苏州太湖三山岛国家湿地公园景观健康评价

Landscape Health Assessment for Sanshan Island of Tai Lake National Wetland Park in Suzhou

苏州太湖三山岛国家湿地公园位于太湖之中,对其进行景观健康评价可以有效评估湿地公园建设绩效,对指导湿地公园的科学性规划与建设具有重要意义。从研究区的景观数据中,提取土地利用类型面积所占比例、道路长度、不同类型坡岸长度等景观评价指标,采用F-score方法和随机森林法,对这些指标进行特征选择,并作为输入特征向量;采用主成分分析方法,分析水质评价指标,将其分为优质、良好、一般和较差4类,并作为输出分类结果。在此基础上,建立支持向量机分类评价模型;通过选择合适的核函数和核参数,实现景观健康评价。研究结果表明,苏州太湖三山岛国家湿地公园景观健康水平具有明显的空间差异性,研究区西部几乎未受到人为干扰,景观健康等级为优质;主岛的东北部位于旅游核心区域,人为干扰程度较重,景观健康等级大多属于一般和较差;主岛东南部建有人工修复湿地,明显改善了该区域的景观健康等级。在模型训练过程中,对比多项式核函数、Sigmoid核函数和高斯核函数三者的分类结果发现,高斯核函数的分类精度最高,训练集精度为90.0%,检核集精度为85.7%。在影响湿地公园景观健康因素众多、野外监测数据较少的条件下,本研究采用的支持向量机方法可以快速取得合理的评价结果,揭示湿地公园景观健康的空间差异。

Sanshan Island of Tai Lake National Wetland Park in Suzhou is Located in the Tai Lake. Landscape health assessment for it could effectively evaluate the performance of wetland park construction and has significant meaning for guiding scientific planning and construction of wetland park. Landscape indices of land use types＇ area percentage, length of roads, and length of different types of sloping banks, which were extracted from landscape data, were used as import feature vector after feature selections using F-score and Random Forest methods. Meanwhile, water quality indexes calculated using principle component analysis method and classified into excellent, good, ordinary and bad classes, were used as output classification result. Thus, model of support vector machine was built by combining landscape indices with water quality indexes. After model training and testing with proper kernel function and kernel parameters, landscape health assessment for this wetland park was implemented. Experimental results showed that there were obvious spatial differences of landscape health level in this wetland park. The west of study area had almost no anthropogenic interference;landscape health level of this area was excellent. The northeast of the main island was core area of tourism and receiving intensive human disturbance. Landscape health levels of this area were ordinary or bad. Constructed wetland was close to the southeast of the main island, which could make significant improvement of landscape health level in this area. During model training, classification accuracy of Gaussian kernel function was better than polynomial and sigmoid kernel function. With Gaussian kernel function, the classification accuracies of training set and test set were 90.0% and 85.7% respectively. Our svm-based method could obtain reasonable assessment result and reveal the spatial difference of landscape health quickly and efficiently, especially under the conditions of complex factors affecting landscape health of the wetland park and scarce field monitoring samples.