三江源区大面积退化草地毒杂草组分覆盖信息多源遥感协同提取方法研究

In our previous NFSC project (41271361) and other works we found that the rangeland degradation monitoring methods, which is represented by remote sensing vegetation index, can hardly reflect the change of community structure. In addition, the physical restriction of sensor itself leads to the result that from the same sensor we cannot obtain data covering large area and with high temporal, spatial and spectral resolution at the same time. In this project, over the "Three-River Headwaters" region we are aiming at: (1) to establish poisonous weed "growth type" and related spectrum acquisition; (2) to study on the complementary properties and fusion methods of multi-source remote sensing data, and to propose methods for extraction poisonous weeds based on multi-source remote sensing synergy technique over large-scale region which is lack of high-resolution remote sensing data; (3) to construct effective feature space for extraction poisonous weeds from their spectral, phenology, biochemical and habitat features; and to study and establish advanced multi-data and multi-feature synergistic toxic weed extraction models based on semantic and machine learning; (4) comparing the performance of different combinations of data and methods, and finally, the methodology for extraction poisonous weed component information over a large area by using multi-source remote sensing will be formed and promoting the application. This project is going to provide important ecosystem structure indicators and technical support for more objective and scientific monitoring rangeland degradation, protecting ecological system and biological diversity, as well as forage and livestock balance management over the large area of arid, semi-arid and alpine regions. This project will also provide demonstration and reference for the researches and applications on the similar regions.

通过基金项目（41271361）及其他研究发现以遥感植被指数为代表的草地退化监测方法难以反映群落结构的变化，以及由于物理限制单一传感器无法同时获得高时、空、谱分辨率并覆盖大区域的遥感数据以供大面积毒杂草组分提取的问题，以三江源为试验区开展研究：（1）建立毒杂草 “生长型”类别和相关地物波谱获取；（2）研究多源遥感数据的互补特性，建立适合大区域缺乏高分数据覆盖的毒杂草提取多源遥感协同方法；（3）对毒杂草光谱、物候、生化、生境及特征提取与选择进行研究，构建有效的特征空间；研究建立基于语义和机器学习的多数据、多特征协同的毒杂草组分提取模型；（4）对比评估各种原始及协同数据、协同方法的提取效果，提出区域尺度毒杂草组分遥感提取优化方法并推广应用，为干旱、半干旱及高寒地区科学客观的草地退化监测、生态系统保护、草畜平衡管理等提供重要的生态系统结构指标和技术支持，为类似地区问题的研究及应用提供借鉴和示范。

以单一传感器和遥感植被指数为代表的草地退化监测方法难以实现对草地群落结构变化的监测。本项目进行了多源遥感数据协同互补的理论方法及大区域草地组分多源遥感协同提取的研究。主要研究内容及结论如下：.1、毒杂草“生长型”类别定义和典型地物光谱获取.通过实地考察，明晰了毒杂草“生长型”情况。获得了草地组分覆盖度样方、地面实测光谱和无人机高光谱图像数据集。构建了“空-天-地”可食牧草和毒杂草“原始光谱、一阶导数光谱、地形特征、生化特征、植被指数、引导滤波和扩张形态学属性剖面特征”等的识别特征体系。.2、多源遥感互补协同相关理论与方法.应用HJ-1A CCD/HSI、Landsat 8、Hyperion、Sentinel-2和 MODIS等多源遥感数据，探索了适用于研究区草地组分信息提取的时-空-谱图像数据融合方法。.（1）引入空间分辨率增强的ASF方法，实现了星载高光谱图像的空间分辨率增强。.（2）基于光谱梯度角匹配同名点、混合像元分解与插值思想建立了ISRME、ASREM和MSREM光谱分辨率增强方法。.（3）探讨时空自适应及改进融合模型在研究区的适应性，生成高时、空分辨率的植被指数时间序列，为研究结合物候特征的草地组分信息识别提供数据基础。 .3、可食牧草、毒杂草组分覆盖度多源遥感协同提取模型方法.（1）提出了基于形状自适应邻域信息的半监督学习模型（SANI-SSL），实现了小样本情况下的无人机高光谱图像草种精细识别。.（2）构建了卫星高光谱新的可食牧草比值植被指数（YSRVI）、归一化植被指数（YSNDVI）和毒杂草光谱反射率累积植被指数（DZAVI）；提出了空间光谱约束再聚类的训练样本扩展算法；建立了多特征复合三核SVM和自适应空间邻域稀疏表达的多维小样本草地组分信息提取优化方法。.（3）建立了生境-光谱多元信息协同的毒杂草空间分布模拟方法。.（4）提出了基于特定对象的高空间分辨率遥感图像的多层多尺度分割优化方法和一种端对端的深度神经网络结构GAMNet。为研究区毒杂草和可食牧草识别提取提供了新方法。.4、采用多种方法对建立的各种模型、方法进行了性能评估，得到了有益结论。