|
High-quality training samples are crucial to the accuracy of land cover classification, and traditional sample collection is mostly manual, which is not only time-consuming and labour-intensive, but also not adaptable to the study area with a large range, so this paper proposes an automatic sample collection method based on migration technology to screen out the training samples of Xiongan New Area with unchanged target years compared to the reference years for land cover mapping in the study area, which solves the problem of shortage of training samples for long time series land classification. The problem of shortage of training samples. In this paper, firstly, 10,000 sample points are randomly and uniformly selected based on the ESA_CCI dataset, and secondly, the spectral differences between the reference year2020 and the target years 2018, 2019,2020 and 2021 Sentinel-2 images are calculated, and the Euclidean distance (ED, Euclidean distance) and spectral angle (SAM. Spectral Angle Mapper) as the best magnitude and similarity metrics for bi-chronological variation monitoring, and samples with similar spectral distances and spectral angles tending to 0 are screened as training samples based on certain thresholds, and then the labelled values of the ESA_CCI dataset are assigned to the training samples, and spectral indices and texture analyses, such as NDVI, MDNWI, and NDBI, are added. Combined with the random forest (RF) classifier in GEE (Google Earth Engine) to complete the land cover mapping of Xiongan New Area, and an overall classification accuracy of 84% was obtained. Overall, the method proposed in this paper has high potential for land cover monitoring without sufficient training samples.
|
