ABSTRACT

Long-term and large-scale lake statistics are meaningful for the study of environment change, but many of the existing methods are labour-intensive and time-consuming. To overcome this problem, a novel method for long-term and large-scale lake extraction by shape-factors- and machine-learning-based water body classification is proposed. An experiment was conducted to extract the lakes in the Yangtze River basin (YRB) from 2008 to 2018 with the Joint Research Centre's Global Surface Water Dataset (JRC GSW) data and OSM data. The results show: 1) The proposed method is automatically and successfully executed. 2) The number of river–lake complexes is between 3008 and 4697, representing 3.56%–5.70% of the total water bodies. 3) The areas of the lakes and rivers in the YRB were obtained, and the accuracy of water classification in each year was stable between 90.2% and 93.6%. Comparing the back propagation neural network, random forest, and support vector machine models, we found that the three machine learning models have similar classification accuracy for the scenario. 4) Fragmented and incomplete small rivers in the JRC GSW data, unchecked training samples, and overlapped shape factors are the three error sources. Future work will focus on addressing these three error sources.

摘要

长期和大规模的湖泊统计数据对于研究环境变化具有重要意义，但许多现有方法耗费大量人力且费时。为了克服这个问题，提出了一种新的基于形状因子和机器学习的水体分类方法，用于长期，大规模的湖泊提取。利用联合研究中心的全球地表水数据集（JRC GSW）数据和OSM数据，从2008年至2018年进行了提取长江流域（YRB）湖泊的实验。结果表明：1）所提出的方法是自动成功执行的。2）河湖综合体的数量在3008至4697之间，占水体总数的3.56％至5.70％。3）获得了黄河三角洲的湖泊和河流面积，每年的水质分类准确率稳定在90.2％至93.6％之间。比较反向传播神经网络，随机森林和支持向量机模型，我们发现这三种机器学习模型对场景的分类精度相似。4）JRC GSW数据中的零散和不完整的小河，未经检查的训练样本和重叠的形状因子是三个误差源。未来的工作将集中于解决这三个错误源。形状因子重叠是三个误差源。未来的工作将集中于解决这三个错误源。形状因子重叠是三个误差源。未来的工作将集中于解决这三个错误源。