Long Time Series High-Quality and High-Consistency Land Cover Mapping Based on Machine Learning Method at Heihe River Basin
Long Time Series High-Quality and High-Consistency Land Cover Mapping Based on Machine Learning Method at Heihe River Basin
复制标题
DOI：
10.3390/rs13081596
复制
发表时间：
2021-04
期刊：
影响因子：
--
通讯作者：
B. Zhong;A. Yang;Kunsheng Jue;Junjun Wu
中科院分区：
文献类型：
--
作者：
B. Zhong;A. Yang;Kunsheng Jue;Junjun Wu
土地覆盖变化的长时间序列是分析长期气候、环境和生态变化的关键。虽然已经公开发布了几个中到精细分辨率的全球土地覆盖数据集，它们在全球尺度上表现出很强的一致性，但它们在区域尺度上存在很大的偏差;此外，2000年之前的高质量土地覆盖数据集不可用，不同数据集之间的分类一致性不是很好。因此，对高质量和一致性的长时间序列土地覆盖数据集的需求很大，但即使在区域尺度上，这些数据集仍然不可用。由8颗连续卫星组成的大地遥感卫星系列卫星图像可追溯到1972年，因此有可能产生一个长时间序列的土地覆盖数据集。此外，新获得的卫星数据具有构建时间序列卫星图像的能力，并且可以采用LCMM等时间序列分析方法来制作高质量的土地覆盖数据集。因此，结合两类卫星数据的优势，提出了一种基于机器学习的时间序列土地覆盖制图方法，并以黑河流域为例进行了验证。首先，使用LCMM方法检索的HRB 2011-2015年高质量土地覆盖数据集，快速准确地制作训练样本。其次，制定了将2011年之后的训练样本转移到更早年份的策略。最后，利用随机森林模型对选取的年度样本进行训练，生成各年度的土地覆盖图。最后，对评价结果进行了综合分析和验证。在这项研究中，最终形成了一个包括1986年，1990年，1995年，2000年，2005年，2010年，2011年，2012年，2013年，2014年和2015年的长时间序列土地覆盖数据集，平均精度约为90%。该数据集具有良好的时间连续性和稳定性，是HRB分辨率为30 m的最长时间序列土地覆盖图。
Long time series of land cover changes (LCCs) are critical in the analysis of long-term climate, environmental, and ecological changes. Although several moderate to fine resolution global land cover datasets have been publicly released and they show strong consistency at the global scale, they have large deviations at the regional scale; furthermore, high-quality land cover datasets from before 2000 are not available and the classification consistency among different datasets is not very good. Thus, long time series of land cover datasets with high quality and consistency are in great demand but they are still unavailable, even at the regional scale. The Landsat series of satellite imagery composed of eight successive satellites can be traced back to 1972 and it is, therefore, possible to produce a long time series land cover dataset. In addition, the newly available satellite data have the capability to construct time series satellite images and a time series analysis method such as LCMM can be employed for making high-quality land cover datasets. Therefore, by taking the advantages of the two categories of satellite data, we proposed a new time series land cover mapping method based on machine learning and it, thereafter, is applied to Heihe River Basin (HRB) for verification purposes. Firstly, the high-quality land cover datasets at HRB from 2011–2015, which were retrieved using the LCMM method, are used for quickly and accurately making training samples. Secondly, a strategy for transferring the training samples after 2011 to earlier years is established. Thirdly, the random forest model is employed to train the selected yearly samples and a land cover map for every year is subsequently made. Finally, comprehensive analysis and validation are carried out for evaluation. In this study, a long time series land cover dataset including 1986, 1990, 1995, 2000, 2005, 2010, 2011, 2012, 2013, 2014, and 2015 is finally made and an average precision of about 90% is achieved. It is the longest time series land cover map with 30 m resolution at HRB and the dataset has good time continuity and stability.