地表热红外异常变化的遥感监测机理研究提供新的思路和方法。

针对利用MODIS等中低遥感图像及绝对热辐射亮温指标无法识别早期森林火灾的问题，研究提出了以自然地理格局的背景温度为基准，构建相对热辐射亮温增量图像，并利用森林分布区小火的时空突变特征和其它高温热源的稳态特征指标，自动识别和提取森林火灾的方法。通过构建自然地域热辐射亮温基线，以中低分辨率遥感图像精校正处理，火区遥感图像亚像元分解与增强为预处理，开展了时间序列图像的上的地表亮温增量”突变“的识别；基于空间邻域”突变”特征和植被参数进一步识别森林草原野火的方法；构建了热辐射亮温增量阙值指标，提出了时间尺度上森林草原初发火、小火识别的阙值方法研究；提出了依据“时间过程突变、空间过程突变”热辐射亮温增量“初发火、小火识别”的原理；研发了初发火的计算机自动识别算法及软件系统。基于南方卫星林火监测分中心的业务系统和数据，选择多个正在发生的典型火场验证了基于“热辐射亮温增量”和“时空突变”提取小火的遥感图像处理理论和方法，结果较为可靠。将识别方法整理，设计逻辑流程、数据流程，并进行编码和形成模块软件，得到人工智能化的林火综合识别模型。研发了森林草原初发火的计算机自动识别算法设计及软件系统，并通过了第三方的软件测试。