面向遥感分类精度评价的样本点空间布局优化方法研究

Accuracy assessment is a key step and important segment for remote sensing data products. Currently, double representations of sampling sites in feature and geographical space are very poor for accuracy assessment of remote sensing classification. This study presents spatial sampling layout optimization method for accuracy assessment of remote sensing classification. Firstly, the spatial stratification and sample types division will be adopted for sampling optimization in feature space. Secondly, the spatial simulated annealing and minimization of the mean of shortest distances criterion will be used for sampling optimization in geographical space, and the optimization results will be evaluated. Finally, the comparative experiments between this method and other spatial sampling methods will be set up to evaluate the accuracy, effectiveness and stability of different methods from three aspects of overall accuracy, root mean square error and standard deviation. Spatial sampling layout optimization method developed in this study will be integrated after improvements. This study can provide a robust operating tool and method for accuracy assessment of remote sensing classification, and support services for the validation of remote sensing products, detailed investigation of soil pollution, survey and monitoring of agricultural product origin, and so on. The study has important scientific significance in the theoretical and technical level and broad application prospect in the practical and application level.

精度评价是遥感数据产品研发的一个关键步骤和重要环节。针对目前遥感分类精度评价样本点特征空间和地理空间双重代表性差的问题，本项目研发一种面向遥感分类精度评价的样本点空间布局优化方法。首先，基于空间分层和样本点特征属性类别划分进行遥感分类精度评价样本点特征空间优化；其次，采用平均最短距离最小化准则构建样本点优化目标函数结合空间模拟退火算法对样本点进行地理空间优化，并对优化结果进行精度评价；最后，研究设置与其他空间抽样方法的对比实验，从总体精度、均方根误差和标准偏差三个方面来评估不同方法的精确性、有效性和稳定性，经过改进完善后集成样本点空间布局优化方法。研究成果从理论技术层面可以为遥感分类精度评价提供一个鲁棒性的操作工具方法，从实践应用层面可以支撑服务于遥感产品真实性检验、土壤污染详查、农产品产地调查监测等，具有重要的科学意义和广阔的应用前景。

精度评价是遥感数据产品合理有效利用的基础和保证。针对目前用于遥感分类精度评价的样本点一般无法实现特征空间和地理空间的均匀或无偏的问题，项目以京津冀遥感分类数据为例，研究提出了一套面向遥感分类精度评价的样本点空间布局优化方法，取得主要成果如下：基于多尺度空间异质性和图像八邻域算法构建了不同的土地利用空间分层模式，并筛选出最佳空间分层方法，结合样本点特征属性类别划分方法集成了样本点特征空间优化方法；耦合平均最短距离最小化准则和空间模拟退火算法研发了样本点地理空间优化方法，进而集成为面向遥感分类精度评价样本点空间布局优化方法——分层均匀抽样方法；研究构建了总体精度、均方根误差和标准偏差的评价指标体系进行评估，研发的分层均匀抽样方法效果优于空间均匀抽样方法、简单随机抽样方法和分层随机抽样方法，兼顾了样本点在特征空间和地理空间的均匀性和代表性，降低了调查和监测成本，为遥感分类精度评价和真实性检验提供了一个鲁棒性的操作方法，应用前景广阔。项目完成了各项研究任务和考核指标，发表学术论文11篇（其中4篇SCI/SSCI论文、3篇EI论文），申请发明专利5项（其中授权发明专利3项），登记计算机软件著作权6项，培养3名硕士研究生，参加5次国内外学术会议，国际学术会议口头报告2次。