湿地序列监测图像的地物目标分类方法

A Ground Object Classification Method Based on Wetland Sequence Monitoring Images

针对固定观测站采集的序列图像中光照变化和背景复杂的问题,将图像序列作为图像在时间维的拓展,提出一种结合颜色空间变换和神经网络的目标分类方法,以辽河口湿地为例开展实地观测试验。基于前期现场调查得到的地物分布信息,将序列图像转换至HSV空间,选择像素小区域作为样本训练多层神经网络,获得更高维度分类依据映射,并完成潮滩生态监测区域目标分类及重要湿地植物翅碱蓬的分布信息提取。对提取结果进行定量验证,该方法整体精确率稳定在90%左右。

A target classification method combining color space transformation and neural network was proposed to address the issues of illumination variation and complex backgrounds in the sequence images collected by fixed observation stations,using image sequence as an extension of the image in time dimension,and taking Liaohe Estuary Wetland as an example of field observation.Based on the distribution information of ground object obtained from previous field investigation,the sequence images were converted to HSV space.The multi-layer neural network was trained by selecting pixel small regions as samples to obtain a higher dimensional classification basis mapping,and achieve the target classification of tidal flat ecological monitoring areas and the extraction of distribution information of important wetland plant Suaeda salsa.Quantitative validation of the extraction results showed that the overall accuracy of this method was stable at around 90%.