针对传统全球数字高程模型修正方法忽略训练样本质量等问题,提出一种基于主动学习的沿海区全球数字高程模型(GDEM)修正方法。该方法首先选择一定数量的代表性样本点作为初始训练集,然后通过聚类批处理模式采样算法,采用迭代方式选取高质量代表点进行模型训练,最后利用选取的代表点构建机器学习模型以实现GDEM修正。以美国克逊维尔为训练区、查尔斯顿为迁移实验区,选取中误差和平均绝对误差验证模型的精度。实验结果表明,与传统的GDEM修正方法相比,本研究方法仅需选择8.57%的采样点即可完成模型训练,且GDEM的中误差降低了3.31%~51.65%、平均绝对误差降低了4.76%~48.72%。在迁移实验区,修正后COPDEM30的中误差从6.52m降至1.68m。相比于传统方法,本研究方法的中误差和平均绝对误差分别降低了24.82%和30.28%,证明了模型具有一定的迁移性。

To address issues such as neglecting sample quality in traditional global digital elevation model (GDEM) correction methods, this paper proposed a GDEM correction method for coastal areas based on active learning. Firstly, a certain number of representative sample points were selected as the initial training set. Then, high-quality representative points were iteratively selected for model training through a clustering-based batch processing sampling algorithm. Finally, a machine learning model was constructed by using all selected representative points to achieve GDEM correction. The accuracy of the model was validated by selecting mean square error and mean absolute error with the coastal areas of Jacksonville and Charleston in the United States as the training area and transfer experimental area respectively. Experimental results show that, compared with traditional GDEM correction methods, the proposed method only requires 8.57% of the sampling points to complete model training. The root mean square error of GDEM is reduced by 3.31% to 51.65% and the mean absolute error is reduced by 4.76% to 48.72%. In the transfer experiment area, the root mean square error of the corrected COPDEM30 is reduced from 6.52 m to 1.68 m. Compared with traditional methods, the root mean square error and mean absolute error of the proposed method are reduced by at least 24.82% and 30.28% respectively, demonstrating that the model has a certain level of transferability.