洞庭湖湿地NPP反演模型优化及其时空变化驱动机制研究

Carbon sequestration capacity of wetland vegetation and its impact on regional climate and environment are the focus of global attention in recent years. Using remote sensing data and driving ecological models has become one of the most important methods for global NPP simulation. Nowadays, low spatial resolution images, such as MODIS data, are mostly used in long time series wetland vegetation NPP research. Although it can reflect the temporal variation of NPP time series, the spatial expression is not very precise. Based on satellite sensors, a variety of models has been developed to estimate NPP, in which Carnegie–Ames–Stanford Approach (CASA) is the most widely used model. In previous studies of estimating wetland NPP using CASA, the potential maximum light utilization mostly used the existing research results, which still need to be constructed and determined according to the actual situation of wetland. Furthermore, it is found that linear correlation models are often used in the study of temporal and spatial variations and driving mechanisms of NPP. However, for regions with complex climate, simple linear correlation models can not accurately express the driving relationship between climate factors and NPP. In view of the existing problems in NPP estimation and analysis of regional wetlands, the project takes Dongting Lake wetland as the research object, obtains time-series Landsat data by using space-time fusion technology, and uses convolutional neural network algorithm to achieve accurate extraction of wetland vegetation information. Then, combined with the actual situation of Dongting Lake wetland, the max was constructed, and the NPP of wetland was estimated using CASA model. Finally, the factor field analysis method was introduced to analyze the temporal and spatial changes and driving factors of NPP in Dongting Lake wetland. The results of the project are expected to provide effective approaches for improving NPP estimation and analysis of regional wetlands.

湿地植被的碳储量及其对区域气候的影响是近年来全球关注的焦点。利用遥感数据与模型目前已成为模拟NPP的主要方法之一。目前长时间序列湿地植被NPP研究多采用低空间分辨率影像，尽管能够反映NPP的时序变化，但空间表达不够精细。另一方面，以往利用CASA估算湿地NPP研究中，最大光能利用率多采用已有的研究成果，尚需综合湿地实际情况进行构建。同时，对于气候复杂地区的时序NPP空间变化与驱动因子分析，采用传统的线性相关模型已经无法满足。项目以洞庭湖湿地为研究对象，拟利用时空融合技术获取时序Landsat数据，并采用卷积神经网络算法实现湿地植被信息的精确提取。在此基础之上，结合洞庭湖湿地实际状况构建了最大光能利用率参数，并驱动CASA模型估算了多年洞庭湖湿地NPP。最后，引入要素场分析方法分析了洞庭湖湿地多年NPP的时空变化及驱动因子。项目结果以期为改善与提高区域湿地NPP估算与分析提供有效的技术方法。

湿地是地球上重要的“碳库”之一，在调节碳平衡以及维持全球气候稳定方面具有不可替代的作用。在湿地斑块破碎化程度不断增大的背景下，利用遥感技术准确地识别湿地类型并估算其净初级生产力（NPP）是目前亟待解决地科学问题。本项目以洞庭湖湿地为研究区，以中等分辨率影像（Landsat）为遥感数据源，基于时空融合技术获得了时间序列的Landsat数据，联合深度学习与集成学习算法实现了湿地植被分类，通过修正CASA模型估算了2000-2019年洞庭湖湿地植被NPP，并利用偏相关方法探明了洞庭湖湿地植被NPP时空演变的驱动机制。研究结果发现，基于遥感云计算的时空融合模型能快速、准确地获取时序Landsat数据，融合影像与真实影像的决定系数R2>0.85；联合深度学习与集成学习在湿地分类过程中既能保持较高的分类精度多年总体分类精度均高85%，也具有较好的鲁棒性和稳定性；利用修正的CASA模型模拟的湿地植被NPP与实测数据具有较高的相关性（R2>0.7，RMSE<21gC.m-2）；洞庭湖湿地区域的年植被NPP呈明显上升趋势（0.86gC.m-2.yr-1，P<0.05），气候变化对洞庭湖湿地植被NPP变化有正贡献（11.95gC.m-2.yr-1），而人类活动则有负贡献（-5.89gC.m-2.yr-1）。